AI accelerators (GPU / ASIC)
AI accelerators (GPU / ASIC) compared across 9 products — generated from the aiinframap product registry (per-metric provenance + as-of on each product).
| Product / Vendor | Status | FP16/BF16 (TFLOPS) | FP8 (TFLOPS) | Memory (GB) | Mem BW (TB/s) | TDP (W) | Interconnect | Process (nm) |
|---|---|---|---|---|---|---|---|---|
| NVIDIA B200 (Blackwell) | — | 2,250 | 4,500 | 192 | 8 | 1,000 | 5th Gen NVLink (NVLink 5.0), 1.8 TB/s bidirectional per GPU | 4 |
| Intel Gaudi 3 | — | 1,835 | 1,835 | 128 | 3.7 | 900 | 24 × 200 GbE Ethernet (1,200 GB/s bidirectional) | 5 |
| AMD Instinct MI300X | — | 1,307.4 | 2,614.9 | 192 | 5.3 | 750 | Infinity Fabric Link (128 GB/s bidirectional per link, 7 links) | 5 |
| AMD Instinct MI325X | — | 1,307 | 2,614 | 256 | 6 | 800 | AMD Infinity Fabric Link (128 GB/s per link) | 5 |
| NVIDIA H100 SXM5 | — | 989.5 | 1,979 | 80 | 3.35 | 700 | NVLink 4.0 (900 GB/s) | 4 |
| NVIDIA H200 SXM | — | 989.5 | 1,979 | 141 | 4.8 | 700 | NVLink 4.0 (900 GB/s bidirectional per GPU) | 5 |
| Google TPU v6e (Trillium) | — | 918 | — | 32 | 1.64 | — | ICI (Interchip Interconnect), 3584 Gbps per chip, 2D Torus topology for 256-chip pods | — |
| AWS Trainium2 | — | 667 | 1,300 | 96 | 2.9 | 500 | NeuronLink v2 | 4 |
| Google TPU v5p | — | 459 | — | 95 | 2.765 | 450 | 4.8 Tbps ICI (Inter-Chip Interconnect), 3D torus topology | 7 |
Updated 6/23/2026
← Back to AI ASICs