Chips · accelerators

NVIDIA’s B200 does 5.3 times the Tesla V100 SXM2’s compute per watt

12 of 18 accelerators shipped by the date first promised. Each figure is quoted from the vendor’s own datasheet, product page or release.

Embed

The frame credits AMD, AWS, Google, Epoch AI, AI data centers and Epoch AI, Data on GPU clusters and links back to its page. Preview it.

Cite
CSV Post on X

16-bit compute per watt, by the year each chip shipped

dense BF16 or FP16 at the vendor’s maximum power; 12 of 26 chips have no published power
0 1,000 2,000 3,000 2017 2018 2019 2020 2021 2022 2023 2024 2025 GFLOPS per watt NVIDIA B200: 2,200 GFLOPS per watt B200 NVIDIA B300 (Blackwell Ultra): 2,000 GFLOPS per watt B300 (Blackwell Ultra) Intel Gaudi 3: 1,864 GFLOPS per watt Gaudi 3 AMD Instinct MI355X: 1,786 GFLOPS per watt Instinct MI355X AMD Instinct MI300X: 1,733 GFLOPS per watt Instinct MI300X NVIDIA H100 SXM: 1,414 GFLOPS per watt H100 SXM NVIDIA H200 SXM: 1,414 GFLOPS per watt H200 SXM AMD Instinct MI300A: 1,290 GFLOPS per watt Instinct MI300A NVIDIA GH200 Grace Hopper Superchip: 990 GFLOPS per watt GH200 Grace Hopper Superchip Intel Gaudi 2: 720 GFLOPS per watt Gaudi 2 AMD Instinct MI250X: 684 GFLOPS per watt Instinct MI250X NVIDIA Tesla V100 SXM2: 417 GFLOPS per watt Tesla V100 SXM2
NVIDIA AMD Intel

Every chip, every figure

dense compute for one chip; memory is high-bandwidth memory; each chip’s page gives the words each figure is quoted from
Chip Shipped 16-bit TFLOPS 8-bit TFLOPS FP4 TFLOPS Memory Bandwidth Power GFLOPS/W Runs at
AMD
Instinct MI250X 2021-11 383 383 INT8 TOPS — 128 GB 3.2 TB/s 560 W 684 7 clusters
Instinct MI300A 2023-12 981 1,960 — 128 GB 5.3 TB/s 760 W 1,290 4 clusters
Instinct MI300X 2023-12 1,300 2,610 — 192 GB 5.3 TB/s 750 W 1,733 3 clusters
Instinct MI325X none on record 1,300 2,610 — 256 GB 6 TB/s 1,000 W 1,300 none tracked
Instinct MI355X 2025-09 2,500 5,000 10,100 288 GB 8 TB/s 1,400 W 1,786 none tracked
AWS
Trainium2 2024-12 667 1,299 — 96 GiB 2.9 TB/s — — 5 data centers
Trainium3 2025-12 671 2,517 2,517 144 GiB 4.9 TB/s — — 2 data centers
Google
TPU v4 2022-10 275 275 INT8 TOPS — 32 GiB 1.2 TB/s — — 2 data centers, 4 clusters
TPU v5e 2023-11 197 393 INT8 TOPS — 16 GB 0.86 TB/s — — 10 data centers
TPU v5p 2024-04 459 459 — 95 GiB 2.77 TB/s — — 10 data centers, 1 cluster
TPU v6e 2024-12 918 918 — 32 GB 1.64 TB/s — — 14 data centers
TPU7x (Ironwood) 2026-03 2,307 4,614 — 192 GiB 7.38 TB/s — — 10 data centers
Huawei
Ascend 910C 2025-03 — — — — — — — 3 data centers
Ascend 950 2026-09 — 1,000 2,000 — — — — 1 data center
Intel
Gaudi 2 2022-06 432 865 — 96 GB 2.4 TB/s 600 W 720 1 cluster
Gaudi 3 2024-09 1,678 1,678 — 128 GB 3.7 TB/s 900 W 1,864 none tracked
Microsoft
Maia 100 none on record — — — 64 GB 1.8 TB/s 700 W — 1 cluster
NVIDIA
Tesla V100 SXM2 2017-08 125 FP16 — — 32 GB 0.9 TB/s 300 W 417 46 clusters
A100 SXM 80GB none on record 312 624 INT8 TOPS — 80 GB 2.04 TB/s 400 W 780 3 data centers, 89 clusters
H100 SXM 2022-09 990 1,979 — 80 GB 3.35 TB/s 700 W 1,414 14 data centers, 71 clusters
GH200 Grace Hopper Superchip 2023-05 990 1,979 — 96 GB 4 TB/s 1,000 W 990 10 clusters
H200 SXM 2024-08 990 1,979 — 141 GB 4.8 TB/s 700 W 1,414 3 data centers, 6 clusters
B200 2025-02 2,200 4,500 9,000 180 GB 7.7 TB/s 1,000 W 2,200 19 data centers
GB200 Grace Blackwell Superchip 2025-05 5,000 10,000 20,000 372 GB 16 TB/s — — 1 cluster
B300 (Blackwell Ultra) 2025-08 2,200 4,500 14,000 270 GB 7.7 TB/s 1,100 W 2,000 30 data centers
Rubin GPU 2026-08 4,000 17,500 35,000 288 GB 19.3 TB/s — — 11 data centers

12 of 18 accelerators shipped by the date first promised

the longest slip was the Google TPU v4, up to 10 months; a shipment is often the day the vendor said it was already shipping, so a slip is an upper bound
2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 today AMD AMD Instinct MI250X: shipped, no dated promise on record Instinct MI250X AMD Instinct MI300A: shipped by the first date promised Instinct MI300A AMD Instinct MI300X: shipped by the first date promised Instinct MI300X AMD Instinct MI325X: past its date, no dated shipment on record Instinct MI325X AMD Instinct MI355X: shipped by the first date promised Instinct MI355X AWS AWS Trainium2: shipped, no dated promise on record Trainium2 AWS Trainium3: shipped by the first date promised Trainium3 Google Google TPU v4: shipped after the first date promised TPU v4 Google TPU v5e: shipped, no dated promise on record TPU v5e Google TPU v5p: shipped, no dated promise on record TPU v5p Google TPU v6e: shipped by the first date promised TPU v6e Google TPU7x (Ironwood): shipped after the first date promised TPU7x (Ironwood) Huawei Huawei Ascend 910C: shipped, no dated promise on record Ascend 910C Huawei Ascend 950: shipped by the first date promised Ascend 950 Intel Intel Gaudi 2: shipped by the first date promised Gaudi 2 Intel Gaudi 3: shipped after the first date promised Gaudi 3 Microsoft Microsoft Maia 100: past its date, no dated shipment on record Maia 100 NVIDIA NVIDIA Tesla V100 SXM2: shipped by the first date promised Tesla V100 SXM2 NVIDIA A100 SXM 80GB: past its date, no dated shipment on record A100 SXM 80GB NVIDIA H100 SXM: shipped by the first date promised H100 SXM NVIDIA GH200 Grace Hopper Superchip: shipped by the first date promised GH200 Grace Hopper Superchip NVIDIA H200 SXM: shipped after the first date promised H200 SXM NVIDIA B200: shipped after the first date promised B200 NVIDIA GB200 Grace Blackwell Superchip: shipped after the first date promised GB200 Grace Blackwell Superchip NVIDIA B300 (Blackwell Ultra): shipped by the first date promised B300 (Blackwell Ultra) NVIDIA Rubin GPU: shipped by the first date promised Rubin GPU
Ring: the end of a promised window. Dot: the shipment, at the end of its stated period. Solid line: a shipment after the first promise. Dashed line: a date moved later.

NVIDIA’s B300 (Blackwell Ultra) runs at 30 of the 95 largest AI data centers

a site with two chip types counts for both; its H100-equivalents are the whole site’s
  1. NVIDIA B300 (Blackwell Ultra) 30 sites
  2. NVIDIA B200 19 sites
  3. NVIDIA H100 SXM 71 GPU clusters to July 2025 14 sites
  4. Google TPU v6e 14 sites
  5. NVIDIA Rubin GPU 11 sites
  6. Google TPU v5e 10 sites
  7. Google TPU v5p 1 GPU cluster to July 2025 10 sites
  8. Google TPU7x (Ironwood) 10 sites
  9. AWS Trainium2 5 sites
  10. NVIDIA A100 SXM 80GB 89 GPU clusters to July 2025 3 sites
  11. Huawei Ascend 910C 3 sites
  12. NVIDIA H200 SXM 6 GPU clusters to July 2025 3 sites
  13. Google TPU v4 4 GPU clusters to July 2025 2 sites
  14. AWS Trainium3 2 sites
  15. Huawei Ascend 950 1 site

Google’s TPU v6e gives the most compute per dollar at list price

dense 16-bit compute for an hour, over the vendor’s own on-demand price per chip-hour
Chip Price a chip-hour TFLOP-hours a dollar Price
Google TPU v6e $2.70 340 on-demand, us-east1 (first US region listed), per chip-hour
Google TPU7x (Ironwood) $12.00 192 on-demand, us-central1, per chip-hour
Google TPU v5e $1.20 164 on-demand, us-central1, per chip-hour
Google TPU v5p $4.20 109 on-demand, us-east5 (first US region listed), per chip-hour
Google TPU v4 $3.22 85 on-demand, us-central2 (Oklahoma), TPU v4 pod, per chip-hour

Epoch AI names these chips, which have no entry here

every other chip name Epoch AI gives maps to one chip above
  • H20 1 data center. NVIDIA publishes no datasheet for the H20, its chip for China, on any page this site could read.
  • AMD Instinct MI210 2 clusters. A PCIe card of AMD’s MI200 family. The record holds the family’s MI250X module.
  • Google TPU v3 2 clusters. A TPU from before 2021. The record starts with TPU v4.
  • NVIDIA Tesla K20X 2 clusters. An NVIDIA chip from before 2017. The record starts with the V100, NVIDIA’s first data-center chip with tensor cores.
  • NVIDIA Tesla K40c 2 clusters. An NVIDIA chip from before 2017. The record starts with the V100, NVIDIA’s first data-center chip with tensor cores.
  • Cerebras CS-1 1 cluster. A whole system built on one wafer-scale chip, not a chip in the sense of this record.
  • Cerebras CS-2 1 cluster. A whole system built on one wafer-scale chip, not a chip in the sense of this record.
  • Google TPU v1 1 cluster. A TPU from before 2021. The record starts with TPU v4.
  • Google TPU v2 1 cluster. A TPU from before 2021. The record starts with TPU v4.
  • GroqChip LPU v1 1 cluster. Not in the record yet: Groq sells cloud access and systems, and the record so far holds chips sold or rented by the largest vendors.
  • Huawei Ascend 910B 1 cluster. Not in the record yet. The record holds Huawei’s Ascend 910C and 950.
  • Intel Data Center GPU Max 1550 1 cluster. Intel’s GPU for supercomputers. The record follows Intel’s Gaudi AI accelerators.
  • NVIDIA A40 PCIe 1 cluster. A graphics and inference card, not one of the NVIDIA training chips this record follows.
  • NVIDIA L40 1 cluster. A graphics and inference card, not one of the NVIDIA training chips this record follows.
  • NVIDIA P100 1 cluster. An NVIDIA chip from before 2017. The record starts with the V100, NVIDIA’s first data-center chip with tensor cores.
  • NVIDIA Quadro RTX 5000 1 cluster. A workstation graphics card, not a data-center AI chip.
  • NVIDIA Tesla K40m 1 cluster. An NVIDIA chip from before 2017. The record starts with the V100, NVIDIA’s first data-center chip with tensor cores.
  • NVIDIA Tesla K80 1 cluster. An NVIDIA chip from before 2017. The record starts with the V100, NVIDIA’s first data-center chip with tensor cores.
  • NVIDIA Tesla P100 PCIe 16GB 1 cluster. An NVIDIA chip from before 2017. The record starts with the V100, NVIDIA’s first data-center chip with tensor cores.
  • NVIDIA Tesla P100 SXM2 1 cluster. An NVIDIA chip from before 2017. The record starts with the V100, NVIDIA’s first data-center chip with tensor cores.
  • PEZY-SC2 1 cluster. A many-core processor in Japanese supercomputers, outside the AI accelerators this record follows.

Chips left out, and why

a figure goes in only with a quote this site could read on the vendor’s own page
  • NVIDIA H20 NVIDIA has no product page or datasheet for it on its own site (nvidia.com/en-us/data-center/h20/ returns 404); only vGPU docs name it.

Sources

accelerators.csv Every dataset → Status → Glossary →
  • Each vendor’s own datasheets, product pages, documentation and releases. Every figure links to its page from the chip’s own page. Each page was read and each quote checked on 23 Oct 2026.
  • NVIDIA's Tensor Core figures marked as sparse are with sparsity, which doubles the dense figure: 2. Specification in sparse. Dense is one-half sparse spec shown. NVIDIA GB200 NVL72
  • Where each chip runs: Epoch AI, AI data centers (95 sites, retrieved 22 Oct 2026), and Epoch AI, GPU clusters (453 clusters after those Epoch marks as superseded; Epoch stopped the list in July 2025). Both CC BY 4.0.