AI Accelerators With the Most HBM per Chip (October 2026)

NVIDIA’s Rubin GPU is the only shipping part in the table with HBM4 memory. NVIDIA’s HGX page lists 288 GB of HBM4 at 22 TB/s per GPU for HGX Rubin NVL8¹, while its Vera Rubin NVL72 page lists the same 288 GB at 19.2 TB/s per GPU, and 20.7 TB of HBM4 across the 72-GPU rack². NVIDIA says Vera Rubin is ramping into full production, with server makers shipping Vera Rubin-based systems². CoreWeave received its first Vera Rubin NVL72 production racks in September 2026, according to NVIDIA³.

AMD’s MI350X carries 288 GB of HBM3E at 8 TB/s peak bandwidth, the same memory as the MI355X, on AMD’s CDNA 4 architecture¹¹. AMD lists its launch date as June 12, 2025 and its typical board power as 1,000 W¹¹. In its 2025 annual report, AMD says large hyperscale customers, OEMs and ODMs deployed its Instinct MI350 Series GPUs during the year¹². It ties with Rubin and the MI355X on capacity and sits after Rubin only because its bandwidth is lower.

The MI355X is the higher-power sibling of the MI350X: AMD lists the same 288 GB of HBM3E and 8 TB/s, a 256 MB last-level cache and a typical board power of 1,400 W, built on TSMC 3nm and 6nm processes¹⁰. It launched on the same date, June 12, 2025¹⁰. AMD’s 2025 annual report says it launched the MI350 Series in June and that cloud providers including Meta and Oracle have since expanded availability of MI350-based infrastructure¹².

NVIDIA’s Blackwell Ultra datasheet gives each GPU in the GB300 NVL72 rack 279 GB of HBM3E at 8 TB/s, and describes 279 GB of HBM3E per Blackwell Ultra chip with up to 37 TB of fast memory per rack⁴. The GB300 NVL72 page lists the rack as “Available Now”, with 20 TB of GPU memory across 72 Blackwell Ultra GPUs⁵. Epoch AI’s hardware dataset records this GPU at 287.5 GB²⁶; the table uses NVIDIA’s own datasheet figure.

In the eight-GPU HGX B300 server, NVIDIA’s datasheet lists each Blackwell Ultra GPU at 270 GB of HBM3E and 7.7 TB/s, lower than the same GPU in the GB300 NVL72 rack⁴. NVIDIA’s HGX page gives the baseboard 2.1 TB of total memory and states that HGX B300 and HGX B200 are shipping now¹. Because NVIDIA states different per-GPU memory for the two systems, each gets its own row.

AMD’s MI325X, launched October 10, 2024, has 256 GB of HBM3E on an 8,192-bit memory interface at a 6 GHz memory clock, for 6 TB/s of peak bandwidth¹³. It keeps the MI300X’s CDNA 3 architecture and 19,456 stream processors but moves from HBM3 to HBM3E¹³¹⁴. AMD’s 2025 annual report lists the MI325 series among its current Instinct data center GPU products¹².

Microsoft’s in-house inference accelerator pairs 216 GB of HBM3e at 7 TB/s with 272 MB of on-chip SRAM, built on TSMC’s 3nm process¹⁵. It isn’t sold to other companies; Microsoft says Maia 200 is deployed in its US Central datacenter region near Des Moines, Iowa, with the US West 3 region near Phoenix, Arizona, coming next¹⁵. It has the most HBM of any hyperscaler’s own chip in the ranked table.

Google’s seventh-generation TPU carries 192 GB of HBM per chip at approximately 7.37 TB/s, according to the TPU7x documentation¹⁶. Each chip is built from two chiplets, each with its own 96 GB of HBM, and a 3D torus interconnect lets slices scale to 9,216 chips¹⁶. Google Cloud made TPU7x generally available on March 31, 2026, after a preview from November 24, 2025¹⁷. Google doesn’t name the HBM generation on the page.

The oldest AMD part in the table, launched December 6, 2023, has 192 GB of HBM3 and 5.3 TB/s¹⁴. AMD’s footnote derives the precise 5.325 TB/s from an 8,192-bit memory bus at a 5.2 Gbps data rate¹⁴. It is the only ranked part on HBM3 rather than HBM3E or HBM4, and AMD’s 2025 annual report still lists the MI300 series among its Instinct products¹².

NVIDIA’s Blackwell datasheet lists each Blackwell GPU in the GB200 NVL72 rack at 186 GB of HBM3E and 8 TB/s, against 180 GB and 7.7 TB/s in the HGX B200 server⁶. The GB200 NVL72 page counts memory per GB200 superchip instead, 372 GB of HBM3E for its two GPUs⁷. NVIDIA said in February 2025 that CoreWeave had become the first cloud provider to make GB200 NVL72-based instances generally available⁸.