Whitepapers & Resources

Read the silicon, not the slogan.

Whitepapers carry the doc number and revision, because that is how a datasheet ages. Documentation is gated to design partners; the abstracts below are public.

01

Whitepapers

WP-001Rev 1.22026-0828 pp

NP-1 Architecture Overview

The die floorplan, the processing-element microarchitecture, the three-tier memory hierarchy, and the 2-D mesh NoC. The document the datasheet points to.

Request access →
WP-002Rev 1.02026-0616 pp

Structured Sparsity in PeakStack

How 2:4 structured sparsity is represented, pruned, and decoded on-chip — and why it is deterministic rather than data-dependent.

Request access →
WP-003Rev 1.12026-0522 pp

INT4 / FP4 Quantization Methodology

Weight-only and activation quantization for the five on-array precisions, with the accuracy budget and the failure cases, stated plainly.

Request access →
WP-004Rev 2.02026-0434 pp

The PeakStack Compiler: from ONNX to Silicon

The lowering pipeline — graph partitioning, operator fusion, mesh placement, and memory liveness — that turns a model export into a resident graph.

Request access →
WP-005Rev 1.02026-0318 pp

Memory Subsystem and Dataflow

The bandwidth and latency tiers in detail, and the weight-stationary versus activation-stationary dataflow choices per workload class.

Request access →

Whitepapers are distributed under NDA to registered design partners. Apply for access.

02

PeakStack SDK

One toolchain across NP-1, NP-2, and NP-4. A model compiled for one runs on all three.

peakc — compiler

Lowers ONNX / PyTorch / TensorFlow Lite into a resident graph. Operator fusion, sparsity planning, mesh placement.

x86_64 · Linux

peakquant — quantizer

Weight-only and activation quantization to INT8 / INT4 / FP8 / FP4, with per-layer calibration and an accuracy report.

CLI · Python

peaksim — simulator

Cycle-accurate model of the die for pre-silicon development and NP-4 bring-up.

RTL-accurate

peakrt — runtime

Host runtime and kernel driver. Loads a compiled graph, manages memory, and drives the PCIe interface.

Linux · RTOS

ONNX Runtime EP

A drop-in ONNX Runtime execution provider so existing pipelines can target NP silicon without a rewrite.

interop

EVK bring-up image

Boot image, kernel, and example graphs for the evaluation kit.

board support

Request the SDK

SDK access is free for design partners and issued with the evaluation kit. We do not gate it behind a marketing funnel — but we do ask who is building with it.

Request access
03

Documentation

DocumentCodeRevStatus
NP-1 Apex datasheetNP1-DS-0011.2Engineering silicon
NP-2 Crest datasheetNP2-DS-0021.0GA
NP-4 Ridge datasheet (target)NP4-DS-0040.9Pre-silicon
EVK-1 evaluation kit guideEVK-1-UG1.1Available
NP-1 reference design (schematics + layout)NP1-RD-0011.0Design partners
PeakStack compiler referencePS-REF-0033.2Available