Compute · AI3Radar

More filters

Manufacturer *Hardware *Compute mode

Clear filters

ANSWER

64 separate series are shown without aggregation.

Coverage: 2017-05-01 – 2026-03-11

TREND

Hardware FP16 peak efficiency

TFLOP/s/W · Log scale

ObservedEstimatedSeparate segments are not joined across breaks

DATA

Comparable data

64 points

Scroll sideways for the complete data.

Hardware FP16 peak efficiency
Date	Subject	Value	Interval	Nature	Method version
5/1/2017	Google TPU v2	0.16 TFLOP/s/W	—	Derived	1
6/21/2017	NVIDIA Tesla V100 PCIe 16 GB	0.45 TFLOP/s/W	—	Derived	1
6/21/2017	NVIDIA Tesla V100 SXM2 16 GB	0.42 TFLOP/s/W	—	Derived	1
6/21/2017	NVIDIA V100	0.42 TFLOP/s/W	—	Derived	1
7/1/2017	NVIDIA Tesla V100 SXM2	0.42 TFLOP/s/W	—	Derived	1
3/27/2018	NVIDIA Tesla V100 DGXS 16 GB	0.5 TFLOP/s/W	—	Derived	1
3/27/2018	NVIDIA Tesla V100 DGXS 32 GB	0.5 TFLOP/s/W	—	Derived	1
3/27/2018	NVIDIA Tesla V100 PCIe 32 GB	0.45 TFLOP/s/W	—	Derived	1
3/27/2018	NVIDIA Tesla V100 SXM2 32 GB	0.42 TFLOP/s/W	—	Derived	1
3/27/2018	NVIDIA Tesla V100 SXM3 32 GB	0.42 TFLOP/s/W	—	Derived	1
5/18/2018	Google TPU v3	0.27 TFLOP/s/W	—	Derived	1
9/20/2018	NVIDIA GeForce RTX 2080 Ti 11GB	0.46 TFLOP/s/W	—	Derived	1
11/26/2019	NVIDIA Tesla V100S PCIe 32 GB	0.52 TFLOP/s/W	—	Derived	1
1/1/2020	Google TPU v4i	0.79 TFLOP/s/W	—	Derived	1
3/1/2020	NVIDIA A100	0.78 TFLOP/s/W	—	Derived	1
5/14/2020	NVIDIA A100 SXM4 40 GB	0.78 TFLOP/s/W	—	Derived	1
6/22/2020	NVIDIA A100 PCIe	1.04 TFLOP/s/W	—	Derived	1
10/5/2020	NVIDIA A40 PCIe	0.5 TFLOP/s/W	—	Derived	1
11/16/2020	NVIDIA A100 SXM4 80 GB	0.78 TFLOP/s/W	—	Derived	1
5/20/2021	Google TPU v4	0.81 TFLOP/s/W	—	Derived	1
8/21/2021	Tesla D1 Dojo	0.91 TFLOP/s/W	—	Derived	1
11/8/2021	AMD Radeon Instinct MI250X	0.77 TFLOP/s/W	—	Derived	1
3/29/2022	NVIDIA GeForce RTX 3090 Ti	0.36 TFLOP/s/W	—	Derived	1
5/10/2022	Intel Habana Gaudi2	0.75 TFLOP/s/W	—	Derived	1
8/11/2022	NVIDIA A800 SXM	0.78 TFLOP/s/W	—	Derived	1
8/22/2022	Biren BR100	1.86 TFLOP/s/W	—	Derived	1
9/20/2022	NVIDIA GeForce RTX 4090	0.73 TFLOP/s/W	—	Derived	1
9/20/2022	NVIDIA H100 PCIe	2.16 TFLOP/s/W	—	Derived	1
9/20/2022	NVIDIA H100 SXM5 80GB	1.41 TFLOP/s/W	—	Derived	1
10/13/2022	NVIDIA L40	0.6 TFLOP/s/W	—	Derived	1
11/3/2022	NVIDIA A800 PCIe 80 GB	1.25 TFLOP/s/W	—	Derived	1
11/8/2022	NVIDIA A800 PCIe 40 GB	1.25 TFLOP/s/W	—	Derived	1
1/4/2023	AMD Instinct MI300A	1.29 TFLOP/s/W	—	Derived	1
3/21/2023	NVIDIA GH100	1.41 TFLOP/s/W	—	Derived	1
3/21/2023	NVIDIA H800 SXM5	1.41 TFLOP/s/W	—	Derived	1
3/21/2023	NVIDIA L4	1.68 TFLOP/s/W	—	Derived	1
5/18/2023	Meta MTIA v1 (MTIA 100)	2.05 TFLOP/s/W	—	Derived	1
8/8/2023	NVIDIA GH200	1.41 TFLOP/s/W	—	Derived	1
8/29/2023	Google TPU v5e	0.88 TFLOP/s/W	—	Derived	1
9/16/2023	NVIDIA L20 PCle	0.43 TFLOP/s/W	—	Derived	1
11/9/2023	NVIDIA HGX H20	0.37 TFLOP/s/W	—	Derived	1
12/6/2023	AMD Instinct MI300X	1.74 TFLOP/s/W	—	Derived	1
12/6/2023	AMD Radeon Instinct MI308X	0.87 TFLOP/s/W	—	Derived	1
12/6/2023	Google TPU v5p	0.85 TFLOP/s/W	—	Derived	1
12/19/2023	MTT S4000	0.22 TFLOP/s/W	—	Derived	1
4/10/2024	Meta MTIA v2 (MTIA 200)	1.97 TFLOP/s/W	—	Derived	1
5/14/2024	Google TPU v6e Trillium	2.42 TFLOP/s/W	—	Derived	1
8/27/2024	Maia 100 (M100)	1.6 TFLOP/s/W	—	Derived	1
9/24/2024	Intel Habana Gaudi3	1.86 TFLOP/s/W	—	Derived	1
10/10/2024	AMD Instinct MI325X	1.31 TFLOP/s/W	—	Derived	1
10/15/2024	Huawei Ascend 910C	1.14 TFLOP/s/W	—	Derived	1
11/15/2024	NVIDIA B200	2.25 TFLOP/s/W	—	Derived	1
11/18/2024	NVIDIA H200 SXM	1.41 TFLOP/s/W	—	Derived	1
12/3/2024	Amazon Trainium2	1.33 TFLOP/s/W	—	Derived	1
2/15/2025	NVIDIA GB200	2.08 TFLOP/s/W	—	Derived	1
3/18/2025	NVIDIA RTX Pro 6000 Blackwell	0.84 TFLOP/s/W	—	Derived	1
6/12/2025	AMD Instinct MI350X	2.31 TFLOP/s/W	—	Derived	1
6/12/2025	AMD Instinct MI355X	1.8 TFLOP/s/W	—	Derived	1
8/22/2025	NVIDIA B300 (Blackwell Ultra)	2.05 TFLOP/s/W	—	Derived	1
8/22/2025	NVIDIA GB300 (Blackwell Ultra)	1.79 TFLOP/s/W	—	Derived	1
11/6/2025	Google TPU v7 Ironwood	2.4 TFLOP/s/W	—	Derived	1
12/2/2025	Amazon Trainium3	0.96 TFLOP/s/W	—	Derived	1
1/26/2026	Microsoft Maia 200	1.69 TFLOP/s/W	—	Derived	1
3/11/2026	Meta MTIA 300	0.75 TFLOP/s/W	—	Derived	1

METHOD

Method

Cadence: Hardware releases
Coverage: 2017-05-01 – 2026-03-11
Value nature: Derived
Version: 1

Theoretical peak from public hardware specifications, not application efficiency