NeuralPeak · Neural Processing Silicon Benchmarks · PeakStack 3.2 · 2026-08
Performance / Measured on silicon

Benchmarks — Numbers with Conditions Attached

Every figure here is measured, not modeled, and carries the conditions that produced it. We publish methodology so the numbers can be reproduced or contested — the two things a benchmark should survive.

Toolchain PeakStack 3.2 Samples median of 1,000 runs Ambient 25 °C

1. Throughput vs. batch

ResNet-50, INT8. Throughput saturates as batch grows and the MAC array stays full; the knee is where latency starts to dominate the calculation.

ResNet-50 · INT8 · images/second

median of 1,000 · 25 °C · log-x axis
012k 24k36k 48k 12 48 1632 64 batch size NP-1 Apex NP-2 Crest
Figure 1. ResNet-50 INT8 throughput vs. batch size. NP-1 saturates at ~48,000 img/s (batch 64); NP-2 at ~12,000 img/s.

Latency vs. batch · NP-1

ResNet-50 · INT8 · microseconds · batch 1–64
0350 7001050 1400 12 48 1632 64 batch size · µs
Figure 2. NP-1 latency vs. batch (ResNet-50, INT8): 310 µs at batch 1 rising to 1,333 µs at batch 64.

2. Energy efficiency

TOPS per watt at nominal voltage, measured at full INT8 utilization. The low-power part wins this chart by design — the flagship trades efficiency headroom for raw throughput.

INT8 · TOPS/W · sustained

throughput ÷ measured rail power · typical envelope
NP-4 Ridge
6.7 TOPS/W
NP-2 Crest
6.4 TOPS/W
NP-1 Apex
6.0 TOPS/W
Figure 3. Sustained INT8 energy efficiency, TOPS per watt.

3. Reference results

A fixed set of models across all three parts, so one number maps to one configuration. FP4/INT4 rows use weight-only quantization unless noted.

WorkloadMetricNP-1 ApexNP-2 CrestNP-4 Ridge
ResNet-50INT8 · batch 1latency310 µs980 µs6.9 ms
ResNet-50INT8 · batch 64throughput48,000 img/s12,000 img/s1,800 img/s
YOLOv8nINT8 · batch 1throughput1,650 fps420 fps55 fps
Llama-3-8BINT4 · decodetokens/s180 tok/s55 tok/s8 tok/s
Llama-3-8BINT4 · prefilltokens/s12,400 tok/s3,900 tok/s620 tok/s
BERT-baseINT8 · batch 32sequences/s21,000 seq/s5,300 seq/s800 seq/s

NP-4 rows are pre-silicon targets. NP-1/NP-2 measured on engineering silicon.

4. Methodology

What was measured, on what, with which software. Reproducibility is the point.

Hardware under test

NP-1
Engineering silicon, rev 1.2, 85 W typical envelope, FC-BGA on reference board
NP-2
General availability, rev 1.0, 20 W typical envelope
NP-4
Pre-silicon — RTL simulation, not physical measurement

Software

Toolchain
PeakStack 3.2, default scheduling passes
Model source
ONNX export; FP4/INT4 via PeakStack quantizer (weight-only, no retrain)
Runtime
PeakStack runtime, host CPU running at stock clocks

Measurement protocol

Warm-up
100 iterations, discarded
Samples
1,000 iterations, median reported (not mean — rejects outliers)
Ambient
25 °C, forced-air reference thermal
Power
Measured at the core rail, not estimated from TDP

What we do not report

Peak MAC-rate
Not a workload; excluded
Unquantized FP32
Not a supported precision on the array; excluded
Extrapolated figures
Every number maps to a run; no scaled estimates
Interpretation note: these are internal measurements on our silicon and our toolchain. Your results will differ by model, precision, batch, and thermal design. If a figure looks wrong, the methodology above is the place to start — and we would rather you contest it than repeat it.