Running a 30B MoE LLM on CPU at 30-50+ tok/s with a 256K Context Window via Hybrid Mamba-2 SSM, L3 NTA Slicing, AVX-512 VNNI Tile Repacking & 2MB HugePages
-
Updated
Nov 1, 2026 - Python
#
zen5
Here are 10 public repositories matching this topic...
PBO tuning and automated Curve Optimizer testing for AMD Ryzen 5000/7000/9000: per-core monitoring, y-cruncher stress tests with crash recovery, Single Core Boost and per-core voltage alignment.
amd winforms stress-test ryzen hardware-monitor overclocking zen3 undervolting curve-optimizer zen4 zen5 curve-shaper precision-boost-overdrive
-
Updated
Nov 1, 2026 - C#
Reverse-engineered AMD Ryzen SMU telemetry and control tools for Linux Zen 3 Vermeer + Zen 5 Granite Ridge.
linux amd reverse-engineering telemetry hardware-monitoring ryzen smu overclocking vermeer zen3 5600x undervolting zen5 9800x3d 9950x3d granite-ridge
-
Updated
Sep 11, 2026 - Python
Per-core Curve Optimizer tuning for Zen 5 desktop CPUs on NixOS
-
Updated
Oct 31, 2026 - Rust
让消费级显卡本地跑动 MoE 架构 30B~35B 大模型（以 Q4_K_M / IQ3_S 这类 4-bit 级量化为基准；Q2/IQ2 等低比特量化质量太低，不在本项目范围）：显存与内存混合承担（非专家权重上显存、占大头的专家权重留内存），速度远超纯 CPU；没有独显或显存不足时纯 CPU 兜底。免装 CUDA/Python/任何运行库，模型可替换。
windows amd nvidia moe quantization avx512 non-commercial expo memory-bandwidth ryzen ddr5 cpu-inference llama-cpp local-llm gguf openai-compatible qwen3 llm-tuning zen5 gpu-offload
-
Updated
Oct 26, 2026 - C++
Optimized Firefox ESR for Windows built from source with PGO/LTO for ZEN5 and AVX2 CPUs
-
Updated
Oct 28, 2026 - Python
AMD NPU Enabler (aie2p) for Strix Point/Krackan (Zen 5, Ryzen AI 7 350) on Ubuntu 26.04. Validated 4.3 TOPS / 68us latency. Includes amdxdna v0.11.0 & XRT v2.23.0.
-
Updated
May 6, 2026 - Python
AVX-512 IPv6 text-to-bytes parser in C for Zen 5 — ~13.5 cycles/address on random inputs, with :: compression and inet_pton validation.
c benchmark performance ipv6 optimization x86-64 ip-address simd vectorization avx512 ipv6-address simd-intrinsics ip-parser avx-512 zen4 zen5 ipv6-parser inet-pton avx512-vbmi avx512-vbmi2
-
Updated
Oct 6, 2026 - C
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.
self-hosted arm64 quantization avx512 inference-engine avx512-vnni openai-api cpu-inference cpp26 llama-cpp local-llm llm-inference qwen gguf speculative-decoding anthropic-api neoverse openai-compatible amd-epyc zen5
-
Updated
Oct 29, 2026 - C++
Add this topic to your repo
To associate your repository with the zen5 topic, visit your repo's landing page and select "manage topics."