AI - Software Tools & Optimizations#
MoE Training Optimization with Primus
Learn how to speed up MoE training on AMD Instinct GPUs with Primus: a fused MoE megakernel, FP8/MXFP8 expert GEMMs, and training-at-scale tuning.
hipCIM 26.06: GPU-Accelerated Whole-Slide Image I/O on AMD Instinct™ GPUs
In this blog you will see how hipCIM 26.06 reads whole-slide images on AMD Instinct™ GPUs, and how fast and faithfully it decodes them.
Bringing Ultra-Fast, Interactive LLM Inference to AMD Instinct™ MI355X GPUs
Accelerate single-request LLM inference on AMD Instinct MI355X with megakernels, speculative decoding, and MXFP4 for 1.9-3.3x faster decode.
AMD Instinct MI455X vs MI355X: A Technical Look at the Advancing AI 2026 Inference Numbers
Explore the MI455X vs MI355X inference numbers from Advancing AI 2026, and learn how each benchmark was measured and what it really means.
DI Series: Serving Kimi-K3 MXFP4 Wide-EP Disaggregated on AMD Instinct MI300X and MI325X
Learn how to serve a 1.4 TiB Kimi-K3 MXFP4 mixture-of-experts model disaggregated across AMD Instinct MI300X and MI325X GPUs.
Automating Performance Bottleneck Identification with the TraceLens Agent
TraceLens Agent is an agentic workflow that examines GPU traces to deliver a prioritized optimization report, detailing root causes and fixes.
UltraQuant on AMD Instinct: More Efficient Agentic Serving for Qwen3.8-MXFP4
A native 4-bit MXFP4 KV cache that speeds up Qwen3.8 decode over 8-bit KV on AMD Instinct MI355X, with no accuracy loss.
Debugging Logprob Mismatches in LLM Reinforcement Learning
Learn to debug logprob mismatches between rollout and training with Qwen3 examples on AMD Instinct MI355X GPUs.
Hyperloom: A Multi-Agent Harness for Autonomous Inference Optimization on AMD GPUs
Hyperloom is a multi-agent harness that autonomously optimizes LLM inference on AMD Instinct GPUs, reaching a median 1.73x throughput gain.
Implementing a High-Performance Custom Diffusion Attention Kernel with FlyDSL
Learn how to implement and optimize flexible, high-performance diffusion attention kernels with FlyDSL.
An Educational GEMM Ladder for Helios GPUs
Build high-performance BF16 GEMM kernels on Helios GPUs with HipKittens, from a naive baseline to optimized schedules.
Iteratively Tuning hipBLASLt TensileLite Kernels: A Smaller Search, a Faster Kernel
Learn why a small iterative search beats a big one-shot sweep for hipBLASLt TensileLite kernels, in less time and with no risk of regression.