AMD Helios Rack-Scale AI System Targets Nvidia with 2.9 Exaflops
ryanshrout · x · 2026-07-24
AMD launched the Helios rack-scale AI compute solution, treating the entire rack as a single accelerator to close the gap with Nvidia in system-level delivery.
Core Specs
- High-Density Integration: A single rack packs 72 Instinct MI455X GPUs and 18 6th-gen EPYC server CPUs.
- Massive Compute: Delivers 2.9 exaflops of FP4 compute and 1.4 exaflops of FP8.
- Memory & Bandwidth: Features 31 terabytes of HBM4 memory and 1.7 petabytes per second of memory bandwidth.
- Networking: Any GPU reaches any other in a single hop. Scale-up uses UALoE over Ethernet at 260 TB/s.
Competitive Edge & Customers
- Open Standards: Unlike proprietary incumbent systems, Helios leans on open standards like OCP, UALink, and Ultra Ethernet.
- Performance Claims: AMD claims 15% higher peak FP4, 50% more memory capacity, and up to 30% more tokens per dollar.
- Customer Roster: OpenAI, Meta, and Anthropic are already on the customer list for gigawatt-scale deployments.
Related event: AMD Launches Rack-Scale AI Platform Helios, Now in Full Production(7 posts)→
More from Infra
- Leaked DeepSeek transcript says the company has only 20,000 H-equivalent cards — fiiiiiist · 2026-07-24
- NVIDIA teases NVFP4 as a cheaper path to LLM inference — NVIDIA Developer · 2026-07-24
- Investor Critique: Etching Transformers Into Silicon Is Inherently Limiting — JosephJacks_ · 2026-07-24
- Huawei’s 4:1 GB300 claim shrinks to 2:1 on memory bandwidth, thread says — zephyr_z9 · 2026-07-24
- NVIDIA introduces NVFP4 for faster LLM inference with less GPU memory — NVIDIA Developer · 2026-07-24
- DeepSeek-V4-Flash reaches 105 tok/s on two 4090D cards after Triton kernel rewrites — iSevenDays · 2026-07-24