neoverse

Helium DPU, ET3000, ET3600, ET2500, ET1600

Educational guide for developers to understand Arm Neoverse processors and Armv9 architecture for server and cloud use cases

My brand new homepage.

llama.cpp build with a faster 8-bit format for CPUs (Arm and x86-64). Runs embeddings, LLMs and MoE models with the same tools and OpenAI-compatible API. Against llama.cpp Q8_0: up to 2× faster prompt processing and, on Arm, about half the memory. Loads BF16 GGUF directly. Linux, Windows, Docker.

Neoverse 系列项目的通用 UI 库

Arm64 LLM inference autotuner: measures quantization, threads and concurrency on your exact Arm CPU with Arm Performix evidence, then hands you the launch command.

Neoverse-based Labs for 2026 IEEE Workshop

2.12x faster IQ4_XS prefill on Arm Neoverse: the missing smmla repack kernel for llama.cpp. IQ4_XS is the smallest standard 4-bit GGUF and had an accelerated matmul path on Intel AMX but none on Arm. Measured stock-vs-patched on free Neoverse N2 CI.

Slash enterprise AI costs by routing simple LLM queries to right-sized local models running on Arm Neoverse CPUs.

ARMONIC is an Autonomous agentic performance optimizer for Arm64 cloud AI. Profiles with Arm Performix (APX), feeds telemetry to an LLM, auto-generates validated patches, and commits optimizations to git — zero human code changes around 98% on moderate work load and upto 93% on heavy workloads speedup on AWS Graviton.

C++20 AArch64 microbenchmark for instruction latency and throughput, compared against llvm-mca's Neoverse V2 model and Arm's published values

LLM server for AVX-512 (Zen 4/5, EPYC) and Arm Neoverse CPUs only. Ryzen 9800X3D prompt reading: 4.13x llama.cpp on Q8_0, 2.05x on Q4_K_M, same perplexity. Lossless speculative decoding, OpenAI and Anthropic API, one docker run.

Prove whether an Arm binary can actually use the chip's matrix hardware. Static ISA verifier + CI gate for AArch64 AI runtimes (SME/SME2, i8mm, dotprod).

First public reproducible KV-cache quantization tradeoff study on AWS Graviton4 CPU with PMU-level mechanism analysis. Arm Create: AI Optimization Challenge, Cloud AI track.

One command to prove how much faster your LLM runs on Arm with KleidiAI. Arm AI Optimization Challenge 2026 — Track 2 (Cloud AI).

int8 vector search for Arm that reaches SDOT and SMMLA. Every number is reproducible in CI on free Arm runners; the README lists the results and the two speed claims I withdrew.

{title} is a feature-rich third-party modification project for {Neoverse Cheats}.