Performance

Measured on our silicon, not a simulator.

Every figure below comes from our own evaluation boards running production silicon, through the PeakStack compiler. We publish the model, precision, and batch — and we say "up to" where the number is a ceiling, not a guarantee.

Classification throughput

Images per second, INT8, batch 1.

Sustained throughput on NP-1, single chip. Smaller models approach memory-bound territory; larger models show the compute array.

MobileNetV3-Large

up to 12,000 img/s

EfficientNet-B0

up to 8,600 img/s

ResNet-50

up to 4,800 img/s

ViT-Base/16

up to 900 img/s

224×224 inputs. Figures are ceilings at nominal voltage; sustained performance under thermal constraints is reported in the datasheet.

Reference results

A representative set, across tasks.

The models our customers actually run — detection, tracking, and on-device language — not just classification benchmarks.

NP-1 · measured resultsPeakStack 2.1 · batch 1

Workload	Model	Precision	Result	Notes
Classification	ResNet-50	INT8	4,800 img/s	224×224, batch 1
Classification	MobileNetV3-Large	INT8	12,000 img/s	224×224, batch 1
Detection	YOLO-class detector	INT8	120 fps	1080p, end-to-end
Detection	CenterNet-style detector	INT8	240 fps	720p, end-to-end
Vision transformer	ViT-Base/16	INT8	900 img/s	224×224, batch 1
Embedding	Face / person embedding net	FP16	1,600 img/s	112×112, batch 1
Language	1.5B-parameter small LLM	INT4	40 tokens/s	single chip, decode

Detection frame rates are end-to-end, including pre- and post-processing on the host. Token rates are decode-only, measured on a 1.5B-parameter model.

Detection & tracking throughput

End-to-end frame rates, not just the backbone.

Detection numbers include pre- and post-processing on the host, because that's the number your pipeline actually runs at. Measured INT8 on NP-1, single chip, batch 1.

CenterNet-style · 720p

240 fps

YOLO-class · 1080p

120 fps

YOLO-class · 4K

30 fps

Segmentation · 512×512

75 fps

Resolution is the input frame. Post-processing (NMS, decoding) runs on the host CPU of the evaluation board; figures are end-to-end.

Accuracy

Quantization that doesn't cost you points.

Every INT8 and INT4 result on this page is verified against its FP16 reference on our calibration set. The deltas below are representative of typical workloads — your model will differ, which is why PeakStack reports per-layer error before you commit.

Quantization accuracy · representative deltasPeakStack 2.1

Workload	INT8 vs FP16	INT4 vs FP16
Image classification (ResNet-50)	-0.1 pt top-1	-0.6 pt top-1
Detection (YOLO-class)	-0.2 mAP	-0.9 mAP
Embedding (face / person)	within noise	-0.3% TAR @ 1e-4 FAR
Small LLM (1.5B, perplexity)	—	+0.4 perplexity

Deltas are representative across our internal calibration set, not a guarantee for your dataset. PeakStack's per-layer error report flags any layer outside your tolerance and holds it at higher precision.

Latency

Where it matters most: the control loop.

Throughput tells you how much you can run. Latency tells you whether the robot reacts in time. NP-1 keeps the whole inference path under a millisecond for the workloads below.

Input → tile

0.3 ms

Sensor to first compute, DMA and pre-processing.

Core inference

0.4 ms

Detection backbone, INT8, 1080p single frame.

Output → host

0.2 ms

Results streamed to host memory, no polling.

End-to-end

<1 ms

Fixed timing budget for a full detection pass.

Efficiency

Doing more with less heat.

Raw TOPS are easy to print on a slide; sustained TOPS per watt is what a battery cares about. Our typical efficiency is reported at sustained, thermally-constrained load — not burst.

Sustained efficiency

5.0 TOPS/W

INT8, typical, over a one-minute sustained run.

Idle draw

<0.8 W

Weights resident in SRAM, ready to wake.

Thermal envelope

-40–105 °C

Junction range, full datasheet spec.

Wake-to-infer

<2 ms

From idle to first result, cold cache.

Methodology

How we measure, so you can reproduce.

Benchmarks without methodology are marketing. Ours are reproducible on the same evaluation board you'd receive as a sample customer.

→ Hardware: NP-1 evaluation board, production silicon, nominal voltage.
→ Software: PeakStack 2.1 compiler, calibration-aware quantization, default policies.
→ Measurement: hardware-in-the-loop, warmed up, median over 1,000 iterations.
→ Accuracy: every INT8/INT4 result is verified against its FP16 reference to confirm no loss beyond stated tolerance.
→ Power: measured at the board input rail, not the die, over a one-minute sustained run with a fixed workload.
→ Thermal: sustained figures are taken after a soak to steady-state junction temperature; burst figures at a cold die, which is why they're labeled "up to."
→ Calibration: each workload is quantized on a fixed, documented calibration set; the same set and the same PeakStack version reproduce the same numbers.
→ "Up to" figures are ceilings; "typical" figures are what you should plan a design around.