AMD publishes educational GEMM ladder for Helios GPUs: 432GB HBM4, 23TB/s bandwidth
simran_s_arora · x · 2026-09-22
AMD's ROCm team and collaborators published a blog building an educational ladder of progressively optimized BF16 GEMM kernels for the new Helios GPU, inspired by Simon Boehm's classic CUDA GEMM worklog.
Key points:
- Hardware: 432 GB HBM4, 23 TB/s bandwidth, 40 PFLOPs FP4 per GPU; 256 WGPs across 8 XCDs, 320KB LDS per WGP; 72-GPU racks with 3.6 TB/s scale-up bandwidth and unified virtual memory, targeting long-context agentic workloads.
- New features: async HBM-to-shared-memory DMA transfers (TDM), simplified cache hierarchy, 32 threads-per-wave thread hierarchy, fine-grained synchronization, and workgroup-cluster launch with multicast.
- Framework: all kernels use HipKittens, with open-source code and examples on GitHub.
More from Infra
- VRAM vs Blackwell Speed: 3090 Ti or 5080 for Local Video Generation? — Apollorax · 2026-09-22
- Liquid AI's LFM2.5 tops mobile benchmarks: 2.32GB memory, 8s latency on iPhone 17 Pro — maximelabonne · 2026-09-22
- AMD Engineers Publish GEMM Optimization Tutorial Blog Inspired by the GEMM Ladder — simran_s_arora · 2026-09-22
- Why Meta hasn't shipped Muse in WhatsApp: not enough hardware for 2B users, says user — zephyr_z9 · 2026-09-22
- Terraform Industries makes high-purity methanol at scale, runs solar-direct electrolyzer under $100/kW — GabGarrett · 2026-09-22
- Deep-dive worklog: optimizing CUDA GEMM from naive kernel to shared-memory tiling — abhijithneil · 2026-09-22