Data-Center AI Accelerator Specs, Performance & Price (July 2026)

Comparison of major data-center AI accelerators (NVIDIA / AMD / Google / Intel / AWS): memory, bandwidth, compute, power and cloud rental price. Specs from official documents, retrieved 22 July 2026.

VendorChipMemoryBandwidth (TB/s)BF16 (TFLOPS)FP8 (TFLOPS)TDP (W)YearCloud rental ($/GPU/hr)
NVIDIAH100 SXM80 GB HBM33.3598919797002022$2.99
NVIDIAH200141 GB HBM3e4.898919797002024$4.39
NVIDIAB200192 GB HBM3e8.02250450010002024$5.89
NVIDIAB300 (Blackwell Ultra)288 GB HBM3eN/AN/AN/A14002026$7.39
AMDMI300X192 GB HBM35.3130726157502023N/A
AMDMI325X256 GB HBM3e6.01307261510002024N/A
AMDMI355X288 GB HBM3eN/AN/A500014002025N/A
GoogleTPU v5p95 GB HBM2e2.77459N/AN/A2023N/A
GoogleTPU v6e (Trillium)32 GB HBM1.64918N/AN/A2024N/A
GoogleTPU v7 (Ironwood)192 GB HBM7.37N/A4600N/A2025N/A
IntelGaudi 3128 GB HBM2e3.67167816789002024N/A
AWSTrainium296 GB2.96551287N/A2024N/A

Method & sources

Specs (memory / bandwidth / BF16 and FP8 compute / TDP / year) are from each vendor's official datasheets and product pages; BF16 and FP8 are dense (no sparsity) figures. Fields the vendor has not disclosed are marked "N/A" — no estimates. Performance is expressed as official BF16 dense TFLOPS; MLPerf submissions are system-level throughput that cannot be cleanly normalised per chip, so they are not used here (see mlcommons.org for cross-reference). Prices are RunPod on-demand cloud rental (USD per GPU per hour, retrieved 22 July 2026), from a single source for consistent comparison; chips RunPod does not list are "N/A", and non-public enterprise purchase prices are excluded. Chips without an official datasheet (e.g. Huawei Ascend, third-party/leaked data only) and unreleased chips are excluded. AI chip specs and cloud prices change frequently — verify on the official pages.

Source: https://www.nvidia.com/en-us/data-center/h200/

Retrieved: 2026-07-22