AMD pushes rack‑scale AI
AMD is advancing rack‑level AI infrastructure through its Helios collaboration with Celestica — a vendor push to capture full rack/system deals rather than component sales. That strategy makes AMD a competitor in pre‑integrated systems conversations with large buyers. (news.futunn.com)
Celestica says it will lead R&D, design and manufacturing of Helios scale‑up networking switches based on the Open Compute Project’s Open‑Rack‑Wide (ORW) form‑factor, and AMD says Helios will be available to customers in late 2026. (corporate.celestica.com) AMD publicly showed Helios at the OCP Global Summit 2025 as a double‑wide, open‑rack reference design that combines AMD Instinct MI450‑series accelerators, EPYC “Venice” CPUs and AMD Pensando networking. (storagenewsletter.com) A full Helios rack is specified to hold 72 MI450/MI455X accelerators and deliver up to about 1.4 exaFLOPS in FP8 and 2.9 exaFLOPS in FP4 precision while aggregating roughly 31 TB of HBM4 memory and ~1.4 PB/s of HBM bandwidth. (tomshardware.com) The design uses Ultra Accelerator Link over Ethernet (UALoE) for scale‑up connectivity, and Celestica says its scale‑up switches will use advanced networking silicon to enable high‑speed interconnect for the MI450 Series. (corporate.celestica.com) Oracle is a named launch partner: Oracle announced it will deploy an initial 50,000 AMD Instinct MI450 GPUs in OCI beginning in Q3 2026, which aligns with Helios‑style rack deployments announced by AMD. (oracle.com) Analysts have already factored Helios into outlooks—TD Cowen put AMD on its “Best Ideas 2026” list with a $290 target citing Helios/MI450 as a catalyst, and Bank of America raised its AMD price target to $300 after the Helios announcements. (finance.yahoo.com)
