[runwangdl]

What

Two of the five MLperf Tiny training fixtures were not the reference model. Regenerated with the exporters from pulp-platform/Onnx4Deeploy#24 and checked against the int8 reference graphs this repo already ships in Tests/Models/MLPerf/.

Benchmark	Reference (shipped int8 graph)	Before	After
KWS	DS-CNN, 1x49x13→12, 22,160 par	1x25x10, 16 ch, 2,732 par, 0.072 M MACs	1x49x10, 64 ch, 23,180 par, 2.657 M MACs
AD	Deep AE, 640→640, 10 Gemm, 265,106 par	128→128, 29,120 par	640→640, 10 Gemm, 267,928 par

DS-CNN also needed the stem's padding fixed: the reference stem is Conv2D(64,(10,4),strides=2,padding='same') and emits a 25x5 map where ours emitted 20x4. With that, MACs land on the published 2.657 M.

Autoencoder — the 640 is 128 log-mel bins × 5 stacked frames, not a single frame. Its 267,928 parameters against 265,106 initializer elements in the shipped int8 graph differ by exactly the BatchNorm affine parameters, which that graph has folded.

The autoencoder moves from L2 to L3

Not a tuning choice. 1046.6 KB of weights plus an equally large gradient-accumulation buffer is 2098 KB of graph-input tensors against GAP9's 1536 KB L2 — minimalloc fails even at --l2 1572864.

It is the only pure MLP among the five benchmarks, so every weight is used exactly once:

model	params	MACs	MACs/param
Autoencoder	268 K	0.26 M	0.99
DS-CNN-S	23 K	2.66 M	115
ResNet-8	78 K	12.5 M	160
MobileNetV1	214 K	7.49 M	35
CCT-2	284 K	35.3 M	124

That ratio is why it is the one benchmark whose parameters, not its activations, set the footprint.

Best measured config: L3 + promote + double-buffer at --l1 122000, cc_stack 4096 — 21.7M cycles per 4 mini-batches against 36.6M for L3 single-buffer (−40%). 122000 rather than the 128000 it used on L2 because an L3 model must leave L1 for the cluster stacks: 122000 + 4096 + 512*8 = 130192 < 131072. At 128000 gvsoc aborts with Allocation failed for allocator 2 (allocator 2 is cluster L1; allocator 1 is L2 — see pi_malloc.h).

DS-CNN keeps only its 128000 L1 entry: at 8× the old fixture it has nothing left to gain from the smaller arena.

Also in here

• Drops two per-model comments that quoted cycle counts measured on the old fixtures (-38.5% / -33.7% figures taken at 0.48M and 0.80M cyc/step, against 10.2M for DS-CNN-S today). Stale numbers in comments are worse than none — code that is wrong fails, a comment that is wrong gets believed.
• Guards the per-mini-batch LOSSLIVE printf with POWER_MEASUREMENT. It sits inside the training loop between two GPIO pulses, so it lands inside the first→last GPIO-high window a PPK2 whole-loop average is taken over. It does not affect the GPIO-high-only average run_power_auto.sh reports — the previously recorded power figures are unaffected — but it does inflate the window average, which is the one energy-per-step has to be computed from. Every loss is still printed after the loop by check_losses.

Measured

GAP9 / gvsoc, Errors: 0 out of 4 throughout:

model	train cycles	opt cycles	deployment
DS-CNN-S	40,641,015	339,228	L2 single-buffer, --l1 128000
Autoencoder	21,717,385	9,854,185	L3 promote + DB, --l1 122000

Full gap9_tiled and training sweep run locally: L2 single-buffer 4/4, L3 single-buffer 9/9, L3 double-buffer 3/3, L3 promote+DB 4/4.

Regenerating the fixtures

Requires pulp-platform/Onnx4Deeploy#24:

Onnx4Deeploy.py -model DSCNN-S        -mode train -o <...>/DSCNN/dscnn_train
Onnx4Deeploy.py -model Autoencoder-AD -mode train -o <...>/Autoencoder/autoencoder_train

Note for anyone re-running these locally

PYTEST_XDIST_WORKER isolates the build directory but not the generated one — TEST_GAP9/Tests/... is shared across configurations. Running the same model at L3 and then at L2 leaves the L3 hex/ readfs artifacts behind and the L2 build picks them up. Wipe the model's generated directory when switching deployment level.

Relationship to #69

Independent. #69 makes the gradient accumulator 6× unrolled and 8-core, which is what takes the autoencoder from 31.8M to 21.7M; this PR is what makes it the right model in the first place.

🤖 Generated with Claude Code

https://claude.ai/code/session_012vb2GVs3qbiyLV1KkV2baN

Siracusa (second commit)

The fixture swap changes the autoencoder for every platform, not just GAP9. Siracusa keeps its own training lists and referenced the autoencoder in two of them, so siracusa-training-tiled-l2 (single- and double-buffer) would have broken.

Siracusa's L2 is 2 MB = 2,097,152 B. The autoencoder's graph-input tensors alone are 2,148,548 B — over the whole level by 51 KB before a single activation, so no --l2 value recovers it. minimalloc fails at the runner's 1,991,744 B capacity. It moves to L3_SINGLEBUFFER_TRAINING_MODELS there as well.

Also drops DS-CNN's 64000 L1 entry to match GAP9, and rewrites the comment justifying the 32 KB autoencoder double-buffer variant — it argued 32 KB forced "~8 of 55 ops" to 2-4 tiles so DB pipelining would fire, measured on the old 29,120-parameter fixture. The reference model is 205 nodes and no longer fits L2 at all.

Reference sites for these two fixtures, for anyone changing them again:

file	references	in CI
test_gap9_tiled_config.py	5	yes
test_siracusa_tiled_config.py	4	yes
benchmark_training.py	3	no — standalone script, L1 values already diverge from the test configs; left alone