[REVISION HISTORY]

AMD expands Helios AI platform and launches Radeon RX 9050

Updated 2 times since CLSTR started tracking revisions of this situation.

What changed

2026-07-28 15:28 UTC → 2026-07-29 08:53 UTC · added ~~removed~~

AMD began full‑scale production of its second‑generation Helios rack‑scale AI platform in July 2026, targeting shipments in the next few months. The system combines up to 72 Instinct MI455X GPUs with up to 18 EPYC “Venice” Zen 6 CPUs – a 2 nm, 256‑core processor – delivering roughly 40 PFLOPS of AI compute at about 235 kW per rack and 31 TB of HBM4 memory. AMD announced a partnership with Anthropic that will install the first 1 GW of GPU capacity in early 2027, scaling to a total of 2 GW later. Additional collaborations with VAST Data, ASUS, Advantech, Cisco and ARBOR extend the stack to edge and enterprise servers. At the same event AMD introduced the Radeon RX 9050 Challenger graphics card, an entry‑level RDNA 4 part offered in 8 GB and a 4 GB variant with reduced memory bandwidth, delivering ~5.3 TFLOPS FP32. Critics note the lower‑spec model mirrors recent controversies over budget GPU memory limits. Schneider Electric and AMD unveiled a joint Helios reference design aimed at "AI factories," supporting up to 10.4 MW of IT load and up to 246 kW per rack. The design incorporates MI455X GPUs, EPYC “Venice” CPUs, Pensando Vulcano NICs, advanced liquid‑cooling from Motivair and digital‑twin tools such as ETAP and EcoStruxure for real‑time monitoring. Separately, AMD’s Ryzen AI Embedded X100 Series (Zen‑5 CPU, RDNA‑3 graphics, XDNA‑2 NPU) powers Foundation Future Industries’ Phantom humanoid robots, offered on a leasing basis. Supermicro launched its H15 server portfolio featuring 6th‑gen EPYC‑9006 processors, Instinct GPUs and Pensando networking, delivering up to 256 cores, doubled PCIe bandwidth and 2.6× memory throughput, positioning the systems as a full‑stack alternative to Nvidia‑based AI infrastructure. AMD also released ROCm.ai, an AI‑native software suite promising 3.3× inference and 2.4× training speedups over ROCm 7. ~~The roadmap foresees Zen 7 (2028) and Zen 8 (2030) CPUs, DDR6‑class memory and up to 2.9 exaflops of low‑precision performance from Helios by Q3 2026.~~

Versions

Only revisions since CLSTR began indexing content versions appear here. Select a version to see what changed compared to the one before it.