Interconnecting Intelligence The facts on AI, from demand to buildout.

What is the fastest AI chip?

The fastest AI chip in Epoch AI's data, AMD Instinct MI455X, delivers 5.0 petaFLOP/s at 16-bit precision.

Updates: when the source publishes · Last checked Nov 21, 2026

The fastest AI chip in Epoch AI's data, AMD Instinct MI455X, delivers 5.0 petaFLOP/s at 16-bit precision. Dense FP16/BF16 tensor performance of 70 AI accelerators by release date, FLOP/s, log scale Data in the table below.
Download image Point at the chart, or tap it, to see the values. The image is sized for X and LinkedIn, with source and address. Source: [13]
Show the data
ChipMakerReleasedFP16/BF16 TFLOP/sMemory bandwidthPower
NVIDIA VR200 (Vera Rubin)NVIDIA2026-09-304,00019.2 TB/s2300 W
Google TPU v8iGoogle2026-08-021,1018.6 TB/s–
AMD Instinct MI455XAMD2026-07-235,00023.3 TB/s–
Meta MTIA 300Meta2026-03-116006.1 TB/s800 W
Amazon Trainium3Amazon AWS2025-12-02671–700 W
Google TPU v7 IronwoodGoogle2025-11-062,3077.4 TB/s960 W
Huawei Ascend 920Huawei2025-10-019004.0 TB/s–
NVIDIA GB300 (Blackwell Ultra)NVIDIA2025-08-222,5008.0 TB/s1400 W
NVIDIA B300 (Blackwell Ultra)NVIDIA2025-08-222,2507.8 TB/s1100 W
AMD Instinct MI355XAMD2025-06-122,5178.0 TB/s1400 W
AMD Instinct MI350XAMD2025-06-122,3108.0 TB/s1000 W
NVIDIA RTX Pro 6000 BlackwellNVIDIA2025-03-185041.8 TB/s600 W
NVIDIA GB200NVIDIA2025-02-152,5008.0 TB/s1200 W
Amazon Trainium2Amazon AWS2024-12-036672.9 TB/s500 W
NVIDIA H200 SXMNVIDIA2024-11-189904.8 TB/s700 W
NVIDIA B200NVIDIA2024-11-152,2507.7 TB/s1000 W
Huawei Ascend 910CHuawei2024-10-158003.2 TB/s700 W
AMD Instinct MI325XAMD2024-10-101,3076.0 TB/s1000 W
Intel Habana Gaudi3Intel2024-09-241,6783.7 TB/s900 W
Maia 100 (M100)Microsoft2024-08-27800–500 W
Google TPU v6e TrilliumGoogle2024-05-149181.6 TB/s380 W
Meta MTIA v2 (MTIA 200)Meta2024-04-101771.0 TB/s90 W
MTT S4000Moore Threads (Tencent)2023-12-191000.8 TB/s450 W
Google TPU v5pGoogle2023-12-064592.8 TB/s540 W
AMD Instinct MI300XAMD2023-12-061,3075.3 TB/s750 W
AMD Radeon Instinct MI308XAMD2023-12-0665410.3 TB/s750 W
NVIDIA HGX H20NVIDIA2023-11-091484.0 TB/s400 W
NVIDIA L2 PCleNVIDIA2023-11-09960.3 TB/s–
NVIDIA L20 PCleNVIDIA2023-09-161200.9 TB/s275 W
Google TPU v5eGoogle2023-08-291970.8 TB/s225 W
NVIDIA GH200NVIDIA2023-08-089904.9 TB/s700 W
Meta MTIA v1 (MTIA 100)Meta2023-05-18510.8 TB/s25 W
AWS Inferentia2Amazon AWS2023-04-131900.8 TB/s–
NVIDIA H800 SXM5NVIDIA2023-03-219893.4 TB/s700 W
NVIDIA GH100NVIDIA2023-03-219903.1 TB/s700 W
NVIDIA L4NVIDIA2023-03-211210.3 TB/s72 W
AMD Instinct MI300AAMD2023-01-049815.3 TB/s760 W
NVIDIA A800 PCIe 40 GBNVIDIA2022-11-083121.6 TB/s250 W
NVIDIA A800 PCIe 80 GBNVIDIA2022-11-033121.9 TB/s250 W
Amazon Trainium1Amazon AWS2022-10-151900.8 TB/s–
NVIDIA L40NVIDIA2022-10-131810.9 TB/s300 W
NVIDIA H100 PCIeNVIDIA2022-09-207562.0 TB/s350 W
NVIDIA GeForce RTX 4090NVIDIA2022-09-203301.0 TB/s450 W
NVIDIA H100 SXM5 80GBNVIDIA2022-09-209903.4 TB/s700 W
Biren BR100Biren2022-08-221,0241.6 TB/s550 W
NVIDIA A800 SXMNVIDIA2022-08-113122.0 TB/s400 W
Intel Habana Gaudi2Intel2022-05-104502.5 TB/s600 W
NVIDIA GeForce RTX 3090 TiNVIDIA2022-03-291601.0 TB/s450 W
AMD Radeon Instinct MI250XAMD2021-11-083833.3 TB/s500 W
Tesla D1 DojoTesla2021-08-2136210.0 TB/s400 W
Google TPU v4Google2021-05-202751.2 TB/s340 W
NVIDIA A100 SXM4 80 GBNVIDIA2020-11-163122.0 TB/s400 W
NVIDIA A40 PCIeNVIDIA2020-10-051500.7 TB/s300 W
NVIDIA A100 PCIeNVIDIA2020-06-223121.9 TB/s300 W
NVIDIA A100 SXM4 40 GBNVIDIA2020-05-143121.6 TB/s400 W
NVIDIA A100NVIDIA2020-03-013121.6 TB/s400 W
Google TPU v4iGoogle2020-01-011380.6 TB/s175 W
NVIDIA Tesla V100S PCIe 32 GBNVIDIA2019-11-261301.1 TB/s250 W
NVIDIA GeForce RTX 2080 Ti 11GBNVIDIA2018-09-201140.6 TB/s250 W
Google TPU v3Google2018-05-181230.9 TB/s450 W
NVIDIA Tesla V100 DGXS 16 GBNVIDIA2018-03-271250.9 TB/s250 W
NVIDIA Tesla V100 PCIe 32 GBNVIDIA2018-03-271120.9 TB/s250 W
NVIDIA Tesla V100 DGXS 32 GBNVIDIA2018-03-271250.9 TB/s250 W
NVIDIA Tesla V100 SXM3 32 GBNVIDIA2018-03-271250.9 TB/s300 W
NVIDIA Tesla V100 SXM2 32 GBNVIDIA2018-03-271250.9 TB/s300 W
NVIDIA Tesla V100 SXM2NVIDIA2017-07-011250.9 TB/s300 W
NVIDIA Tesla V100 SXM2 16 GBNVIDIA2017-06-211250.9 TB/s300 W
NVIDIA Tesla V100 PCIe 16 GBNVIDIA2017-06-211120.9 TB/s250 W
NVIDIA V100NVIDIA2017-06-211250.9 TB/s300 W
Google TPU v2Google2017-05-01460.6 TB/s280 W

Key figures

Points on the chart
70
VR200 (Vera Rubin)
Release date: Sep 30, 2026; Tera FLOP/s, FP16/BF16 (1,000 = 1 petaFLOP/s): 4,000 TFLOP/s
Google TPU v8i
Release date: Aug 2, 2026; Tera FLOP/s, FP16/BF16 (1,000 = 1 petaFLOP/s): 1,101 TFLOP/s
MI455X
Release date: Jul 23, 2026; Tera FLOP/s, FP16/BF16 (1,000 = 1 petaFLOP/s): 5,000 TFLOP/s
Meta MTIA 300
Release date: Mar 11, 2026; Tera FLOP/s, FP16/BF16 (1,000 = 1 petaFLOP/s): 600 TFLOP/s

Worked out from the chart's data, as of Nov 21, 2026.

How it's measured

Epoch AI's Machine Learning Hardware database lists accelerators' peak dense 16-bit (FP16/BF16) tensor performance from their datasheets, with release dates.

Datasheet peaks are upper limits; real workloads reach less, and memory and networking matter as much. The data is licensed under CC BY 4.0 and refreshes daily.

On the chart: Peak dense performance from datasheets as collected by Epoch AI; real workloads reach less.

Sources

The data

Download as CSV · Free to use with credit to Interconnecting Intelligence and the original source.

ChipMakerReleasedFP16/BF16 TFLOP/sMemory bandwidthPower
NVIDIA VR200 (Vera Rubin)NVIDIA2026-09-304,00019.2 TB/s2300 W
Google TPU v8iGoogle2026-08-021,1018.6 TB/s–
AMD Instinct MI455XAMD2026-07-235,00023.3 TB/s–
Meta MTIA 300Meta2026-03-116006.1 TB/s800 W
Amazon Trainium3Amazon AWS2025-12-02671–700 W
Google TPU v7 IronwoodGoogle2025-11-062,3077.4 TB/s960 W
Huawei Ascend 920Huawei2025-10-019004.0 TB/s–
NVIDIA GB300 (Blackwell Ultra)NVIDIA2025-08-222,5008.0 TB/s1400 W
NVIDIA B300 (Blackwell Ultra)NVIDIA2025-08-222,2507.8 TB/s1100 W
AMD Instinct MI355XAMD2025-06-122,5178.0 TB/s1400 W
AMD Instinct MI350XAMD2025-06-122,3108.0 TB/s1000 W
NVIDIA RTX Pro 6000 BlackwellNVIDIA2025-03-185041.8 TB/s600 W
NVIDIA GB200NVIDIA2025-02-152,5008.0 TB/s1200 W
Amazon Trainium2Amazon AWS2024-12-036672.9 TB/s500 W
NVIDIA H200 SXMNVIDIA2024-11-189904.8 TB/s700 W
NVIDIA B200NVIDIA2024-11-152,2507.7 TB/s1000 W
Huawei Ascend 910CHuawei2024-10-158003.2 TB/s700 W
AMD Instinct MI325XAMD2024-10-101,3076.0 TB/s1000 W
Intel Habana Gaudi3Intel2024-09-241,6783.7 TB/s900 W
Maia 100 (M100)Microsoft2024-08-27800–500 W
Google TPU v6e TrilliumGoogle2024-05-149181.6 TB/s380 W
Meta MTIA v2 (MTIA 200)Meta2024-04-101771.0 TB/s90 W
MTT S4000Moore Threads (Tencent)2023-12-191000.8 TB/s450 W
Google TPU v5pGoogle2023-12-064592.8 TB/s540 W
AMD Instinct MI300XAMD2023-12-061,3075.3 TB/s750 W
AMD Radeon Instinct MI308XAMD2023-12-0665410.3 TB/s750 W
NVIDIA HGX H20NVIDIA2023-11-091484.0 TB/s400 W
NVIDIA L2 PCleNVIDIA2023-11-09960.3 TB/s–
NVIDIA L20 PCleNVIDIA2023-09-161200.9 TB/s275 W
Google TPU v5eGoogle2023-08-291970.8 TB/s225 W
NVIDIA GH200NVIDIA2023-08-089904.9 TB/s700 W
Meta MTIA v1 (MTIA 100)Meta2023-05-18510.8 TB/s25 W
AWS Inferentia2Amazon AWS2023-04-131900.8 TB/s–
NVIDIA H800 SXM5NVIDIA2023-03-219893.4 TB/s700 W
NVIDIA GH100NVIDIA2023-03-219903.1 TB/s700 W
NVIDIA L4NVIDIA2023-03-211210.3 TB/s72 W
AMD Instinct MI300AAMD2023-01-049815.3 TB/s760 W
NVIDIA A800 PCIe 40 GBNVIDIA2022-11-083121.6 TB/s250 W
NVIDIA A800 PCIe 80 GBNVIDIA2022-11-033121.9 TB/s250 W
Amazon Trainium1Amazon AWS2022-10-151900.8 TB/s–
NVIDIA L40NVIDIA2022-10-131810.9 TB/s300 W
NVIDIA H100 PCIeNVIDIA2022-09-207562.0 TB/s350 W
NVIDIA GeForce RTX 4090NVIDIA2022-09-203301.0 TB/s450 W
NVIDIA H100 SXM5 80GBNVIDIA2022-09-209903.4 TB/s700 W
Biren BR100Biren2022-08-221,0241.6 TB/s550 W
NVIDIA A800 SXMNVIDIA2022-08-113122.0 TB/s400 W
Intel Habana Gaudi2Intel2022-05-104502.5 TB/s600 W
NVIDIA GeForce RTX 3090 TiNVIDIA2022-03-291601.0 TB/s450 W
AMD Radeon Instinct MI250XAMD2021-11-083833.3 TB/s500 W
Tesla D1 DojoTesla2021-08-2136210.0 TB/s400 W
Google TPU v4Google2021-05-202751.2 TB/s340 W
NVIDIA A100 SXM4 80 GBNVIDIA2020-11-163122.0 TB/s400 W
NVIDIA A40 PCIeNVIDIA2020-10-051500.7 TB/s300 W
NVIDIA A100 PCIeNVIDIA2020-06-223121.9 TB/s300 W
NVIDIA A100 SXM4 40 GBNVIDIA2020-05-143121.6 TB/s400 W
NVIDIA A100NVIDIA2020-03-013121.6 TB/s400 W
Google TPU v4iGoogle2020-01-011380.6 TB/s175 W
NVIDIA Tesla V100S PCIe 32 GBNVIDIA2019-11-261301.1 TB/s250 W
NVIDIA GeForce RTX 2080 Ti 11GBNVIDIA2018-09-201140.6 TB/s250 W
Google TPU v3Google2018-05-181230.9 TB/s450 W
NVIDIA Tesla V100 DGXS 16 GBNVIDIA2018-03-271250.9 TB/s250 W
NVIDIA Tesla V100 PCIe 32 GBNVIDIA2018-03-271120.9 TB/s250 W
NVIDIA Tesla V100 DGXS 32 GBNVIDIA2018-03-271250.9 TB/s250 W
NVIDIA Tesla V100 SXM3 32 GBNVIDIA2018-03-271250.9 TB/s300 W
NVIDIA Tesla V100 SXM2 32 GBNVIDIA2018-03-271250.9 TB/s300 W
NVIDIA Tesla V100 SXM2NVIDIA2017-07-011250.9 TB/s300 W
NVIDIA Tesla V100 SXM2 16 GBNVIDIA2017-06-211250.9 TB/s300 W
NVIDIA Tesla V100 PCIe 16 GBNVIDIA2017-06-211120.9 TB/s250 W
NVIDIA V100NVIDIA2017-06-211250.9 TB/s300 W
Google TPU v2Google2017-05-01460.6 TB/s280 W

More Chips charts

  1. AI chip sales by designer
  2. NVIDIA data center revenue
  3. TSMC monthly revenue
  4. Taiwan AI supply chain revenue
  5. Taiwan's server and chip exports
  6. US server imports by country

All Chips charts

Latest Chips facts

Chips Deal

On September 14, 2026, Anderon finalized the agreement with the U.S. Department of Commerce. The Department made $650 million available at closing and $350 million more will be granted on milestones; IBM contributed $1.0 billion of cash into Anderon.

Anderon LLCIBMU.S. Department of Commerce Source: International Business Machines Corp 10-Q, filed Oct 22, 2026

See every Chips fact