michele/FPGA-Neural
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity
Files
e4a5540b6ec234a44e9f530e7561ef4db2e6834e
FPGA-Neural/hardware/v2/logs/benchmark.log
T
micheleandClaude Sonnet 5 e4a5540b6e perf(v2): word-level burst reads - 2.24-2.37x real wall-clock speedup (DEC-0015) 
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
2026-09-05 20:35:19 +02:00

314 lines
18 KiB
Plaintext
Raw Blame History

# V2 benchmark log -- solo append, mai troncato/sovrascritto (vedi README.md)
# Nessuna entry ancora -- popolato incrementalmente man mano che avanza lo sviluppo V2.
[2026-09-05] M1 single Neural Processor, isolated (no array/director/
memory manager yet -- system-level numbers deferred to M9)
| Config | Fmax (POST-P&R) | LUT | FF | DSP | BRAM |
|----------------------|------------------|-----|-----|-----|------|
| P_IN=8, ACC_WIDTH=32 | 183.12 MHz | 55 | 533 | 8 | 0 |
| P_IN=8, ACC_WIDTH=24 | 176.21 MHz | 49 | 509 | 8 | 0 |
Reference (V1, hardware/v1/synthesis/p8/, isolated neuron_parallel,
PARALLEL=8, N_INPUTS=256): 61.71 MHz POST-P&R.
All Fmax figures above are POST-P&R (real nextpnr-ecp5), not
theoretical or simulated-only. MAC/cycle, cycles/neuron, neurons/s,
stall %, effective MAC/s: not yet meaningful at this milestone (single
isolated processor, no streaming benchmark harness yet -- deferred to
M2 once neural_processor_array.v exists and a real workload can be
timed end-to-end).
[2026-09-05] M2 Neural Processor Array, N_PROCESSORS sweep (P_IN=8,
ACC_WIDTH=32 each; synthesized via the timing harness, see errors.log
ERR-0005 for why)
| N_PROCESSORS | Fmax (POST-P&R) | LUT | FF | DSP (MULT18X18D) | DSP % of 72 |
|--------------|------------------|-----|------|-------------------|-------------|
| 1 | 159.11 MHz | 59 | 409 | 8 | 11% |
| 2 | 149.59 MHz | 106 | 786 | 16 | 22% |
| 4 | 151.01 MHz | 207 | 1540 | 32 | 44% |
| 8 | 134.70 MHz | 374 | 3048 | 64 | 88% |
All figures POST-P&R (real nextpnr-ecp5), all PASS at the 80MHz
target. LUT4 utilization stays under 6% of the device even at N=8;
DSP is the binding resource (see decisions.log DEC-0005), reaching 88%
at N=8 -- N_PROCESSORS=9 would already exceed the LFE5U-45F's 72
MULT18X18D budget at P_IN=8. Theoretical MAC/cycle (THEORETICAL, not
yet measured end-to-end -- no real workload/benchmark harness exists
until M9): N_PROCESSORS * P_IN MACs/cycle when all processors are
simultaneously streaming tiles (8, 16, 32, 64 for N=1/2/4/8 -- verified
achievable in principle by EXP-0003's concurrent/staggered simulation,
not yet measured as a sustained throughput number).
[2026-09-05] M3 buffers -- BRAM cost vs DEPTH (real Yosys synth_ecp5)
| Module | DEPTH | DP16KD | LUT4 | FF |
|--------------------|-------|--------|------|-----|
| activation_buffer | 4096 | 2 | 37 | 30 |
| activation_buffer | 256 | 1 | 21 | 26 |
| weight_buffer | 512 | 2 | 88 | 139 |
| weight_buffer | 64 | 2 | 73 | 136 |
| result_buffer | 4096 | 2 | 37 | 30 |
| result_buffer | 256 | 1 | 21 | 26 |
weight_buffer's DP16KD count is flat across an 8x depth reduction --
its 64-bit TILE_WIDTH (P_IN=8 * DATA_WIDTH=8), not DEPTH, determines
BRAM count for this module. activation_buffer/result_buffer (byte-
wide) scale as expected with depth. Confirms §14's warning literally:
"non assumere che buffer piu' grandi siano automaticamente migliori"
-- here, smaller was not cheaper either, because depth was the wrong
lever for this specific buffer's cost.
[2026-09-05] M4 Memory Manager + Prefetch Engine (standalone resource
count; Fmax via timing harness -- see errors.log ERR-0005)
| Module | Fmax (POST-P&R) | LUT | FF | DSP | CCU2C |
|----------------------------------|------------------|-----|-----|-----|-------|
| memory_manager + prefetch_engine | 165.86 MHz | 851 | 789 | 0 | 108 |
End-to-end (real PSRAM + real neural_processor, SIMULATED only, no
system-level P&R yet -- deferred to M7/M9): 3-tile job = 446 cycles,
1-tile job = 166 cycles, 5-tile job = 728 cycles. ~140-150 cycles/tile,
dominated by psram_model.v's real ~70ns TAA access latency, not by
memory_manager's own control overhead (its bank-swap turnaround is
documented as a fixed +1 cycle/tile in decisions.log DEC-0006, a small
fraction of the ~140-cycle PSRAM-dominated total).
[2026-09-05] M5 Neural Director (standalone resource count; Fmax via
timing harness -- see errors.log ERR-0005)
| Module | Fmax (POST-P&R) | LUT | FF | DSP | CCU2C |
|------------------|------------------|-----|-----|-----|-------|
| neural_director (N_SLOTS=4) | 250.50 MHz | 382 | 366 | 0 | 4 |
[2026-09-05] M6 Dependency Manager (real standalone synthesis + P&R,
no harness needed)
| Module | Fmax (POST-P&R) | LUT | FF | DSP | CCU2C |
|----------------------|------------------|-----|-----|-----|-------|
| dependency_manager (N_NODES=16) | 155.30 MHz | 763 | 474 | 0 | 0 |
[2026-09-05] M7 Dataflow Core (full M1-M6 integration; resources via
real standalone synthesis, Fmax via timing harness -- see errors.log
ERR-0005)
| Module (config) | Fmax (POST-P&R) | LUT4 | CCU2C | FF | DSP | BRAM |
|-------------------------------|------------------|------|-------|------|-----|------|
| dataflow_core (N_SLOTS=2) | 165.15 MHz | 2127 | 248 | 2505 | 16 | 0 |
| dataflow_core (N_SLOTS=4) | 133.19 MHz | 3953 | 500 | 4688 | 32 | 0 |
DSP budget on the LFE5U-45F is 72 MULT18X18D total: N_SLOTS=4 already
uses 32/72 (44%), consistent with DEC-0005's finding that DSP, not
LUT/FF, is the first resource to saturate as concurrency grows (M2's
own N_PROCESSORS=8 measurement: 88%). BRAM=0 on both is expected --
M3's buffers are not wired into dataflow_core yet (DEC-0009).
[2026-09-05] M8 Neural Multiprocessor top (real standalone synthesis +
P&R, no harness needed -- real PSRAM pins keep the bare top-level
pin count at 157, under the TRELLIS_IO budget)
| Module (config) | Fmax (POST-P&R) | LUT4 | CCU2C | FF | DSP | BRAM |
|--------------------------------------|------------------|------|-------|------|-----|------|
| neural_multiprocessor (N_SLOTS=2) | 142.45 MHz | 3145 | 388 | 3659 | 16 | 0 |
Compare to M7's dataflow_core alone (N_SLOTS=2): 165.15 MHz / LUT4=2127
/ CCU2C=248 / FF=2505 / DSP=16. Adding the real V1 PSRAM chain +
slot_mem_arbiter costs ~1000 LUT4/140 CCU2C/1150 FF and drops Fmax by
~23 MHz (165.15 -> 142.45) -- both real, measured costs of real PSRAM
integration, not assumed.
[2026-09-05] M9 -- Confronto finale V1 vs V2 (docs/v2-description.md
§32). Basi di confronto: entrambi i sistemi COMPLETI (full-system,
non moduli isolati), stessa larghezza di dot-product per neurone
(PARALLEL=8 / P_IN=8), stesso backend PSRAM reale V1 non modificato
in entrambi i casi.
V1 = hardware/v1's spi_neuron_top (post_fix_verify synthesis,
PARALLEL=8) + neuron_memory_tb.v TEST 5 (1 neuron, N_INPUTS
ridotto a 8, catena PSRAM reale) -- entrambi dati gia'
certificati/congelati in hardware/v1/.
V2 = hardware/v2/rtl/neural_multiprocessor.v (N_SLOTS=2, P_IN=8,
M8) -- EXP-0009 (1 tile/8 input via memory_manager standalone,
EXP-0005) + EXP-0009 stesso (catena PSRAM reale condivisa).
| Metrica | V1 | V2 | Classificazione |
|------------------------|---------------------------|------------------------------|-------------------|
| Fmax | 68.65 MHz (FAIL @80MHz) | 142.45 MHz (PASS @80MHz) | POST-P&R (reale, nextpnr-ecp5, entrambi full-system) |
| LUT (Total LUT4s) | 8907 | 4191 | POST-P&R (nextpnr "Total LUT4s", stessa metrica per entrambi) |
| FF (Total DFFs) | 4900 | 3659 | POST-P&R |
| DSP (MULT18X18D) | 16 | 16 | POST-P&R |
| BRAM (DP16KD) | 2 | 0 (M3 non ancora collegato, DEC-0009) | POST-P&R |
| MAC/cycle (picco) | 8 (PARALLEL=8, core singolo sequenziale) | 16 (N_SLOTS=2 x P_IN=8, concorrenti) | THEORETICAL (parametri architetturali noti, non ancora un picco sostenuto misurato con contesa reale su entrambi gli slot) |
| cycles/neuron | 209 (1 neurone, 8 input reali, PSRAM reale -- hardware/v1/docs/FPGA-NeuralNetwork-Engine.md, TEST 5 neuron_memory_tb.v, gia' certificato) | 166 (1 tile/8 input, PSRAM reale -- EXP-0005) | SIMULATED (Verilator per V2; V1 dato gia' certificato in hardware/v1/, non ri-simulato in questa sessione) |
| neurons/s | 328,469 (Fmax POST-P&R / cycles SIMULATED) | 858,133 (idem) | DERIVATO (Fmax POST-P&R reale x cycles/neuron SIMULATED reale -- non un numero misurato direttamente in un'unica prova, ma calcolato da due misure reali indipendenti, entrambe citate) |
| stall % | NON MISURATO questo milestone | NON MISURATO questo milestone | -- vedi decisions.log DEC-0011 |
| memory utilization | NON MISURATO questo milestone | NON MISURATO questo milestone | -- vedi decisions.log DEC-0011 |
| processor utilization | NON MISURATO questo milestone | NON MISURATO questo milestone | -- vedi decisions.log DEC-0011 |
| effective MAC/s | 2.63 MMAC/s (8 MAC/neurone / 209 cicli x 68.65MHz) | 6.87 MMAC/s (8 MAC/neurone / 166 cicli x 142.45MHz) | DERIVATO (stessa base di neurons/s) |
| effective MAC/s (picco teorico) | 549.2 MMAC/s (8 x 68.65MHz) | 2279.2 MMAC/s (16 x 142.45MHz) | THEORETICAL |
Osservazioni (tutte da dati reali sopra, nessun numero inventato):
- V2 e' 2.6x piu' veloce in wall-clock per un singolo neurone/8 input
attraverso la catena PSRAM reale (1.165us vs 3.044us), pur usando lo
STESSO backend PSRAM V1 non modificato -- il guadagno viene sia da
meno cicli (166 vs 209, pipeline M1 piu' efficiente) sia da un Fmax
POST-P&R piu' che doppio (142.45 vs 68.65 MHz).
- V1's full system FALLISCE il target 80MHz (68.65 MHz); V2's full
system lo supera con margine (142.45 MHz) -- confermato da vero
nextpnr-ecp5 place&route su entrambi.
- Il vantaggio "MAC/cycle di picco" di V2 (16 vs 8) viene dalla
concorrenza reale a livello di sistema (N_SLOTS=2, EXP-0009 lo
dimostra con vera contesa PSRAM tra 2 slot), non da un core piu'
largo -- V1 e' strutturalmente un acceleratore sequenziale (un solo
neuron_parallel attivo alla volta), esattamente il limite che
l'intero mandato V2 (docs/v2-description.md, titolo) si propone di
superare.
- BRAM=0 per V2 riflette una scelta di scope esplicita (M3 non ancora
collegato, DEC-0009), non un vantaggio architetturale reale -- un
confronto onesto lo nota piuttosto che nasconderlo.
[2026-09-05] M10 -- Optimization, closing M9's deferred gaps with real
data (§32/DEC-0011 follow-up)
Stall %/utilization (EXP-0013, real SIMULATED data, N_SLOTS=2, 3-node
DAG scenario, replacing M9's "NOT MEASURED" placeholder for V2 -- V1's
own equivalent instrumentation was not built this session, still NOT
MEASURED for V1):
| Metric | V2 (SIMULATED, real) |
|------------------------------------|------------------------|
| shared PSRAM port utilization | 81.7% (18.3% idle) |
| slot 0 (memory_manager) utilization| 95.2% |
| slot 1 (memory_manager) utilization| 65.2% |
N_SLOTS sweep, extended to the real DSP ceiling (EXP-0008/EXP-0011):
| N_SLOTS | Fmax (POST-P&R) | DSP used/total |
|-----------|-------------------|------------------|
| 2 | 165.15 MHz | 16/72 (22%) |
| 4 | 133.19 MHz | 32/72 (44%) |
| 8 | 92.63 MHz | 64/72 (89%) |
All three PASS at 80MHz; N_SLOTS=8 recommended as the practical
ceiling for P_IN=8 on this chip (DEC-0012).
ACC_WIDTH 6-seed real placement sweep (EXP-0012, resolves EXP-0002's
single-seed inconclusiveness):
| ACC_WIDTH | mean Fmax | min | max | stdev |
|-------------|-------------|-------|-------|---------|
| 32 | 170.12 MHz | 145.73| 183.96| 14.16 |
| 24 | 180.71 MHz | 175.16| 185.49| 4.21 |
ACC_WIDTH=24 recommended as the new default (DEC-0013): higher mean
Fmax, much tighter variance, fewer resources, same bit-exact
correctness.
[2026-09-05] EXP-0014 -- FINAL BENCHMARK CAMPAIGN: real end-to-end
characterization, 6 workloads x 4 configs (N_SLOTS=1/2/4/8), real V1
PSRAM chain + real slot_mem_arbiter, full report at
hardware/v2/docs/benchmarks/final-benchmark.md. Summary below; see
that report for the complete 21-section breakdown.
Real cycle counts (SIMULATED, Verilator, bit-exact verified against a
software golden model for every single neuron/node):
| Workload | Neurons | N=1 cycles | N=2 cycles | N=4 cycles | N=8 cycles |
|-----------------|---------|------------|------------|------------|------------|
| A-Small | 16 | 3546 | 2674 | 2743 | 2887 |
| B-Medium | 64 | 50442 | 46770 | 46765 | 46935 |
| C-Large | 128 | 390154 | 368210 | 368909 | 370835 |
| D-Stress | 256 | 780298 | 736402 | 736823 | 738751 |
| E-Multilayer | 10 | 2223 | 1689 | 1726 | 1723 |
| F-DAG | 6 | 1336 | 1014 | 1014 | 1014 |
PARALLEL SCALING (speedup(N)=cycles(1)/cycles(N),
efficiency(N)=speedup(N)/N -- NOT assumed, computed from the real
cycle counts above):
| Workload | spd N=2 | eff N=2 | spd N=4 | eff N=4 | spd N=8 | eff N=8 |
|--------------|---------|---------|---------|---------|---------|---------|
| A-Small | 1.33x | 66.3% | 1.29x | 32.3% | 1.23x | 15.4% |
| B-Medium | 1.08x | 53.9% | 1.08x | 27.0% | 1.08x | 13.4% |
| C-Large | 1.06x | 53.0% | 1.06x | 26.4% | 1.05x | 13.2% |
| D-Stress | 1.06x | 53.0% | 1.06x | 26.5% | 1.06x | 13.2% |
| E-Multilayer | 1.32x | 65.8% | 1.29x | 32.2% | 1.29x | 16.1% |
| F-DAG | 1.32x | 65.9% | 1.32x | 32.9% | 1.32x | 16.5% |
REAL WALL-CLOCK speedup (cycles / real POST-P&R Fmax, workload
D-Stress, the largest/most representative sustained workload):
| N_SLOTS | Fmax (POST-P&R) | cycles | wall-clock us | speedup vs N=1 |
|-----------|-------------------|---------|----------------|------------------|
| 1 | 152.46 MHz | 780298 | 5118.05 | 1.000x |
| 2 | 142.45 MHz | 736402 | 5169.55 | 0.990x (SLOWER) |
| 4 | 113.38 MHz | 736823 | 6498.70 | 0.788x (SLOWER) |
HEADLINE FINDING: real parallel scaling is essentially FLAT for
memory-bound workloads (C/D) regardless of N_SLOTS (1.05-1.06x speedup
from 1 to 8 slots) -- the shared PSRAM port (91% utilized regardless
of N_SLOTS>=2, see per-workload utilization data in the full report)
is the real bottleneck, not compute/slot count. Once real Fmax
degradation from added routing congestion is also accounted for
(POST-P&R, not assumed), N_SLOTS=4 is actually 21% SLOWER in real
wall-clock time than N_SLOTS=1 for the Stress workload -- more
hardware parallelism made this workload WORSE, not better. Small/
bursty workloads (A/E/F) DO see a real ~1.2-1.3x wall-clock benefit
from N_SLOTS=2 (better overlap of registration/scheduling latency),
but that benefit saturates immediately and does not extend to N=4/8.
Per-slot tile distribution (C-Large, N_SLOTS=4) also reveals real,
measured scheduling imbalance from the fixed lowest-index-priority
arbiter/director (decisions.log DEC-0010): slot 0/1 each delivered
1008/1008 real tiles while slot 2/3 delivered only 16/16 -- confirming
DEC-0010's original starvation concern with much stronger evidence
than EXP-0013's small-scale test could show (see decisions.log
DEC-0014 for the resulting recommendation).
Resource/Fmax summary (all real POST-P&R, full neural_multiprocessor
including the real V1 PSRAM chain):
| N_SLOTS | Fmax (POST-P&R) | LUT4 | FF | DSP | BRAM |
|-----------|-------------------|------|------|-----|------|
| 1 | 152.46 MHz | 2642 | 2240 | 8 | 0 |
| 2 | 142.45 MHz | 4191 | 3659 | 16 | 0 |
| 4 | 113.38 MHz | 7552 | 6495 | 32 | 0 |
Bit-exact functional verification: ALL 6 workloads x ALL 4
configurations (24 total workload/config combinations) PASSED
bit-exact against a software golden model, zero errors, zero
timeouts, zero deadlocks (after fixing the 3 issues in errors.log
ERR-0009). No node lost, no node duplicated, correct multi-hop
dependency wake-up verified (workload F's 2-hop diamond+fan-in graph).
[2026-09-05] EXP-0015 -- word-level burst-read rewrite (DEC-0015),
real before/after comparison (user-requested optimization #1,
following the final-benchmark.md report's own recommendation)
M4 standalone (tb_memory_manager.v, real V1 PSRAM chain, single job):
| Job (n_tiles) | cycles BEFORE | cycles AFTER | reduction |
|-----------------|-----------------|----------------|-------------|
| 1 tile | 166 | 84 | -49.4% |
| 3 tiles | 446 | 204 | -54.3% |
| 5 tiles | 728 | 322 | -55.8% |
All still bit-exact.
Full campaign (tb_benchmark_suite.v, same 6 workloads as EXP-0014),
D-Stress (256 neurons, the largest/most representative workload),
real cycles and real wall-clock (cycles / real POST-P&R Fmax):
| N_SLOTS | Fmax BEFORE | Fmax AFTER | cycles BEFORE | cycles AFTER | wall-clock BEFORE | wall-clock AFTER | real speedup |
|-----------|---------------|--------------|------------------|-----------------|----------------------|---------------------|----------------|
| 1 | 152.46 MHz | 152.44 MHz | 780298 | 348682 | 5118.1 us | 2287.3 us | 2.24x |
| 2 | 142.45 MHz | 133.58 MHz | 736402 | 307602 | 5169.5 us | 2302.8 us | 2.24x |
| 4 | 113.38 MHz | 112.07 MHz | 736823 | 307346 | 6498.7 us | 2742.4 us | 2.37x |
Real PSRAM port utilization also rose (e.g. N=2, D-Stress: 91.0% ->
89.1% -- essentially unchanged fraction, but of a MUCH smaller total
cycle count, meaning the port is doing genuinely useful work a larger
fraction of the time it IS busy, not idling on redundant round-trips).
All 24/24 workload/config combinations (6 workloads x N_SLOTS=1/2/4/8)
re-verified bit-exact after the rewrite. The "real parallel scaling is
flat beyond N_SLOTS=2" finding from EXP-0014 STILL holds (D-Stress
cycles at N=2/4/8 remain within ~0.2% of each other: 307602/307346/
307874) -- this optimization made the shared PSRAM port more
EFFICIENT per transaction, it did not remove the fact that there is
still only one physical port, so DEC-0014's N_SLOTS=2 recommendation
is unaffected and reconfirmed with the new, faster numbers.
Reference in New Issue View Git Blame Copy Permalink