Which AI chip has the best performance-per-watt in 2026?

Chip leaderboard

Best perf-per-watt (TFLOPS/W)

NVIDIA Jetson AGX Thor leads at 15.9 TFLOPS/W. 43 chips ranked.

How much work each chip does per watt. Straight from the vendor datasheet: best dense throughput divided by rated power. Wafer-scale and rack-scale chips look artificially strong because their datasheet numbers roll in shared overhead the per-chip numbers ignore.

#	Chip	Value	Inputs
1	NVIDIA Jetson AGX Thor NVIDIA	15.9 TFLOPS/W	FP8 TFLOPS: 2,070 · TDP: 130 W
2	NVIDIA DRIVE Thor NVIDIA	15.4 TFLOPS/W	FP8 TFLOPS: 2,000 · TDP: 130 W
3	Google TPU v7 (Ironwood) Google	13.2 TFLOPS/W	FP8 TFLOPS: 9,228 · TDP: 700 W
4	Google TPU v6 (Trillium) Google	9.18 TFLOPS/W	FP8 TFLOPS: 1,836 · TDP: 200 W
5	NVIDIA B100 (Blackwell) NVIDIA	5 TFLOPS/W	FP8 TFLOPS: 3,500 · TDP: 700 W
6	AMD Instinct MI350 series AMD	4.6 TFLOPS/W	FP8 TFLOPS: 4,600 · TDP: 1000 W
7	NVIDIA B200 (Blackwell) NVIDIA	4.5 TFLOPS/W	FP8 TFLOPS: 4,500 · TDP: 1000 W
8	AMD Instinct MI355X AMD	3.57 TFLOPS/W	FP8 TFLOPS: 5,000 · TDP: 1400 W
9	NVIDIA B300 (Blackwell Ultra) NVIDIA	3.57 TFLOPS/W	FP8 TFLOPS: 5,000 · TDP: 1400 W
10	NVIDIA GB300 (Blackwell Ultra) NVIDIA	3.57 TFLOPS/W	FP8 TFLOPS: 5,000 · TDP: 1400 W
11	AMD Instinct MI300X AMD	3.49 TFLOPS/W	FP8 TFLOPS: 2,614 · TDP: 750 W
12	NVIDIA GB200 Grace Blackwell Superchip NVIDIA	3.33 TFLOPS/W	FP8 TFLOPS: 9,000 · TDP: 2700 W
13	Huawei Ascend 910D Huawei	3 TFLOPS/W	FP8 TFLOPS: 2,400 · TDP: 800 W
14	NVIDIA H100 Tensor Core GPU NVIDIA	2.83 TFLOPS/W	FP8 TFLOPS: 1,979 · TDP: 700 W
15	NVIDIA H200 Tensor Core GPU NVIDIA	2.83 TFLOPS/W	FP8 TFLOPS: 1,979 · TDP: 700 W
16	NVIDIA H800 NVIDIA	2.83 TFLOPS/W	FP8 TFLOPS: 1,979 · TDP: 700 W
17	Cerebras WSE-3 (Wafer-Scale Engine 3) Cerebras Systems	2.72 TFLOPS/W	FP16 TFLOPS: 62,500 · TDP: 23000 W
18	AMD Instinct MI325X AMD	2.61 TFLOPS/W	FP8 TFLOPS: 2,614 · TDP: 1000 W
19	AWS Trainium 2 Amazon	2.6 TFLOPS/W	FP8 TFLOPS: 1,300 · TDP: 500 W
20	Tenstorrent Blackhole Tenstorrent	2.48 TFLOPS/W	FP16 TFLOPS: 745 · TDP: 300 W
21	NVIDIA L40S NVIDIA	2.09 TFLOPS/W	FP8 TFLOPS: 733 · TDP: 350 W
22	Intel Gaudi 3 Intel Corporation	2.04 TFLOPS/W	FP8 TFLOPS: 1,835 · TDP: 900 W
23	Biren BR100 Biren Technology	1.86 TFLOPS/W	FP16 TFLOPS: 1,024 · TDP: 550 W
24	SambaNova SN40L SambaNova Systems	1.79 TFLOPS/W	FP8 TFLOPS: 1,250 · TDP: 700 W
25	Huawei Ascend 910C Huawei	1.6 TFLOPS/W	FP16 TFLOPS: 800 · TDP: 500 W
26	Google TPU v5p Google	1.53 TFLOPS/W	BF16 TFLOPS: 459 · TDP: 300 W
27	AWS Inferentia 2 Amazon	1.52 TFLOPS/W	FP8 TFLOPS: 380 · TDP: 250 W
28	Google TPU v4 Google	1.43 TFLOPS/W	BF16 TFLOPS: 275 · TDP: 192 W
29	Google TPU v5e Google	1.16 TFLOPS/W	BF16 TFLOPS: 197 · TDP: 170 W
30	Huawei Ascend 910B Huawei	1 TFLOPS/W	FP16 TFLOPS: 400 · TDP: 400 W
31	NVIDIA T4 Tensor Core GPU NVIDIA	0.929 TFLOPS/W	FP16 TFLOPS: 65 · TDP: 70 W
32	Tenstorrent Wormhole Tenstorrent	0.925 TFLOPS/W	FP16 TFLOPS: 148 · TDP: 160 W
33	Tesla Dojo D1 Tesla	0.905 TFLOPS/W	FP8 TFLOPS: 362 · TDP: 400 W
34	NVIDIA A10 Tensor Core GPU NVIDIA	0.833 TFLOPS/W	FP16 TFLOPS: 125 · TDP: 150 W
35	NVIDIA A100 Tensor Core GPU NVIDIA	0.78 TFLOPS/W	FP16 TFLOPS: 312 · TDP: 400 W
36	NVIDIA A800 NVIDIA	0.78 TFLOPS/W	FP16 TFLOPS: 312 · TDP: 400 W
37	NVIDIA H20 NVIDIA	0.74 TFLOPS/W	FP8 TFLOPS: 296 · TDP: 400 W
38	Intel Gaudi 2 Intel Corporation	0.72 TFLOPS/W	BF16 TFLOPS: 432 · TDP: 600 W
39	AMD Instinct MI250X AMD	0.684 TFLOPS/W	FP16 TFLOPS: 383 · TDP: 560 W
40	Groq LPU (Language Processing Unit) Groq	0.501 TFLOPS/W	FP16 TFLOPS: 188 · TDP: 375 W
41	AWS Trainium (Trn1) Amazon	0.42 TFLOPS/W	BF16 TFLOPS: 210 · TDP: 500 W
42	NVIDIA V100 Tensor Core GPU NVIDIA	0.417 TFLOPS/W	FP16 TFLOPS: 125 · TDP: 300 W
43	Cambricon MLU370-X8 Cambricon Technologies	0.384 TFLOPS/W	FP16 TFLOPS: 96 · TDP: 250 W

43 chips ranked. 17 more are in our catalog but don't publish the numbers needed for this ranking.

Other rankings: Most HBM memory per chip · Highest memory bandwidth · Highest FP8 dense throughput · Highest FP16 dense throughput · Highest INT8 dense throughput · Cheapest per TFLOP at launch (list price) · Most transistors per die