NeuralPeak · Neural Processing Silicon Announcements & Changelog · 2026-08
News & Changelog

Dated, So You Can Tell What Changed

Announcements are dated; the toolchain is versioned. If a number moved, the changelog says when and why.

Toolchain PeakStack 3.3.0 Silicon NP-1 rev 1.2

1. Announcements

2026-08-14
Release

PeakStack 3.3 — FP4 sparsity and an ONNX Runtime bridge

2:4 structured-sparsity kernels now cover FP4, a new scheduler pass improves mesh placement for transformer workloads, and a drop-in ONNX Runtime execution provider lets existing pipelines target NP silicon without a rewrite.

2026-06-30
Product

NP-2 “Crest” reaches general availability

The mid-range part moves from sampling to GA across commercial and industrial temperature grades, with the full datasheet and reference design released to design partners.

2026-04-09
Milestone

NP-1 first silicon boots; bring-up complete

Engineering samples pass functional bring-up across the full die — PCIe, both memory PHYs, and all sixteen processing elements. Performance figures now come from hardware, not simulation.

2026-01-22
Milestone

NP-1 Apex tapes out

The flagship die is released to the foundry. Tape-out closes the architecture phase; the compiler team now works against a fixed target.

2025-10-08
Program

Design partner program opens; EVK-1 announced

First evaluation boards are committed to early design partners, and the mutual-NDA program is open for applications.

2. PeakStack changelog

v3.3.0
2026-08-14

FP4 sparsity · ONNX Runtime EP

  • 2:4 structured-sparsity kernels extended to FP4
  • New scheduler pass for 2-D mesh placement on transformer graphs
  • Drop-in ONNX Runtime execution provider (x86_64 Linux)
v3.2.1
2026-07-02

Bugfix release

  • Fixed LPDDR5X refresh scheduling under sustained 85 °C load
  • Fixed INT4 calibration regression in low-activation tail layers
v3.2.0
2026-06-30

Quantization & sparsity GA

  • FP4 / INT4 weight-only quantization out of preview
  • 2:4 structured sparsity out of preview
  • Per-layer energy and utilization counters in the runtime
v3.1.0
2026-03-18

Attention fusion

  • Attention op fusion; KV cache residency in global SRAM
  • NP-2 Crest target added
v3.0.0
2025-12-02

First public toolchain

  • ONNX importer, graph partitioner, and NP-1 target
  • Cycle-accurate simulator (peaksim) for pre-silicon work

3. Silicon revisions

NP-1 · rev 1.2
2026-08

Engineering silicon

  • Yield and characterization fixes; current benchmark baseline
NP-1 · rev 1.0
2026-01

First silicon

  • Full functional bring-up achieved 2026-04-09
NP-2 · rev 1.0
2026-06

General availability

  • Commercial and industrial temperature grades