Database section · 55 records shown
Chips
GPUs, AI accelerators, CPUs and superchips — every record carries a labeled spec sheet, its manufacturing chain, and its connections through the stack.
Field notes
Fifty-five chips, one question: compute is abundant, so what's scarce? Memory, packaging and power. Read any spec sheet here with that lens — the headline FLOPS matter less than the gigabytes beside them — and the basis labels tell you what's measured versus marketed.
Start with these three
NB
NVIDIA B200
The current workhorse of the frontier — two dies pretending to be one.
GT
Google TPU v7 (Ironwood)
The strongest proof that the non-NVIDIA path works at scale.
AI
AMD Instinct MI355X
The challenger that finally made it a real race.
Memory bandwidth by release year
The argument of this whole section in one picture: per-chip memory bandwidth has climbed roughly tenfold in eight years, and every step up this curve is an HBM generation. Click any point to open its record; wafer-scale on-chip memory is excluded as a different animal.
Vendor-spec memory bandwidth per chip (TB/s) against first-availability year, from the records below — dense HBM parts only.
Reading tips: memory bandwidth (TB/s) usually predicts real AI performance better than raw compute; TDP is the chip's power budget in watts; the process node tells you which factory generation makes it. Click any chip for the full labeled spec sheet.
Figures below are the key facts; every value on a record page carries its basis (vendor spec / claim / reported / estimate). Missing values mean the vendor has not published one.
| Chip | Class | Year | Process | Memory | Mem BW (TB/s) | TDP (W) | Status | ★ |
|---|---|---|---|---|---|---|---|---|
| AMD Instinct MI400 / MI450AMD | GPU | 2026 | — | 432 GB HBM4 | — | — | Ramping | |
| AMD Instinct MI355XAMD | GPU | 2025 | TSMC N3P | 288 GB HBM3E | 8.00 | 1,400 | Shipping | |
| AMD EPYC 9005 (Turin)AMD | CPU | 2024 | TSMC N4/N3 | — | — | — | Shipping | |
| AMD Instinct MI325XAMD | GPU | 2024 | TSMC N5/N6 | 256 GB HBM3E | 6.00 | 1,000 | Shipping | |
| AMD Instinct MI300AAMD | APU | 2023 | TSMC N5/N6 | 128 GB HBM3 (unified) | 5.30 | 760 | Shipping | |
| AMD Instinct MI300XAMD | GPU | 2023 | TSMC N5/N6 | 192 GB HBM3 | 5.30 | 750 | Shipping | |
| AMD Instinct MI250XAMD | GPU | 2021 | TSMC N6 | 128 GB HBM2e | 3.20 | 560 | Prior generation | |
| AWS Trainium3AWS Annapurna | AI Accelerator | 2026 | 3 nm-class | — | — | — | Announced | |
| AWS Graviton4AWS Annapurna | CPU | 2024 | — | — | — | — | Shipping | |
| AWS Trainium2AWS Annapurna | AI Accelerator | 2024 | — | 96 GB HBM3 | 2.90 | — | Shipping | |
| AWS Inferentia2AWS Annapurna | AI Accelerator | 2023 | — | 32 GB HBM | — | — | Shipping | |
| AWS TrainiumAWS Annapurna | AI Accelerator | 2022 | — | 32 GB HBM | — | — | Prior generation | |
| Biren BR100Biren | AI Accelerator | 2022 | TSMC N7 | — | — | — | Historical | |
| Cambricon Siyuan 590Cambricon | AI Accelerator | 2023 | — | — | — | — | Shipping | |
| Cerebras WSE-3Cerebras | Wafer-Scale | 2024 | TSMC N5 | 44 GB on-chip SRAM | 21,000 | 23,000 | Shipping | |
| Etched SohuEtched | AI Accelerator | 2026 | — | 144 GB HBM3E | — | — | In development | |
| FuriosaAI RNGDFuriosaAI | AI Accelerator | 2024 | TSMC N5 | 48 GB HBM3 | — | 150 | Ramping | |
| Google TPU v7 (Ironwood)Google | AI Accelerator | 2025 | — | 192 GB HBM3E | 7.37 | — | Ramping | |
| Google AxionGoogle | CPU | 2024 | — | — | — | — | Shipping | |
| Google TPU v6e (Trillium)Google | AI Accelerator | 2024 | — | 32 GB HBM | 1.60 | — | Shipping | |
| Google TPU v5eGoogle | AI Accelerator | 2023 | — | 16 GB HBM2 | 0.82 | — | Shipping | |
| Google TPU v5pGoogle | AI Accelerator | 2023 | — | 95 GB HBM2e | 2.77 | — | Shipping | |
| Google TPU v4Google | AI Accelerator | 2021 | TSMC N7 | 32 GB HBM2 | 1.20 | 192 | Prior generation | |
| Google TPU v1–v3Google | AI Accelerator | 2015 | 28 nm (v1) | — | — | — | Historical | |
| Graphcore Bow IPUGraphcore | AI Accelerator | 2022 | TSMC N7 (WoW) | 900 MB SRAM | — | — | Historical | |
| Groq LPUGroq | AI Accelerator | 2020 | GF 14 nm | 230 MB SRAM | — | — | Shipping | |
| Huawei Ascend 910CHuawei (HiSilicon) | AI Accelerator | 2025 | SMIC N+2 (7 nm-class) | 128 GB HBM2E | — | — | Shipping | |
| Intel Xeon 6Intel | CPU | 2024 | Intel 3 | — | — | — | Shipping | |
| Intel Gaudi 3Intel (Habana) | AI Accelerator | 2024 | TSMC N5 | 128 GB HBM2E | 3.70 | 900 | Shipping | |
| Intel Gaudi 2Intel (Habana) | AI Accelerator | 2022 | TSMC N7 | 96 GB HBM2E | 2.45 | 600 | Prior generation | |
| Meta MTIA v2Meta | AI Accelerator | 2024 | TSMC N5 | — | — | 90 | Shipping | |
| Microsoft Cobalt 100Microsoft | CPU | 2024 | — | — | — | — | Shipping | |
| Microsoft Maia 100Microsoft | AI Accelerator | 2024 | TSMC N5 | 64 GB HBM2E | 1.80 | 700 | Shipping | |
| NVIDIA RubinNVIDIA | GPU | 2026 | TSMC N3 (reported) | 288 GB HBM4 | — | — | Ramping | |
| NVIDIA Rubin CPXNVIDIA | AI Accelerator | 2026 | — | 128 GB GDDR7 | — | — | Announced | |
| NVIDIA Vera CPUNVIDIA | CPU | 2026 | — | — | — | — | Announced | |
| NVIDIA B300 (Blackwell Ultra)NVIDIA | GPU | 2025 | TSMC 4NP | 288 GB HBM3E | 8.00 | 1,400 | Ramping | |
| NVIDIA B100NVIDIA | GPU | 2024 | TSMC 4NP | 192 GB HBM3E | 8.00 | 700 | Historical | |
| NVIDIA B200NVIDIA | GPU | 2024 | TSMC 4NP | 192 GB HBM3E | 8.00 | 1,000 | Shipping | |
| NVIDIA GB200 Grace BlackwellNVIDIA | Superchip | 2024 | TSMC 4NP | 384 GB HBM3E + 480 GB LPDDR5X | 16.00 | 2,700 | Shipping | |
| NVIDIA H200NVIDIA | GPU | 2024 | TSMC 4N | 141 GB HBM3E | 4.80 | 700 | Shipping | |
| NVIDIA BlueField-3NVIDIA | DPU | 2023 | — | — | — | — | Shipping | |
| NVIDIA GH200 Grace HopperNVIDIA | Superchip | 2023 | TSMC 4N | 96 GB HBM3 or 144 GB HBM3E + 480 GB LPDDR5X | 4.90 | 1,000 | Shipping | |
| NVIDIA Grace CPUNVIDIA | CPU | 2023 | TSMC 4N | 480 GB LPDDR5X | — | 500 | Shipping | |
| NVIDIA H800NVIDIA | GPU | 2023 | TSMC 4N | 80 GB HBM3 | 3.35 | 700 | Historical | |
| NVIDIA L40SNVIDIA | GPU | 2023 | TSMC 4N | 48 GB GDDR6 | 0.864 | 350 | Shipping | |
| NVIDIA A800NVIDIA | GPU | 2022 | TSMC N7 | 40/80 GB HBM2e | 2.00 | 400 | Historical | |
| NVIDIA H100NVIDIA | GPU | 2022 | TSMC 4N | 80 GB HBM3 | 3.35 | 700 | Shipping | |
| NVIDIA A100NVIDIA | GPU | 2020 | TSMC N7 | 40/80 GB HBM2e | 2.00 | 400 | Prior generation | |
| NVIDIA V100NVIDIA | GPU | 2017 | TSMC 12FFN | 16/32 GB HBM2 | 0.9 | 300 | Historical | |
| Qualcomm Cloud AI / AI200Qualcomm | AI Accelerator | 2026 | — | — | — | — | Announced | |
| Rebellions REBELRebellions | AI Accelerator | 2025 | Samsung | — | — | — | In development | |
| SambaNova SN40LSambaNova | AI Accelerator | 2023 | TSMC N5 | 64 GB HBM + 1.5 TB DDR | — | — | Shipping | |
| Tenstorrent BlackholeTenstorrent | AI Accelerator | 2024 | — | 32 GB GDDR6 | — | — | Shipping | |
| d-Matrix Corsaird-Matrix | AI Accelerator | 2024 | — | — | — | — | Ramping |
Mem BW = peak memory bandwidth, vendor spec. TDP = maximum configured thermal design power. "—" means the vendor has not published a figure. Download: chips.csv · JSON