machine learning hardware
We present key data on over 180 AI accelerators, such as graphics processing units (GPUs) and tensor processing units (TPUs), used to develop and deploy machine learning models in the deep learning era.
Charts and analysis of these data →
ml_hardware.csv
183 rows · 40 columns · Updated 2026-11-09
| Hardware name | Manufacturer | Type | Notes | Release date | Release price (USD) | Tensor-FP16/BF16 performance (FLOP/s) | FP8 performance (FLOP/s) | FP4 performance (FLOP/s) | Memory (bytes) | Memory bandwidth (byte/s) | Intranode bandwidth (byte/s) | Internode bandwidth (bit/s) | TDP (W) | FP64 (double precision) performance (FLOP/s) | FP32 (single precision) performance (FLOP/s) | TF32 (TensorFloat-32) performance (FLOP/s) | FP16 (half precision) performance (FLOP/s) | INT16 performance (OP/s) | INT8 performance (OP/s) | INT4 performance (OP/s) | Foundry | Die Size (mm^2) | Base clock (MHz) | Boost clock (MHz) | Memory clock (MHz) | Memory bus (bit) | Tensor cores | Process size (nm) | Transistors (millions) | Link to datasheet | Source for the price | ML models | Last modified | Max performance | Energy efficiency | Total processing performance (bit-OP/s) | Price-performance | ML OP/s | Inference Performance Index (GB300e) |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| NVIDIA VR200 (Vera Rubin) | NVIDIA | GPU | Source for release date (2026-09-30, early-access availability announcement): https://blogs.nvidia.com/blog/coreweave-agentic-ai-vera-rubin/ NVIDIA reports CoreWeave availability and Cognition production workloads by this date. First racks arrived earlier in September; not a universal GA date. | 2026-09-30 | 4000000000000000.0 | 1.75e+16 | 3.5e+16 | 288000000000.0 | 19200000000000.0 | 250000000000000.0 | https://nvdam.widen.net/s/7hztspzswk/gpu-architecture-datasheet-vera-rubin-nvidia-us-5198950-web https://www.nvidia.com/en-us/data-center/vera-rubin-nvl72/ https://dam-cdn.nvd.orangelogic.com/AssetLink/56p68o47y6f1yucpubl0ump1c0aif422.pdf https://www.nvidia.com/en-us/data-center/nvlink/ | 2026-10-21 00:02:33+00:00 | 3.5e+16 | 1.75e+16 | 2.177 | ||||||||||||||||||||||||
| Google TPU v8i | TPU | 2026-08-02 | 17500.0 | 1101000000000000.0 | 1.01e+16 | 1.01e+16 | 288000000000.0 | 8601000000000.0 | 2400000000000.0 | TSMC | https://cloud.google.com/blog/products/compute/tpu-8t-and-tpu-8i-technical-deep-dive, https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/eighth-generation-tpu-agentic-era/ | 2026-10-21 14:38:14+00:00 | 1.01e+16 | 8.08e+16 | 62914285714.28571 | 1.01e+16 | 0.912 | ||||||||||||||||||||||
| AMD Instinct MI455X | AMD | GPU | Inference Capacity Index update (2026-10-20): Saved score; see Scored inputs. Audited specs + stated modeling assumptions. Frozen Helios resource projection with FP4 experts; vendor-specific collective execution is uncalibrated. 72-GPU scale-up domain Interconnect convention: Intranode bandwidth is the aggregate bidirectional equivalent (2 × the study's one-way outbound scale-up bandwidth), consistent with the existing NVIDIA Airtable entries. Fabric assumptions: 72-chip switched scale-up. Release/availability: 2026-07-23. Basis: Vendor-listed launch date. AMD lists Launch Date 7/23/2026; does not establish universal system availability.. Source: https://www.amd.com/en/products/accelerators/instinct/mi400/mi455x.html Other populated fields retained from the supplied Airtable export where a name match exists. Derived performance/efficiency fields are left blank for Airtable to recompute, rather than carrying stale calculated values. | 2026-07-23 | 5000000000000000.0 | 2.01e+16 | 4.03e+16 | 432000000000.0 | 23300000000000.0 | 3600000000000.0 | https://www.amd.com/en/products/accelerators/instinct/mi400/mi455x.html https://www.amd.com/content/dam/amd/en/documents/products/accelerators/instinct/amd-instinct-mi455x_brochure.pdf | 2026-10-21 00:14:55+00:00 | 4.03e+16 | 2.01e+16 | 2.893 | ||||||||||||||||||||||||
| Meta MTIA 300 | Meta | Other | Primarily used for ranking and recommendation (R&R) inference and training. | 2026-03-11 | 600000000000000.0 | 1200000000000000.0 | 216000000000.0 | 6100000000000.0 | 800.0 | TSMC | https://ai.meta.com/blog/meta-mtia-scale-ai-chips-for-billions/ | 2026-05-27 18:59:10+00:00 | 1200000000000000.0 | 1500000000000.0 | 9600000000000000.0 | 1200000000000000.0 | |||||||||||||||||||||||
| Microsoft Maia 200 | Microsoft | GPU | Source for release date and specs: https://blogs.microsoft.com/blog/2026/01/26/maia-200-the-ai-accelerator-built-for-inference/ | 2026-01-26 | 5072000000000000.0 | 1.0145e+16 | 216000000000.0 | 7100000000000.0 | 2000000000000.0 | 750.0 | TSMC | 140000.0 | https://blogs.microsoft.com/blog/2026/01/26/maia-200-the-ai-accelerator-built-for-inference/ https://arxiv.org/html/2608.24664v1 | 2026-10-21 14:40:32+00:00 | 1.0145e+16 | 13526666666666.666 | 4.058e+16 | 5072000000000000.0 | 0.677 | ||||||||||||||||||||
| Amazon Trainium3 | Amazon AWS | GPU | Source for TDP: https://newsletter.semianalysis.com/p/aws-trainium3-deep-dive-a-potential to do: fill in memory specs | 2025-12-02 | 671000000000000.0 | 2517000000000000.0 | 2517000000000000.0 | 700.0 | 183000000000000.0 | 671000000000000.0 | TSMC | https://awsdocs-neuron.readthedocs-hosted.com/en/latest/about-neuron/arch/neuron-hardware/trainium3.html | 2026-10-22 17:37:52+00:00 | 2517000000000000.0 | 3595714285714.2856 | 2.0136e+16 | 2517000000000000.0 | 0.349 | |||||||||||||||||||||
| Google TPU v7 Ironwood | TPU | source for release date: https://www.cnbc.com/2025/11/06/google-unveils-ironwood-seventh-generation-tpu-competing-with-nvidia.html Google does not officially state TPUv7's TDP. In the blog post announcing TPUv7 (https://blog.google/products/google-cloud/ironwood-tpu-age-of-inference/ ), Google disclosed that TPUv7 is 29.3x more efficient than the TPUv2 in peak FLOP/s per TDP (measuring peak FLOP/s as either FP8 or bf16 FLOP/s when FP8 is not supported). TPUv2 does 46 TFLOP/s peak and has a TDP of 280 W. TPUv7 does 4614 TFLOP/s peak (FP8). So the TPUv7's TDP is (4614 TFLOP/s / 46 TFLOP/s) * 280 W / 29.3x efficiency = ~959 W. Note that this may be a slightly approximate figure due to rounding on Google's part. We round to 960 W. | 2025-11-06 | 12000.0 | 2307000000000000.0 | 4614000000000000.0 | 192000000000.0 | 7370000000000.0 | 1200000000000.0 | 960.0 | https://blog.google/products/google-cloud/ironwood-tpu-age-of-inference/ https://docs.cloud.google.com/tpu/docs/tpu7x | https://epoch.ai/data/ai-chip-sales-documentation/methodology#google | Gemini 3 Pro | 2026-10-21 14:42:52+00:00 | 4614000000000000.0 | 4806250000000.0 | 3.6912e+16 | 192250000000.0 | 4614000000000000.0 | 0.742 | |||||||||||||||||||
| Huawei Ascend 920 | Huawei | NPU | 2025-10-01 | 900000000000000.0 | 4000000000000.0 | SMIC | 6.0 | https://www.tomshardware.com/pc-components/gpus/huawei-introduces-the-ascend-920-ai-chip-to-fill-the-void-left-by-nvidias-h20 | 2025-07-30 16:53:19+00:00 | 900000000000000.0 | 900000000000000.0 | ||||||||||||||||||||||||||||
| NVIDIA GB300 (Blackwell Ultra) | NVIDIA | GPU | Individual GPUs are sometimes called Blackwell Ultra, but systems are referred to as GB300. GB300 is only available in systems containing two or more GPUs. Performance figures are per individual GPU, not per GB300 superchip. Performance figures are listed without sparsity. https://resources.nvidia.com/en-us-blackwell-architecture/blackwell-ultra-datasheet Memory capacity conflict: the current NVIDIA Ultra PDF lists 279 GB, while other NVIDIA system docs list 288 GB. The 288 GB (decimal) figure is retained as a study assumption for the Inference Capacity Index (reference denominator for each model). Installed vs usable/SKU interpretation is unresolved; no automatic reserve subtraction. | 2025-08-22 | 43000.0 | 2500000000000000.0 | 5000000000000000.0 | 1.5e+16 | 288000000000.0 | 8000000000000.0 | 1800000000000.0 | 1400.0 | 1300000000000.0 | 80000000000000.0 | 1250000000000000.0 | 165000000000000.0 | 1.5e+16 | TSMC | 4.0 | https://resources.nvidia.com/en-us-dgx-systems/dgx-b300-datasheet https://www.nvidia.com/en-us/data-center/gb300-nvl72/ https://dam-cdn.nvd.orangelogic.com/AssetLink/1k0p832eq8r5ca0u5383ie5o4tp3bst1.pdf | GPT-5.5,Inkling,Inkling-Small,Beam 1.0 | 2026-10-26 19:42:38+00:00 | 1.5e+16 | 10714285714285.715 | 6.112e+16 | 58139534883.72093 | 5000000000000000.0 | 1.0 | |||||||||||
| NVIDIA B300 (Blackwell Ultra) | NVIDIA | GPU | Individual GPUs are sometimes called Blackwell Ultra, but systems are referred to as B300. Performance figures are listed without sparsity. | 2025-08-22 | 2250000000000000.0 | 4500000000000000.0 | 1.35e+16 | 288000000000.0 | 7750000000000.0 | 1800000000000.0 | 1100.0 | 1200000000000.0 | 75000000000000.0 | 1100000000000000.0 | 157500000000000.0 | 1.35e+16 | TSMC | 640.0 | 4.0 | 208000.0 | https://developer.nvidia.com/blog/inside-nvidia-blackwell-ultra-the-chip-powering-the-ai-factory-era/ https://resources.nvidia.com/en-us-blackwell-architecture/blackwell-ultra-datasheet https://dam-cdn.nvd.orangelogic.com/AssetLink/1k0p832eq8r5ca0u5383ie5o4tp3bst1.pdf https://docs.nvidia.com/dgx/dgxb300-user-guide/introduction-to-dgxb300.html | https://epoch.ai/data/ai-chip-sales-documentation/methodology#nvidia | Arcee Trinity Large,Composer 2,Composer 2.5 | 2026-10-21 01:36:38+00:00 | 1.35e+16 | 12272727272727.273 | 4500000000000000.0 | 0.541 | |||||||||||
| AMD Instinct MI355X | AMD | GPU | release announcement (source for release date): https://www.tomshardware.com/pc-components/gpus/amd-announces-mi350x-and-mi355x-ai-gpus-claims-up-to-4x-generational-gain-up-to-35x-faster-inference-performance | 2025-06-12 | 30000.0 | 2516600000000000.0 | 5033200000000000.0 | 1.00663e+16 | 288000000000.0 | 8000000000000.0 | 1075200000000.0 | 1400.0 | 2340000000000000.0 | 5033200000000000.0 | 1.00663e+16 | TSMC | 3.0 | https://www.amd.com/content/dam/amd/en/documents/instinct-tech-docs/product-briefs/amd-instinct-mi355x-gpu-brochure.pdf https://www.amd.com/content/dam/amd/en/documents/instinct-tech-docs/product-briefs/amd-instinct-miI355x-platform-brochure.pdf | https://seekingalpha.com/news/4472735-amd-jumps-after-report-of-plans-to-raise-price-of-instinct-mi350-ai-chips | 2026-10-21 14:33:19+00:00 | 1.00663e+16 | 7190214285714.286 | 3.744e+16 | 167773333333.33334 | 5033200000000000.0 | 0.552 | |||||||||||||
| AMD Instinct MI350X | AMD | GPU | release announcement (source for release date): https://www.tomshardware.com/pc-components/gpus/amd-announces-mi350x-and-mi355x-ai-gpus-claims-up-to-4x-generational-gain-up-to-35x-faster-inference-performance | 2025-06-12 | 25000.0 | 2309600000000000.0 | 4600000000000000.0 | 9200000000000000.0 | 288000000000.0 | 8000000000000.0 | 1000.0 | 2340000000000000.0 | 4614000000000000.0 | 9227500000000000.0 | TSMC | 3.0 | https://www.amd.com/content/dam/amd/en/documents/instinct-tech-docs/product-briefs/amd-instinct-mi350x-gpu-brochure.pdf | https://seekingalpha.com/news/4472735-amd-jumps-after-report-of-plans-to-raise-price-of-instinct-mi350-ai-chips | 2026-06-25 16:29:09+00:00 | 9227500000000000.0 | 9227500000000.0 | 3.744e+16 | 184560000000.0 | 4614000000000000.0 |