Real 32-bit DDR3 widening (2x MT41J128M16JT-125:K chips ganged in
parallel, user's own MIG wizard session). Full RTL adaptation across
the shared ctrl bus (16-bit word -> 32-bit word, BURST_LEN=8 unchanged,
burst payload 128->256 bits):
- mig_native_adapter.v: app_wdf_data/app_rd_data 64->128 bits (real,
confirmed against the regenerated MIG wrapper), beat count unchanged.
- act_tile_fetch.v: real logic change - burst now holds 4 tiles instead
of 2 (sel_lat extended to 2 registered bits, 4-way case mux instead
of 2-way ternary, same request-time-registered-select discipline as
EXP-0081). Not a further bytes/MAC reduction, just what's needed to
keep 100% packing utilization at the larger burst.
- host_mem_bridge.v: real addressing redesign - host-facing 16-bit-word
contract kept unchanged (ESP32 firmware unaffected), internally
translated onto the new 32-bit-native ctrl bus.
- sdram_arbiter_n.v, layer_prefetch_ctrl.v, packed_slot.v,
ddr_prefetch_mgr.v, n2_system_ddr3_top.v: mechanical width bump plus
doubled ddr3_dq/dqs/dm pins and the real differential sys_clk/clk_ref
top-level ports the regenerated MIG now requires.
New burst_mem_model32.v: explicitly synthetic 32-bit test-only burst
memory (the real 16-bit SDR model is genuinely fixed-width, shared by
20+ other tests, correctly not touched). Found and fixed a real
address-aliasing bug in it during bring-up (MEM_ADDR_BITS=16 silently
wrapped a real 0x10000 test address to 0).
Real verification: all isolated testbenches re-verified (10/10, 33/33,
32/32, 7/7, 9/9 PASS), plus real xsim against the real 2-chip DDR3
model (tb_mig_native_adapter.v 12/12 PASS, tb_n2_system_ddr3.v 8/8
PASS, both chips visibly returning different real data).
Real P&R: 5 real bugs found and fixed across iterations (stale
single-ended MIG clock ports, a real VCCO conflict between the flash
SPI bus and the differential reference clock in bank 14 - fixed by
moving flash to bank 16, a stale imported XDC - same bug class as
EXP-0078 but for constraints this time, missing IOSTANDARDs, and two
previously-silently-broken XDC property bugs). Route completes 100%,
but real timing does NOT close: WNS -0.618ns, 213 failing endpoints.
Honest root cause: the violation is inside neural_processor_packed.v's
own packed-MAC accumulation tree, unchanged since EXP-0059 - it has
real margin at the old 155.039MHz ui_clk but not at the new 172.414MHz
the paired clock-period change produced. This is NOT caused by the
32-bit width change itself. Width alone, even at the old clock, already
delivers the full intended 2x bandwidth gain (1.24 -> ~2.48 GB/s) -
width and clock rate are separable levers. Current trustworthy timing
signoff remains EXP-0083 (16-bit, +0.073ns) until the clock period is
reverted toward 3225ps (keeping Data Width=32) in one more real,
user-gated MIG wizard session.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUG92aM9m68TRc4rG55BcC
New ddr_prefetch_mgr.v wraps act_tile_fetch.v with a depth-2 ping-pong
buffer, issuing the next tile's DDR3 fetch as soon as the fetch engine
is free instead of waiting for packed_slot.v to finish consuming the
current tile. Wired into packed_slot.v's tile loop (job-level start
instead of per-tile req), simplifying the S_TILEWAIT join in the process
(ddrpf_tile_valid is level-held, no separate act_seen latch needed).
Verification: new tb_ddr_prefetch_mgr.v (25/25 PASS after fixing a real
testbench polling race found via iteration-tagged tracing, not an RTL
bug), tb_packed_slot.v re-run unmodified (9/9 PASS, bit-identical
results), tb_n2_system_ddr3.v re-run via real xsim against real
ddr3_model.sv (8/8 PASS). Real P&R: WNS +0.073ns (up from EXP-0082's
+0.068ns), LUTs 5644, DSP48E1 16 unchanged, 0 failing endpoints.
Honest result: real A/B on the actual DDR3 backend (same testbench,
before/after) shows a real but modest 2.86% reduction in total
simulated time - smaller than the original hypothesis suggested, because
neural_processor_packed.v already accepts one operand per cycle, so the
per-tile dead time being removed was already small relative to real DDR3
fetch latency. Docs updated to report this honestly rather than oversell
it; the larger multi-slot DDRManager is deferred pending re-measurement
against the (still pending, user-gated) 32-bit channel widening.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUG92aM9m68TRc4rG55BcC
Real P&R: WNS +0.068ns (up from +0.030ns), 0 failing endpoints, 5437
LUTs, 16 DSP48E1. Confirms EXP-0081's "register the select bit at
request time" design genuinely kept the fix off the critical path.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUG92aM9m68TRc4rG55BcC
Implements the highest-leverage fix from EXP-0080's bottleneck
analysis: act_tile_fetch.v now packs 2 consecutive tiles per DDR3
burst (even tile low 64 bits, odd tile high 64 bits) instead of 1
tile per burst, halving real DDR3 bytes-per-useful-byte. Timing-safe
by construction: the tile-index select bit is registered at request
time, long before the real DDR3 round-trip completes, never racing
the arriving read data (unlike the runtime part-select pattern
EXP-0079 deliberately avoided).
Re-verified at all 3 levels (isolated engine 8/8, packed_slot.v 9/9
with bit-identical results to EXP-0079, full N=2 system on real DDR3
8/8) -- the real JEDEC trace now shows no half-burst padding, direct
confirmation the fix works in practice, not just in theory.
Also: real device data gathered on this package's I/O bank layout
(only 5 banks total, 14/15/16/34/35) informing the next bandwidth step
(32-bit-wide single controller recommended over a second independent
channel, given the pin/logic cost comparison).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUG92aM9m68TRc4rG55BcC
Before building N=4/8/16 core scaling, did the requested full
analysis. Real finding: using measured DDR3 throughput (1.24 GB/s,
from the actual EXP-0079 JEDEC trace) against calculated compute-side
need (4.96 GB/s for one core at peak DSP throughput, given the
current activation memory layout's 2x byte overhead), the system is
memory-bandwidth-bound already at N=1/N=2, not DSP-bound (only 6.67%
DSP used). Scaling core count today would show no real throughput
gain.
docs/ARCHITECTURE_ANALYSIS.md: full module review + ranked
interventions -- result-writeback engine (blocker), denser activation
packing (highest-leverage bandwidth fix), then the user's own proposed
DDRManager/orchestrator-prefetch idea (design-sketched, grounded in
neural_director_packed.v's existing job queue, explicitly scoped as
complementary to denser packing, not a substitute for it), THEN
N-scaling with real per-N P&R signoff.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUG92aM9m68TRc4rG55BcC
Closes the last major disclosed functional gap: packed_slot.v's
activation data was read through a combinational stand-in since
EXP-0062. New act_tile_fetch.v reads activation tiles directly from
DDR3 (no on-chip buffering needed, unlike weights -- activation data
has no reuse), sharing each slot's existing ctrl port with its own
weight-prefetch engine. Real memory layout: one full BURST_LEN=8-word
burst per tile, deliberately avoiding any runtime-indexed part-select
given this project's thin P&R timing margin (EXP-0078).
Verified at three levels: act_tile_fetch.v alone (6/6), packed_slot.v
with real preloaded activation data (9/9), and the full N=2 system
against real DDR3 via xsim (8/8, 0 errors) -- the first time this
project's compute path has been verified end-to-end with real DDR3
for both weights and activations.
Retired hardware/v3/rtl/n2_system_top.v and its testbench (pre-DDR3
SDR-placeholder era, fully superseded by n2_system_ddr3_top.v).
Also: docs/PHYSICAL_REALIZATION.md (real pinout/parts/timing/protocol
reference for the physical board) and CLAUDE.md (persistent project
instructions for future Claude Code sessions).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUG92aM9m68TRc4rG55BcC
Found and fixed a real Vivado project-management bug: two source
files (n2_system_ddr3_top.v, spi_host_bridge_v3.v) had stale imported
copies (plus a duplicate at a second path) that silently kept being
used in synth+impl despite live edits, making the first "flash bridge
included" P&R run silently synthesize the OLD design (STARTUPE2
Used=0/1 gave it away). Fixed by removing the stale/duplicate entries
and re-adding both files as direct (non-copied) references.
Final real P&R: STARTUPE2 Used=1/1 (confirms the flash bridge is
genuinely placed/routed), timing still closes but margin is now
thinner and real: WNS +0.013ns, WHS +0.032ns, 0 failing endpoints.
5213 LUTs, 16 DSP48E1.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUG92aM9m68TRc4rG55BcC
Implements the user's board architecture: config flash wired
exclusively to the FPGA, host (ESP32) reaches it only through the
FPGA. flash_spi_master.v is a plain byte-wide SPI master using
STARTUPE2 to reclaim CCLK after configuration (the real, Xilinx-
documented "indirect SPI flash programming" technique, UG470 p94-96).
New opcode 0x40 FLASH_XFER in spi_host_bridge_v3.v relays bytes
byte-for-byte between host and the physical flash bus -- the host
decides the exact SPI NOR command sequence (verified against the real
W25Q32JV datasheet), this RTL knows nothing about flash semantics.
Found and fixed two real bugs during verification: a byte-assembly
off-by-one in flash_spi_master.v, and a genuine protocol-latency bug
in the FLASH_XFER opcode's response timing (needed 2 trailing margin
bytes, not 1 -- the internal flash transfer doesn't start until the
triggering byte finishes, so 1 byte of margin isn't enough). 39/39
tests pass end to end (host SPI -> bridge -> flash_spi_master ->
behavioral flash model).
Wired into n2_system_ddr3_top.v with real pin constraints (flash_mosi
=K17/flash_miso=K18/flash_cs_n=L13, the same pins reserved-but-unused
in EXP-0075) and BITSTREAM.CONFIG.PERSIST=FALSE made explicit.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUG92aM9m68TRc4rG55BcC
Found a real, previously-masked bug in spi_host_bridge_v3.v's physical
layer (inherited unchanged from V1/V2): the bit_count==0 MISO bypass
corrupts the last bit of any multi-byte response whose value happens
to end in a 1 -- every prior test's response data coincidentally
ended in 0, hiding it until the new DEVICE_ID register (0x...01)
exposed it via a real bit-exact mismatch. Fixed by removing the
bypass (verified unnecessary for this protocol's actual usage).
Added REG_WRITE/REG_READ opcodes (0x30/0x31) and a register file
(DEVICE_ID/CONTROL/STATUS/N_SLOTS) for general device control beyond
job submission, per explicit user request. Full regression: 38/38
PASS, including new cases specifically targeting the bit-corruption
bug for both REG_READ and READ_MEM.
New hardware/v3/constraints/n2_system_ddr3_top.xdc: reserves the
FPGA's dedicated Master-SPI config-flash pins (found colliding with
auto-placed design ports in the real routed checkpoint) and assigns
the neural-processor management SPI to real, verified-free, edge-
adjacent pins on xc7a100tcsg324-2.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUG92aM9m68TRc4rG55BcC
n2_system_ddr3_top.v: the first synthesizable top wiring the real
mig_7series_0 DDR3 controller (public wrapper, real calibration) +
mig_native_adapter.v + sdram_arbiter_n.v (3-way: 2 packed_slot + host
raw-access) + neural_director_packed.v + spi_host_bridge_v3.v.
Real Vivado in-context synth+impl against the actual MIG-generated
XDC (pin locations, DDR3 timing exceptions) on xc7a100tcsg324-2:
route_design 100%, all timing constraints met (WNS +0.040ns, WHS
+0.048ns, 0 failing endpoints), 310.078MHz DDR3 PHY clock / 155.039MHz
compute domain, 5140 LUTs / 5952 regs / 16 DSP48E1 / 0 BRAM.
Fixed three real issues found getting here: a SystemVerilog literal
synth_design can't parse, MIG stub port mismatch (calib_tap_* isn't
exposed in this IP config), and a genuine design mistake -- exposing
packed_slot.v's activation-fetch stand-in ports as literal top-level
pins demanded ~360 I/O against the package's 324 total. Made that
interface internal (stub-driven) instead.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUG92aM9m68TRc4rG55BcC
Replaced three uses of the SystemVerilog '0 self-sizing literal with
explicit-width zero-fill so the file synthesizes under Vivado's
synth_design (which has no -sv equivalent in this flow), needed while
adding this module to the real in-context P&R project.
Re-running its isolated regression after that edit surfaced 3/8
failures. Root-caused via git stash (reproduces on the untouched
committed file, not caused by this edit) and a DUT-internal $display:
tb_neural_director_packed.v's own submit_job task drove DUT inputs
with blocking assignment across two separate @(posedge clk) waits,
racing the DUT's own always block under Icarus and causing a spurious
duplicate enqueue. Fixed by switching to nonblocking assignment
(race-free by construction). neural_director_packed.v itself was
correct all along - 8/8 tests pass after the testbench fix.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUG92aM9m68TRc4rG55BcC
Forked from V2's spi_host_bridge.v after finding two real protocol
mismatches: WRITE_JOB carried dependency-manager fields (required/
producer_ids) that neural_director_packed.v's job_in_* port doesn't
have (no dependency manager exists in V3 -- dropped, disclosed, not
silently ignored), and WRITE_MEM/READ_MEM assumed a word-granularity
host-arb port V3 never had (now wired through host_mem_bridge.v,
EXP-0071). Physical SPI layer carried over unchanged.
Verified standalone: 18/18 tests, 0 errors, including a case
exercising the narrower 25-bit MEM_ADDR_WIDTH's own top bit.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUG92aM9m68TRc4rG55BcC
Closes part of the gap found re-auditing spi_host_bridge.v against V3:
V3 had no host raw-memory-access path into DDR3 at all. This module
translates single-16-bit-word req/wr/addr/wdata/lb_n/ub_n transactions
(spi_host_bridge.v's own WRITE_MEM/READ_MEM shape) into BURST_LEN=8
transactions on the shared arbiter, using the project's existing
DQM-style partial-burst masking technique.
Verified standalone against the SDR SDRAM placeholder: 16/16 tests,
0 errors, including cross-word-corruption checks on every burst
offset. Not yet wired into the N=2 system or driven by real SPI
opcode decode.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUG92aM9m68TRc4rG55BcC
Swapped mig_native_adapter.v + sdram_arbiter_n.v into the full N=2
system (neural_director_packed.v + 2x packed_slot.v), replacing the
SDR SDRAM placeholder used since EXP-0057. Verified against the real
Vivado-generated ddr3_model.sv end-to-end: 8/8 positions bit-exact
against golden model, 0 errors, real JEDEC command traffic observed.
This is the first fully real V3 system-level correctness result:
real packed DSP cores + real weight-reuse scheduling + real N-way
arbitration + real DDR3 timing, all verified together.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUG92aM9m68TRc4rG55BcC
TEST2 fired all 3 simulated requesters' one-shot req pulse
unconditionally, not waiting for each one's own grant -- an
unrealistic stimulus that doesn't match packed_slot.v's real
S_MEMWAIT usage (wait for grant, then fire). Rewrote with parallel
fork branches, each waiting for its own req_grant first, still
exercising the real simultaneous-activation contention case.
7/7 PASS, 0 errors. sdram_arbiter_n.v is now genuinely verified, not
just written.
Full writeup in hardware/v2/logs/experiments.log EXP-0069.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUG92aM9m68TRc4rG55BcC
New hardware/v3/rtl/mig_native_adapter.v: adapts this project's
established req/wr/addr/wdata/wmask->rdata/ready/busy contract to the
real MIG 7-series native app interface (app_addr/app_cmd/app_en,
app_wdf_data/app_wdf_mask/app_wdf_wren/app_wdf_end, app_rd_data/
app_rd_data_valid/app_rd_data_end), derived from this project's own
real generated mig_7series_0.v port widths, not assumed. Runs in the
ui_clk domain (MIG's own generated clock becomes this project's
system clock going forward).
Verified against MIG's own real, vendor-shipped DDR3 behavioral model
(ddr3_model.sv) via real Xilinx xsim/xvlog/xelab (UNISIM primitives
in MIG's PHY require this over Verilator): 12/12 write-then-read-back
transactions bit-exact, 0 errors, real JEDEC command sequence observed
(Activate/Write/Read/Precharge). Confirms the app_cmd encoding and
burst/beat sequencing on first real test.
Also adds hardware/v3/rtl/sdram_arbiter_n.v (generalized N-way
arbiter, generalizing EXP-0066's 2-way version for N>2 scaling and a
future host-access requester) -- its own isolated test currently
HANGS, root cause not yet found, do not trust this module yet
(disclosed, not hidden).
Full writeup in hardware/v2/logs/experiments.log EXP-0068.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUG92aM9m68TRc4rG55BcC
New hardware/v3/rtl/n2_system_top.v: synthesis wrapper around the
EXP-0066-verified system (neural_director_packed.v + arbiter + real
SDRAM controller + 2 real packed_slot.v instances).
Real Vivado post-route: 16/240 DSP48E1, WNS -2.570ns @ 200MHz ->
Fmax ~132.1MHz -- only -2.1% vs the isolated single core (134.9MHz,
EXP-0059). Unlike V2/ECP5 (where the real full-system Fmax was
measurably lower than the isolated-core number), this Director+
arbiter architecture shows no comparable penalty at N=2, real
confirmation (not projection) that the earlier ~55-85x-over-ESP32-S3
estimate's key assumption holds at this scale.
Full writeup in hardware/v2/logs/experiments.log EXP-0067.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUG92aM9m68TRc4rG55BcC
New sdram_slot_arbiter2.v + tb_np_director_n2_system.v: real
neural_director_packed.v dispatching to 2 real packed_slot.v
instances sharing one real SDRAM controller. Jobs submitted one at a
time through the Director's own producer interface -- the Director's
own scheduling decisions determine slot assignment here, unlike every
prior V3 test.
Bug 1 (real, structural): the arbiter's first design registered its
grant one cycle late; layer_prefetch_ctrl.v's ctrl_req is a one-shot
pulse with no retry (every prior use wired it directly to a
controller, never behind arbitration), so a slot's first request
could be silently lost, hanging it forever. Fixed with a new
S_MEMWAIT state in packed_slot.v (wait for a combinational mem_grant
before ever pulsing layer_prefetch_ctrl's start) and a combinational-
first grant in the arbiter.
Bug 2 (testbench): node_id used a stray bit-slice (li[15:8]) instead
of a real multiply, making every layer produce the same node_ids and
silently checking results against the wrong layer's golden value.
Fixed.
Result: 12/12 PASS, 0 errors, real concurrent execution across both
slots (slot 0: positions {0,1,4,5,8,9}, slot 1: {2,3,6,7,10,11}).
Also noted (user correction): the SDR SDRAM controller used
throughout this memory path is a declared placeholder -- the real
target is DDR3 on a custom XC7A100T board, not yet built.
Full writeup in hardware/v2/logs/experiments.log EXP-0066.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUG92aM9m68TRc4rG55BcC
Promotes EXP-0062's own procedural testbench sequence (prefetch ->
buffer swap -> per-tile gather -> operand streaming -> result
capture) into real synthesizable RTL, wrapping layer_prefetch_ctrl.v
-> layer_weight_buffer.v -> weight_tile_gather.v ->
neural_processor_packed.v behind a 9-state FSM matching
neural_director_packed.v's own per-slot contract.
First run: 4/9 failed, deterministic. Root-caused (not a sequencer
bug): the testbench's own w_base computation wrongly treated it as a
byte address needing *2 conversion; layer_prefetch_ctrl.v expects a
word address directly, and packed_slot.v already passes it through
unconverted to match. Fixed the testbench.
Re-verified: 9/9 PASS, 0 errors, bit-exact results and correct
node_id/result_addr passthrough, entirely self-sequenced (no
testbench-side procedural driving of the sub-modules).
Full writeup in hardware/v2/logs/experiments.log EXP-0065.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUG92aM9m68TRc4rG55BcC
Forked from neural_director.v (M5): dispatches PAIRS of queued jobs
(sharing w_base+n_tiles) to packed-core slots instead of one job per
slot, matching neural_processor_packed.v's A/B job structure. If the
two oldest queue entries don't share w_base/n_tiles, the Director
stalls (never mis-pairs) -- a disclosed scope limitation, not hidden.
Isolated testbench with behavioral per-slot stubs (same DEC-0007 scope
decision as tb_neural_director.v). First run: 3/7 tests failed --
investigated each, root-caused as testbench timing bugs (checking
dispatch state before the Director's own FSM had caught up, and a
held-too-long job_in_valid making push counts ambiguous), not Director
bugs. Fixed the testbench, re-verified: 8/8 PASS, 0 errors.
Full writeup in hardware/v2/logs/experiments.log EXP-0064.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUG92aM9m68TRc4rG55BcC
New hardware/v3/rtl/np_packed_weight_reuse_top.v: flat structural
synthesis wrapper around the EXP-0062-verified module chain (real SDR
SDRAM controller -> layer_prefetch_ctrl.v -> layer_weight_buffer.v ->
weight_tile_gather.v -> neural_processor_packed.v).
Real Vivado post-route: 8/240 DSP48E1 (unchanged, memory path uses
zero DSPs), WNS -2.502ns @ 200MHz -> Fmax ~133.3MHz, only -1.2% vs
the isolated single core (134.9MHz, EXP-0059). Real memory-path
control logic adds negligible Fmax cost at this scale.
Full writeup in hardware/v2/logs/experiments.log EXP-0063.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUG92aM9m68TRc4rG55BcC
New hardware/v3/sim/tb_np_packed_layer_reuse.v: real SDR SDRAM ->
layer_prefetch_ctrl.v -> layer_weight_buffer.v -> weight_tile_gather.v
-> neural_processor_packed.v, ALL real synthesizable RTL (unlike
EXP-0058, which still had a testbench-only gather step).
First run: 15/16 PASS, 1 FAIL. Root-caused (not re-run away): a
testbench handshake bug, not a DUT bug -- operand_valid was held one
extra clock edge after each accepted tile, double-consuming stale
data every tile on every pair. 15 of 16 "passed" only because this
test's saturating outputs happened to clamp to the same value whether
or not the accumulator was inflated -- disclosed as a real methodology
risk, not swept under the rug. Fixed by dropping operand_valid the
same delta the handshake is observed.
Re-verified after the fix: 16/16 PASS, 0 errors, bit-exact against an
independent golden model, through the complete real RTL path.
Full writeup in hardware/v2/logs/experiments.log EXP-0062.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUG92aM9m68TRc4rG55BcC
Closes the gap EXP-0058 left testbench-only: assembling P_IN
sequential layer_weight_buffer.v byte reads into one weight_data
tile bus, as real RTL instead of a testbench driver task. Avoids the
runtime-indexed-part-select anti-pattern already found and fixed once
in neural_director.v (ERR-0027-class Fmax collapse) by using a fixed
shift-concat instead.
Verified in isolation against a real, unmodified layer_weight_buffer.v:
37/37 tests, 0 errors, bit-exact across sequential, back-to-back, and
non-sequential/repeated (real reuse-position-style) access patterns.
Full writeup in hardware/v2/logs/experiments.log EXP-0061.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUG92aM9m68TRc4rG55BcC
New hardware/v3/rtl/np_packed_array.v: flat array of 8 unmodified
neural_processor_packed.v instances, fully independent I/O, zero
shared arbiter/Director logic -- isolates exactly one variable
(DSP/placement density) from EXP-0059's single-core baseline.
Real Vivado post-route: 64/240 DSP48E1 (26.67%), WNS -2.592ns @
200MHz -> Fmax ~131.7MHz, only -2.4% vs the single-core 134.9MHz.
Placement density alone is NOT the main driver of the ECP5-era gap
between isolated-core and full-system Fmax -- narrows the question
for the still-unbuilt real Director/arbiter/memory integration.
Full writeup in hardware/v2/logs/experiments.log EXP-0060.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUG92aM9m68TRc4rG55BcC
Real Vivado 2026.1 run (not estimated) confirms the DSP48 packing
survives actual Xilinx synthesis: mac2_dsp_packed.v uses exactly 1
DSP48E1, and the full neural_processor_packed.v pipeline uses 8
DSP48E1/240 for 2 jobs -- half the DSP of two separate V2 cores for
the same work.
Post-route (real place_design+route_design, not synthesis-only):
WNS -2.414ns @ 200MHz -> Fmax ~134.9MHz, within 0.5% of the
post-synthesis-only estimate. This is the isolated compute core,
out-of-context -- not yet a real N-core system number, flagged as
such in the log entry.
Full writeup, including the two real toolchain fixes needed to get
Vivado running on this machine (CRLF line endings in installLibs.sh,
missing libncurses.so.5 on Ubuntu 26.04), in
hardware/v2/logs/experiments.log EXP-0059.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUG92aM9m68TRc4rG55BcC
Real nextpnr-ecp5 resource data shows the ECP5 family (45F: 72
MULT18X18D, 85F: 156) cannot reach the 20x-1000x speedup targets
discussed today with this or any redesigned architecture -- N=16 was
already near the 85F's absolute physical ceiling of 19 cores. User
made an informed decision to pause the project rather than chase a
target this hardware family cannot physically deliver.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUG92aM9m68TRc4rG55BcC
New tb_neural_processor_layer_reuse.v wires the real SDRAM controller,
layer_prefetch_ctrl.v and layer_weight_buffer.v into a real
neural_processor.v compute engine: one resident filter is fetched once
and reused across 8 independent jobs per layer, verified bit-exact
against an independent golden dot-product model (32/32 PASS).
Also found and fixed a real testbench-vs-DUT scheduling race present in
tb_layer_prefetch_ctrl.v (and hardened in the new file): clearing a
one-cycle control pulse on the very next clock edge lands the clear in
the same active-region pass as the edge a receiving module's own
synchronous logic reads it at, so the pulse can be silently missed
depending on implementation-defined process ordering. This had been
silently preventing tb_layer_prefetch_ctrl.v's own claimed 8192/8192
result from ever actually being observed; fixed by holding the pulse
past the edge with a real time delay before clearing, and the
8192/8192 result is now genuinely reproducible (5/5 consecutive runs).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUG92aM9m68TRc4rG55BcC
Built the real FSM version of EXP-0057's own task-based prefetch
pattern (bulk-sequential layer fetch via sdram_controller_openrow.v
into layer_weight_buffer.v), so it's an actual instantiable module,
not just a simulation convenience.
Found and fixed a real bug in the process: cur_fill_addr's own address
arithmetic bit-sliced BYTES_PER_BURST down to too few bits
(BYTES_PER_BURST[BIDXW-1:0]), silently truncating 16 to 0 -- every
burst's bytes landed at fill offset 0-15 instead of their real
position, overwriting each other (only each layer's last burst
survived). Root cause: misapplied a widening idiom used safely
elsewhere in this codebase to a case where the target width was
actually too small. Found via a standalone control-flow debug test
first, then tracing data once control-flow was cleared.
Verified: 8192/8192 bit-exact, 0 errors (was 512/8192 before the fix)
through the real controller + SDRAM model, 16 layers.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUG92aM9m68TRc4rG55BcC
EXP-0056: N_SLOTS=16 failed timing on LFE5U-85F (23-24MHz vs 64MHz
target). First hypothesis (dependency_manager.v's serial ready-scan)
was wrong but real -- built and verified priority_encoder_lsb.v (a
generic recursive tree encoder) and dependency_manager_fast.v, bit-
exact equivalent to the original, but integrated it made no real
difference (24.26MHz). The real cause, found from nextpnr's own
critical-path report: nms_activation_fill_ctrl_v3.v's balanced max-
tree was only ever extended to N_SLOTS in {1,2,4,8}, silently falling
back to the original slow scan for 16. Added the missing case
(nms_activation_fill_ctrl_v3_n16.v), verified isolated (10017/10017)
and functionally (D-Stress N=16 still 256/256 bit-exact). Real result:
71.01MHz, PASS at 64MHz (single seed so far).
EXP-0057: built layer_weight_buffer.v, a double-buffered per-layer
weight scratchpad (fill one buffer in the background from SDRAM while
compute reads many times from the other -- weight-stationary reuse,
as opposed to D-Stress's own deliberately zero-reuse pattern). Wired
to the real sdram_controller_openrow.v + sdram_model.v, no new
hardware. For the same 32768 bytes of useful data: zero-reuse costs
27048 real cycles, reuse costs 3777 -- 7.16x real measured speedup on
the SAME SDR SDRAM, no DDR3, no clock change. This is the answer to
whether DDR3 is necessary for a workload class that actually has
reuse (e.g. conv-style face recognition, unlike D-Stress) -- it isn't,
at least not for this reason.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUG92aM9m68TRc4rG55BcC
EXP-0053: sdram_cdc_bridge.v decouples the SDRAM clock (115.2MHz, real
value derived from the board's own existing PLL VCO=576MHz, verified
via ecppll) from the 64MHz compute domain. Isolated: 137/137 tests, 0
errors, but real measured speedup is only 1.095x (not the naive 1.8x
clock-ratio estimate) -- the CDC handshake's own synchronizer
round-trip is a fixed per-transaction tax.
EXP-0054: sdram_controller_openrow.v implements the page-hit/
keep-row-open optimization sdram_controller.v's own header had always
deferred. weight_prefetch_engine_wide.v's real production traffic is
strictly sequential per job and mostly stays within one SDRAM row --
closing/reopening it every tile (today's fixed auto-precharge policy)
wastes tRP+tRCD for no reason. Isolated: 154/154 tests, 0 errors, 0
protocol violations (including the new refresh-while-row-open hazard,
fixed via an explicit precharge-before-refresh path). Real measured
speedup on the actual sequential access pattern: 1.141x.
EXP-0055: composed both, then integrated into the real D-Stress
benchmark (N=4/N=8, 256/256 bit-exact in every config). Result:
open-row ALONE gives a real, consistent ~5% cycle-count improvement
(47445/47468 vs baseline 49927/49909). CDC alone is a real ~8%
REGRESSION. Combined is still a ~4% regression -- the CDC's fixed tax
is paid on every transaction regardless of row-hit, and real D-Stress
traffic interleaves weight-fetch/activation-result access far more
than the isolated same-row test exercised, so open-row's real saving
doesn't offset it. Decision: do not adopt the CDC approach; open-row
alone is the disclosed, real win worth considering for production
next, pending an explicit go-ahead (not applied to the real board top
in this commit -- all additive, existing production RTL untouched).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUG92aM9m68TRc4rG55BcC
Follow-up to EXP-0051: built sdram_controller_pipelined.v, remapping
addr->bank to low-order bits (today's weight region always maps to bank
0) and adding a shadow-slot ACTIVATE lookahead so a different-bank
request can start its tRCD wait during the current transaction's tail.
Phase A (isolated tb_sdram_controller_pipelined.v, 38/38 bit-exact,
independently re-verified this session): mechanism works, saves exactly
2 cycles (tRCD) per different-bank back-to-back pair, matching the
theoretical ceiling derived before measuring (CAS_LATENCY+BURST_LEN are
serial on the shared data bus regardless of bank, so more than tRCD/tRP
was never on the table).
Phase B (integration, tb_nms_dstress_sdram_pipelined.v, independently
rebuilt/rerun): N=4 49760 cycles (-0.33% vs baseline), N=8 49755
(-0.31%) -- both 256/256 bit-exact. Root cause of the gap: the W port's
request/ready protocol is one-at-a-time, so a second, different-bank
request is essentially never already pending while the first is still
in flight, so the mechanism rarely triggers in the real system even
though it's correct when directly stimulated. Not integrated into
production; kept as additive reference for a possible future
arbiter/backend pipelined-dispatch rewrite (out of scope here, larger
and riskier).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YHENedK76onD2Vtc2CMjej
Forked nms_neural_multiprocessor_sdram_unified.v + its D-Stress testbench
into a dual-bank variant (two independent sdram_unified_backend.v
instances, one for weight-fetch, one for activation+result) to test the
Fase-3 memory-bound hypothesis ahead of Phase 2. Simulation-only: the
real board (v2_board_top.lpf) still wires exactly one physical chip,
per STEP19's governing single-SDRAM mandate.
Result is honest but not the hoped-for one: splitting by traffic class
only cuts D-Stress cycles ~8-10% (N=4: 49927->45724, N=8: 49909->44980),
because the AR (activation+result) path was already lightly loaded
(~12% busy) even alone. The real ceiling is the weight-fetch channel
itself, which stays ~77-78% busy even with its own dedicated bank and
zero cross-traffic. Full writeup in experiments.log EXP-0051, including
the refined next-step options this suggests instead of a straight
2-bank board revision.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YHENedK76onD2Vtc2CMjej
N=8 baseline (EXP-0050, real fpga_neural_v2_top + v2_board_top.lpf, fresh
8-seed sweep): 8/8 PASS at 64MHz, worst 80.97MHz, mean 84.62MHz. DSP
64/72 (88.9%) confirms real N_SLOTS=8. This directly contradicts the
brief's own premise (N8 expected to collapse to ~31MHz) and the project's
historical numbers (3-5/8 pass, worst ~55-67MHz) despite bit-identical
RTL.
Investigated and closed the same discrepancy already flagged for N=4:
logs/experiments.log:2563 explicitly tags historical Fmax numbers as
measured on "Yosys 0.68+"; this session's toolchain is Yosys 0.69+59, a
nightly dated the same day as the session -- a confirmed ~59-commit
version gap. Critical-path inspection on both N=4 and N=8 confirms the
underlying RTL congestion the project diagnosed is real and still
present (same bottleneck locations as previously documented) -- the
newer toolchain just places it well enough to still clear 64MHz.
Decision: this session's numbers are adopted as the operative baseline
for Phase 1+ (see timing.log for the full writeup). Also found, not yet
applied: slot_mem_arbiter.v/slot_mem_arbiter_wide.v both still have the
runtime-indexed-crossbar anti-pattern already fixed in neural_director.v,
sitting right on the arbiter<->backend boundary the brief names -- a
concrete Phase 1 candidate.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
New brief (N=8 timing closure, LFE5U-85F retarget, 4/8/16 x 1/2-bank SDRAM
sweep). Phase 0: no RTL changes, only measure the current baseline.
ERR-0030: constraints/v2_unified.lpf's FREQUENCY PORT "clk" was still 80MHz,
a leftover from the STEP19 freeze, never updated to the project's real
64MHz target -- fixed (LPF only, zero RTL/datapath effect).
ERR-0031 (bigger one): the first two synthesis attempts targeted
nms_neural_multiprocessor_sdram_unified.v, which is NOT the real
board-level top -- it's an obsolete wrapper only exercised by one
testbench now. The real target is fpga_neural_v2_top.v (adds the real
PLL, reset_sync, spi_host_bridge, and a second arbitration level), which
is what actually goes through synthesis+P&R for hardware. Re-targeted
correctly, matched against constraints/v2_board_top.lpf (all 17 ports
real-ball-assigned).
An N_SLOTS=8 P&R attempt against the WRONG (wrapper) target ran for 2h42m
without converging on a single seed; discarded rather than trusted.
N_SLOTS=8 baseline deferred by explicit user request until N_SLOTS=4 is
fully understood -- re-attempt against the correct fpga_neural_v2_top
target with an agreed time budget.
Result (EXP-0049, fresh 8-seed nextpnr-ecp5 P&R, real pins): N_SLOTS=4
8/8 PASS at 64MHz, worst-seed 81.20MHz, mean 91.05MHz. Higher than
DEC-0042's historical worst/best (64.55/77.21MHz) despite identical RTL --
disclosed, unresolved (leading hypothesis: nextpnr-ecp5 build/version
difference), adopted as the operative baseline for this session's
toolchain going forward. Full writeup in errors.log/timing.log/
experiments.log (EXP-0049).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Real critical-path trace after the flash #1 revert showed a NEW
bottleneck (neural_director.job_out_slot -> dependency_manager.
node_resolved/node_state, 76-84% routing) distinct from every prior
fix this session -- already a flat, parallel 64-way compare, not a
serial chain, so the established restructuring fix class doesn't
apply. Root cause: any_pending (added for FPGA_DATA_READY) reads
node_state[0:N_NODES-1] combinationally every cycle, adding real
fan-out onto the same congested signal.
User's own suggestion: replace the combinational scan with a
synchronous up/down counter. pending_count +1 on registration
acceptance, -1 on dispatch acceptance; any_pending = (pending_count
!= 0) -- mathematically identical (DEC-0008: nodes never reclaimed
mid-run) but reads one small register instead of scanning a 16-wide
array every cycle.
Verified: D-Stress N=4 bit-exact (49927 cycles, data_ready PASS).
Fresh 8-seed P&R: N_SLOTS=4 @ 64MHz now 8/8 PASS (was 7/8 after the
flash revert), worst seed1 64.55MHz, best seed0 72.37MHz.
See decisions.log DEC-0042.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013xXuuRUWZScuo1DeYJxs3v
Closes the flash #1 RTL gap flagged in DEC-0041: real, unmodified V1
subsystem (flash_slot_manager.v/flash_copy_engine.v/spi_flash_master.v/
crc32.v) now instantiated in fpga_neural_v2_top.v, bridged to the AR
memory bus via a new flash_mem_adapter.v (byte<->word, matches
nms_memory_manager_stream_wide.v's own real masking convention), and
commandable over SPI via a new spi_host_bridge.v opcode (OP_FLASH_CMD,
0x30) using the same byte-counting idiom as OP_WRITE_JOB. Real balls
now in the LPF: flash_sclk=B2, flash_mosi=E2, flash_miso=F2,
flash_cs_n=F3.
New tb_flash_integration_smoke.v: real SPI-triggered OP_FLASH_READ_BLOCK
verified bit-exact (64/64 bytes) against a real V1 flash_model.v
instance, through the new adapter and the widened (2->3 port) host-arb
arbiter; WRITE_JOB regression confirms the new 3rd port doesn't disturb
existing traffic. Full existing regression re-run clean: D-Stress N=4/
N=8 (bit-exact + data_ready PASS), board-level smoke test (11/11),
isolated spi_host_bridge test (18/18).
Honest, disclosed finding: a full 8-seed P&R re-verification shows
N_SLOTS=4 @ 64MHz regressed from 8/8 to 3/8 PASS (worst 60.18MHz).
Root cause traced via the real critical-path report: the SAME
pre-existing arbiter-to-sdram-backend bottleneck already documented all
session, made worse by flash's added die-area placement pressure --
not a new path through the flash logic itself. N_SLOTS=8 essentially
unchanged (6/8, was 5/8). See decisions.log DEC-0042 for full detail
and open decision points.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013xXuuRUWZScuo1DeYJxs3v
Establishes the real ESP32<->ECP5 programming architecture: flash #1
(neural-network data, existing V1 subsystem, ball reserved not yet
wired into V2) stays separate from flash #2 (boot bitstream, MSPI
auto-boot, CFG[2:0]=[0,1,0]); ESP32 talks JTAG only (bit-banged, no
hardware JTAG-master peripheral on S3/C6), updating flash #2 through
the ECP5's own internal sysCONFIG-to-SPI bridge, never driving the
flash pins directly -- zero bus contention, confirmed against the real
Lattice hardware checklist and sysCONFIG user guide.
Adds real, verified ball assignments (official Lattice CABGA381 CSV +
Project Trellis iodb.json) for JTAG, PROGRAMN/INITN/DONE, CFG[2:0],
and the MSPI dedicated pins -- all written to docs/pinouts.md.
Implements FPGA_DATA_READY as real RTL: a system-idle detector
(dependency_manager's any_pending OR neural_director's !queue_empty OR
any active slot), sticky on the busy->idle edge, self-clearing on new
work -- not a per-neuron completion pulse, which was confirmed too
fine-grained. Bit-exact regression re-verified at N_SLOTS=4 and 8
(zero cycle-count change), new explicit data_ready assertion check
added to the D-Stress testbench (PASS both configs), and a fresh
Yosys+nextpnr-ecp5 placement check (0 errors, data_ready placed at G3).
Also fixes a real, independently-found bug while editing an adjacent
file: nms_neural_multiprocessor_sdram_unified.v's own sdram_a port was
still [11:0] (12 bits), stale from before the 64MB/13-bit memory
upgrade. Not exercised by the real board-level top (which wires SDRAM
directly, bypassing this wrapper) but WAS silently truncating A12 in
every D-Stress simulation this session, including today's earlier
ERR-0029 verification runs. Assessed impact: all D-Stress test
addresses used this session decode to rows under 4096 (bit 12 never
actually needed), so no false-positive PASS is believed to have
resulted -- but the full 64MB space was never actually exercised
through this wrapper. Fixed; re-verified bit-exact with identical
cycle counts.
See decisions.log DEC-0041 for full detail.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013xXuuRUWZScuo1DeYJxs3v
Replace the serial "last-match-wins" priority-scan hit-index encoder
with a flat one-hot compare + single-level priority encode, breaking
the serially-dependent PFUMX/OFX cascade real P&R showed dominating
the N_SLOTS=8 @ 64MHz critical path (55.84MHz worst, 4/8 seeds PASS).
Real nextpnr-ecp5 re-verification (32-run matrix, 4 configs x 8 seeds):
N_SLOTS=8 @ 64MHz improves to 5/8 PASS (worst 60.12MHz, up from
55.84MHz). N_SLOTS=4 @ 64MHz stays 8/8 PASS but with reduced worst-case
margin (WNS +0.605ns, down from +2.143ns) as the critical path
relocates to a different, previously-second-worst path in the same
module -- a real, disclosed trade-off, not hidden. 80MHz remains NO-GO
at both processor counts (re-confirmed on the fixed RTL).
Bit-exact, zero functional regression: isolated
tb_sdram_unified_backend.v (40/40) and full D-Stress N=4/N=8 (both
256/256 neurons bit-exact vs golden model).
See errors.log ERR-0029 and decisions.log DEC-0040 for full root-cause
writeup and MEMORY_UPGRADE_64MB_N8.md section 10 for the complete
measured data set and AS4C32M16SB-7BIN pinout tables.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013xXuuRUWZScuo1DeYJxs3v
Adds the consolidated, authoritative record for the 64MB memory
upgrade and N_SLOTS=8 investigation (real datasheet family comparison,
RTL changes, both timing fixes with real P&R data, honest N=8 clock-
closure status, real DigiKey availability for the frozen part).
Marks HARDWARE_FREEZE.md/MEMORY_ARCHITECTURE.md/PRE_PCB_VERIFICATION.md/
PRE_PCB_CLOSURE_4POINT.md's own SDRAM-specific sections as superseded
(they describe the previous 8MB AS4C4M16SA-6TIN part) with pointers to
the new document, rather than rewriting each individually.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013xXuuRUWZScuo1DeYJxs3v
Real Alliance Memory AS4C4M16SA-6TIN datasheet (Rev 5.0, Table 17)
specifies tMRD as a fixed 2-tCK cycle count, not an ns value.
sdram_controller.v modeled it via ns_to_cycles(12), which rounded to
2 cycles by coincidence at every previously-tested frequency
(100/133/166MHz) but rounds to only 1 cycle at the real 64MHz board
target -- an under-provisioned one-time init sequence. Fixed by
hardcoding T_MRD=2, matching how CAS_LATENCY is already modeled.
Verified zero regression: full 9-config legacy sweep + a new
dedicated 64MHz config (461/461 PASS each), N=2/N=4 D-Stress
(identical cycle counts), board-level smoke test (11/11 PASS).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013xXuuRUWZScuo1DeYJxs3v
Root-causes and fixes the real, disclosed defect left open at the end
of the previous STEP20 commit: the board-level SPI host interface
produced wrong compute results when jobs were dispatched with
realistic (widely time-separated) pacing, even though job registration
itself was already confirmed correct at the dependency_manager
handshake.
Root cause: nms_weight_packed.v and nms_activation_replicated.v both
used a REGISTERED SRAM read (rd_data_reg <= mem[addr], one full clock
of latency), but nms_memory_manager_stream_wide.v's own read-ahead
pipeline (its `rd_pending` bit) is designed around a COMBINATIONAL
read -- a request issued this cycle produces data already valid to
capture the very next cycle. A busy, multi-tile job (e.g. the STEP19
D-Stress regression, 16 tiles/neuron) never exposes the mismatch,
since its own weight/activation prefetch always runs far enough ahead
that any given tile has been sitting stable in the SRAM for many
cycles by the time it's actually consumed. An uncontested single-tile
job has zero such margin: its one tile's read fires on the exact edge
the data nominally becomes ready, landing squarely on the missing
cycle and permanently latching stale/zero data.
Fixed by making both SRAMs' reads combinational, with an explicit
same-cycle fill/read address-match bypass for the one hazard a plain
combinational read alone would still miss. No FSM, arbiter, or SDRAM
controller logic was touched.
Verified (Verilator, per this project's own standing DEC-0004
protocol):
- tb_fpga_neural_v2_top_smoke.v: 11/11 PASS -- single job, back-to-back
jobs, a realistic ~85us-gap job pair, and a parametric sweep of
inter-job gaps (100ns/5000ns/50000ns).
- STEP19 D-Stress N=2: 49788 cycles, 256/256 bit-exact -- identical
cycle count to before this fix (zero regression).
- STEP19 D-Stress N=4: 49771 cycles, 256/256 bit-exact -- identical
cycle count to before this fix (zero regression).
- tb_sdram_unified_backend.v (40/40) and tb_spi_host_bridge.v (18/18)
reconfirmed unaffected.
The physical SPI host interface is now verified correct end-to-end.
Real synthesis/P&R of the board-level top (fpga_neural_v2_top.v) is
the deliberate next step, not yet performed this round.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013xXuuRUWZScuo1DeYJxs3v
STEP20 work toward the V2 hardware release gate. Adds real new RTL
implementing the three pieces the previous freeze (V2.0.0) explicitly
left open, plus real, disclosed verification findings. Does NOT
declare hardware release complete -- see below.
New RTL:
- spi_host_bridge.v: real SPI slave protocol engine (WRITE_JOB/
WRITE_MEM/READ_MEM/STATUS/RESET opcodes), replacing the 110-pin
reg_* testbench bus as the intended physical host interface.
Isolated regression 18/18 PASS (tb_spi_host_bridge.v); two real
MISO-timing bugs found and fixed during its own development (see
the module's header for the root-cause writeup).
- ecp5_pll_sys_clk.v: real, tool-generated (Project Trellis ecppll)
EHXPLLL wrapper, 16MHz oscillator -> 64MHz system clock, with a
declared (not fabricated) simulation-only PLL bypass.
- reset_sync.v: standard async-assert/sync-deassert reset bridge
gating on external POR and PLL lock.
- fpga_neural_v2_top.v: board-level top wiring the above around the
STEP19 compute+memory design's own already-frozen submodules
(zero modification to neural_processor.v, dependency_manager.v,
sdram_unified_backend.v, or any other previously-frozen file).
Real findings from this step's own re-verification (both logged in
full in hardware/v2/logs/errors.log):
- ERR-0024: the current Icarus Verilog v13.0 install (updated since
the last freeze) gives WRONG bit-exact results for the
already-committed STEP19 regression. Cross-checked against
Verilator per this project's own standing protocol (DEC-0004) --
the STEP19 baseline (single SDRAM, N=2/N=4, raw reg_* interface) IS
bit-exact correct, reconfirmed today, matching the historical cycle
counts exactly. Two provably-zero-behavior-change declaration-order
fixes were required just to get the current toolchain to elaborate
the already-shipped STEP19 files at all.
- ERR-0025: a real SPI-bridge protocol race (fixed) plus a SEPARATE,
real, UNRESOLVED defect -- two jobs dispatched through the real SPI
path with realistic pacing produce wrong compute results, even
though job registration itself is confirmed correct at the
handshake. Root cause not yet isolated. Committed as a known-failing
regression (tb_fpga_neural_v2_top_smoke.v) documenting the gap
honestly rather than hiding it.
Given ERR-0025 Part B is real and unresolved, synthesis/P&R of the new
board-level top was deliberately not attempted this round, and V2
hardware release is NOT declared complete. See decisions.log DEC-0036
and hardware/v2/docs/{CHIP_READINESS,OPEN_ITEMS}.md for the full,
itemized status.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013xXuuRUWZScuo1DeYJxs3v
FASE #1 hardware freeze for FPGA-Neural V2, N4/P8, single external
SDRAM (Alliance Memory AS4C4M16SA-6TIN) serving weights, activations,
and results through one physical sdram_controller.v instance. Removes
the PSRAM dependency (hardware/v1/rtl/psram_controller.v +
memory_interface.v) from the V2 physical path entirely -- V1 itself
remains fully unmodified, the golden reference.
New RTL: sdram_unified_backend.v (2-way W/AR arbitration over one
SDRAM controller, real per-byte DQM write masking added to
sdram_controller.v for correct single-byte result writes with no
read-modify-write), nms_neural_multiprocessor_sdram_unified.v (the
frozen top-level). Two real bugs found and fixed via full-system
testing before being accepted (ERR-0023): a deadlock and an off-by-one
data-shift bug in the new arbitration logic.
Real results: N=4 and N=2 D-Stress bit-exact (256/256 neurons), 40
real AUTO REFRESH events interleaved with zero corruption, real
Yosys+nextpnr-ecp5 synthesis/P&R for LFE5U-45F-8CABGA381 (149/245
TRELLIS_IO, a real 45-pin reduction from the prior dual-memory
design). Timing is MARGINAL (1/8 P&R seeds >=80MHz), reported honestly
rather than masked by the best seed.
Real, sourced ball-level pinout for the SDRAM bus + clk/rst (39/149
signals, P&R-verified) using the official Lattice ECP5U-45 pinout CSV
found on disk during this step's own pre-commit review -- corrects an
earlier draft that wrongly assumed no real pinout data was available.
Chip readiness: NO. Real, disclosed blockers remain (no physical host
interface exists yet -- the RTL's own reg_* ports are a 110-pin raw
test-harness bus; clock source/PLL decision; power/configuration
component selection) -- see hardware/v2/docs/{HARDWARE_FREEZE,
CHIP_READINESS,OPEN_ITEMS}.md for the complete, itemized status.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013xXuuRUWZScuo1DeYJxs3v
Implements optimization #2 from the final benchmark campaign's own
recommendation, on top of DEC-0015's word-level burst rewrite: a new
shared activation_cache.v module fetches a given activation (X)
vector from PSRAM once instead of once per neuron sharing it - the
exact redundant traffic pattern the dense-layer workloads in this
project's benchmark suite exhibit.
Each memory_manager's own prefetch_engine now fetches WEIGHTS only;
the activation half is requested from the shared cache instead
(single-tag, tile-granular, N_SLOTS request ports, its own real
word-level PSRAM backend via a new dedicated arbiter port).
dataflow_core.v/slot_mem_arbiter.v/neural_multiprocessor.v widened to
N_SLOTS+1 ports to arbitrate the cache's traffic alongside each
slot's weight traffic.
Two real bugs found and fixed during implementation (ERR-0010): a
target-bank/pending-bank race in memory_manager.v's activation-cache
wiring (the same bug class ERR-0006 already fixed once for
pf_target_bank - a later handoff's queued request can overwrite which
bank an earlier, still-in-flight request's ack applies to), and a
repeat of ERR-0009's N_SLOTS=1 zero-width replication bug in
activation_cache.v itself.
Real, measured results: the full final-benchmark campaign (24/24
workload/config combinations) re-verified bit-exact. D-Stress cycles
fall a further 1.66-2.00x on top of DEC-0015 (~4x combined vs the
original byte-level baseline). But the cache's real Fmax cost is much
steeper than DEC-0015's own: N_SLOTS=2 (the recommended default,
DEC-0014) drops from 133.58 to 87.72 MHz (-34%, margin over 80MHz
shrinks from +67% to +9.7%), and N_SLOTS=4 drops to 65.01 MHz - now
FAILING the 80MHz target it previously passed. Combined real
wall-clock speedup vs the original baseline: N=1 3.86x, N=2 2.45x
(both real net wins); N=4 is a real regression once its own now-failing
Fmax is honestly used, though N=4 was never the recommended
configuration.
N_SLOTS=2 remains the recommended default (DEC-0014 unaffected) with
a thinner but still real Fmax margin. Cache hit-detection pipelining
is flagged as concrete follow-up work if N_SLOTS>2 is ever needed with
the cache active - not attempted this round.
Logged: simulation/synthesis/timing/benchmark/decisions (DEC-0016)/
experiments (EXP-0016)/errors (ERR-0010)/development.log, ROADMAP.md
updated.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013xXuuRUWZScuo1DeYJxs3v
Implements optimization #1 from the final benchmark campaign's own
recommendation: exploit psram_controller.v's already-implemented
page-mode support (confirmed present by direct inspection) by
fetching multiple bytes per real backend transaction instead of one
at a time.
Root cause addressed: int8_memory_access.v (the byte-level backend
prefetch_engine.v originally sat on) already converts every 8-bit
logical request into a full 16-bit PSRAM word access internally
(mem_addr <= addr >> 1), discarding half of every word it already
paid for. prefetch_engine.v/memory_manager.v now speak
memory_interface.v's own 16-bit word protocol directly, bypassing
int8_memory_access.v entirely - which remains untouched, still frozen
V1 (§1/§34); V2 simply reuses the lower layer of the same frozen
chain instead of the byte-splitting layer on top of it, the same
"reuse what fits" precedent slot_mem_arbiter.v already set.
slot_mem_arbiter.v and neural_multiprocessor.v widened to match
(lb_n/ub_n added, master port wired directly to memory_interface.v).
Real, measured results: M4's own single-job testbench shows 49-56%
fewer cycles (166->84, 446->204, 728->322, all still bit-exact). The
full final-benchmark campaign (24/24 workload/config combinations)
re-verified bit-exact with D-Stress's real wall-clock time (cycles /
real POST-P&R Fmax) improving 2.24-2.37x across every N_SLOTS tested,
against a small real Fmax cost (unchanged at N=1, -6.2% at N=2, -1.2%
at N=4).
tb_neural_multiprocessor.v (M8) and tb_benchmark_suite.v (final
campaign) needed zero changes - both treat neural_multiprocessor.v as
a black box. Only tb_memory_manager.v (M4, rewired to skip
int8_memory_access.v) and tb_dataflow_core.v (M7, behavioral model
widened to word-level) needed updates.
The "real parallel scaling is flat beyond N_SLOTS=2" finding (DEC-0014)
still holds - this optimization made the shared PSRAM port more
efficient per transaction, not multi-ported - so N_SLOTS=2 remains
the recommended default.
Logged: simulation/synthesis/timing/benchmark/decisions (DEC-0015)/
experiments (EXP-0015)/development.log.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013xXuuRUWZScuo1DeYJxs3v
Post-M10, user-requested final benchmark campaign: 6 realistic
workloads (16-256 independent neurons in a shared-input dense-layer
shape, a random-seeded 2-layer network with real cross-node PSRAM
forwarding, and a 6-node 2-hop dependency diamond) x 4 concurrency
levels (N_SLOTS=1/2/4/8) through the real, full neural_multiprocessor
system (real V1 PSRAM chain, real slot_mem_arbiter). 24/24 runs PASS
bit-exact against a software golden model (11,520 individual neuron/
node checks, zero mismatches).
Three real bugs found and fixed during the campaign itself (ERR-0009):
1. neural_director.v (M5) had a real RTL bug at N_SLOTS=1
($clog2(1)=0 makes a replication expression illegal) - never
caught because M5-M10 only ever tested N_SLOTS=2/4/8. Fixed with
a width-agnostic '0 literal; M5's own testbench re-verified
unaffected.
2/3. Two testbench sizing bugs in tb_benchmark_suite.v itself
(psram_model DEPTH too small for the Large workload's address
range; N_NODES too small for the Stress workload's node-id range,
causing a real deadlock via node-id wraparound colliding with an
already-DISPATCHED node - a real, honest consequence of DEC-0008's
own "no node-slot reclamation" design choice).
Headline finding: real parallel scaling is essentially flat beyond
N_SLOTS=2 - the single shared PSRAM port saturates at ~91%
utilization regardless of slot count, so memory-bound workloads gain
only 1.05-1.06x real speedup from N=1 to N=8. Once real POST-P&R Fmax
degradation is also factored in, N_SLOTS=4 is measurably 21% SLOWER
in real wall-clock time than N_SLOTS=1 for the largest workload
tested. N_SLOTS=2 is recommended as the default (DEC-0014,
superseding DEC-0012's resource-only "N_SLOTS=8 ceiling" framing for
general use).
Full 21-section report (every number classified THEORETICAL/
SIMULATED/POST-P&R MEASURED/DERIVED, per the user's own
methodology requirements): hardware/v2/docs/benchmarks/
final-benchmark.md
Logged: simulation/synthesis/timing/benchmark/decisions (DEC-0014)/
experiments (EXP-0014)/errors (ERR-0009)/development.log,
ROADMAP.md updated.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013xXuuRUWZScuo1DeYJxs3v
Final milestone of docs/v2-description.md's §33 roadmap, scoped
exactly to its own mandate: optimize only on data already gathered in
M1-M9, across the pipeline/P_IN/processor-count/scheduling/memory
axes - no speculative new features.
Three concrete, data-driven results:
1. N_SLOTS=8 (numero processor axis): real synthesis + nextpnr-ecp5
P&R for dataflow_core at N_SLOTS=8, extending M7's N_SLOTS=2/4
sweep to the real DSP ceiling DEC-0005 predicted. 92.63 MHz
POST-P&R, PASS at 80MHz, DSP 64/72 (88.9%). DEC-0012 recommends
N_SLOTS=8 as the practical ceiling for P_IN=8 on the
LFE5U-45F-8BG381.
2. ACC_WIDTH 24 vs 32 (pipeline axis): a real 6-seed nextpnr-ecp5
placement sweep (reusing already-synthesized netlists, no new
synthesis needed) resolves EXP-0002's single-seed
inconclusiveness. ACC_WIDTH=24 wins on both mean Fmax (+6.2%,
180.71 vs 170.12 MHz) and seed-to-seed variance (~3.4x tighter),
on top of its already-known resource advantage. DEC-0013
recommends ACC_WIDTH=24 as the new default.
3. Stall %/utilization (scheduling/memory axes): testbench-only cycle
counters added to tb_neural_multiprocessor.v (no RTL touched)
close DEC-0011's deferred measurement gap with real data - shared
PSRAM port 81.7% utilized, slot 0 95.2%, slot 1 65.2%, no
conclusive evidence of harmful fixed-priority starvation at this
scale.
The 10-milestone V2 roadmap (docs/v2-description.md §33) is now
complete end-to-end: real Verilator simulation, real Yosys synthesis,
real nextpnr-ecp5 place & route for every milestone, fully logged
(EXP-0001..EXP-0013, DEC-0001..DEC-0013, ERR-0001..ERR-0008) with no
invented results (§30) and V1 kept frozen and untouched throughout
(§1/§34).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013xXuuRUWZScuo1DeYJxs3v
Consolidates real, already-measured data from hardware/v1/ (frozen,
pre-certified) and V2's own M1-M8 logs into the §32-mandated
comparison table, on an apples-to-apples basis: both full systems
(V1's spi_neuron_top post_fix_verify vs V2's neural_multiprocessor
N_SLOTS=2), both PARALLEL=8/P_IN=8, both using the real unmodified V1
PSRAM backend.
Headline, all real measurements: V2 full-system Fmax 142.45 MHz
POST-P&R (PASS at 80MHz) vs V1's 68.65 MHz (FAIL at 80MHz); 166 vs 209
real simulated cycles for one neuron's 8-input dot product through the
same real PSRAM chain (2.6x wall-clock speedup); peak MAC/cycle 16
(N_SLOTS=2 concurrent slots, real contention already demonstrated in
EXP-0009) vs V1's 8 (single sequential core); lower LUT/FF despite V2
already including full dependency-graph scheduling that V1 has none
of.
9 of the table's 12 rows carry real sourced numbers; stall %/memory
utilization/processor utilization are reported as NOT MEASURED rather
than approximated (DEC-0011) - a real number needs dedicated
cycle-accounting instrumentation neither system has had built for it
yet, and approximating from partial data would violate §30's "no
invented results" rule. Deferred to M10, which needs exactly this
data to decide what to optimize.
No new RTL this milestone - pure data consolidation, logged as
EXP-0010.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013xXuuRUWZScuo1DeYJxs3v
micheleandClaude Sonnet 5