Humanity Statsv1
Live
Top 50

Die energieeffizientesten KI-Chips

Sortiert nach Rechenleistung je verbrauchtem Watt.

Microsoft Maia 200 leads with 1.4 x 10^13 OP/s per W, ahead of NVIDIA B300 (Blackwell Ultra) (1.3 x 10^13 OP/s per W) and NVIDIA GB300 (Blackwell Ultra) (1.1 x 10^13 OP/s per W).

Efficiency has improved by roughly a third per year across generations. It compounds with the power constraint: at a fixed grid connection, a more efficient chip is directly more compute, which is why efficiency gains show up in data-center capacity rather than in electricity bills.

#NameManufacturerReleasedEfficiency
01Microsoft Maia 200MicrosoftJan 26, 20261.4×10¹³
02NVIDIA B300 (Blackwell Ultra)NVIDIAAug 22, 20251.3×10¹³
03NVIDIA GB300 (Blackwell Ultra)NVIDIAAug 22, 20251.1×10¹³
04AMD Instinct MI350XAMDJun 12, 20259.2×10¹²
05NVIDIA B200NVIDIANov 15, 20249×10¹²
06NVIDIA GB200NVIDIAFeb 15, 20258.3×10¹²
07AMD Instinct MI355XAMDJun 12, 20257.2×10¹²
08NVIDIA L4NVIDIAMar 21, 20236.7×10¹²
09Google TPU v6e TrilliumGoogleMay 14, 20244.8×10¹²
10Google TPU v7 IronwoodGoogleNov 6, 20254.8×10¹²
11Meta MTIA v1 (MTIA 100)MetaMay 18, 20234.1×10¹²
12Meta MTIA v2 (MTIA 200)MetaApr 10, 20243.9×10¹²
13Biren BR100BirenAug 22, 20223.7×10¹²
14Cambricon MLU270-S4CambriconJun 20, 20193.7×10¹²
15Amazon Trainium3Amazon AWSDec 2, 20253.6×10¹²
16GroqChip LPU v1—Oct 15, 20223.5×10¹²
17AMD Instinct MI300XAMDDec 6, 20233.5×10¹²
18NVIDIA RTX Pro 6000 BlackwellNVIDIAMar 18, 20253.4×10¹²
19T-Head Hanguang 800AlibabaSep 25, 20193×10¹²
20NVIDIA H200 SXMNVIDIANov 18, 20242.8×10¹²
21NVIDIA GH200NVIDIAAug 8, 20232.8×10¹²
22NVIDIA H100 SXM5 80GBNVIDIASep 20, 20222.8×10¹²
23NVIDIA GH100NVIDIAMar 21, 20232.8×10¹²
24NVIDIA H800 SXM5NVIDIAMar 21, 20232.8×10¹²
25AMD Instinct MI325XAMDOct 10, 20242.6×10¹²
26Amazon Trainium2Amazon AWSDec 3, 20242.6×10¹²
27AMD Instinct MI300AAMDJan 4, 20232.6×10¹²
28Cambricon Siyuan 590CambriconOct 31, 20242.6×10¹²
29Cambricon MLU370-S4/S8CambriconNov 3, 20212.6×10¹²
30Cambricon MLU370-S8CambriconNov 3, 20212.6×10¹²
31Cambricon MLU370-S4CambriconNov 3, 20212.6×10¹²
32NVIDIA A800 PCIe 80 GBNVIDIANov 3, 20222.5×10¹²
33NVIDIA H100 PCIeNVIDIASep 20, 20222.2×10¹²
34Huawei Ascend 910CHuaweiOct 15, 20242.1×10¹²
35NVIDIA H100 NVLNVIDIAMar 14, 20242.1×10¹²
36NVIDIA A100 PCIeNVIDIAJun 22, 20202.1×10¹²
37Huawei Ascend 910 (320 TFLOPS)HuaweiAug 23, 20192.1×10¹²
38Huawei Ascend 310HuaweiOct 10, 20182×10¹²
39Baidu Kunlun RG800Kunlunxin BaiduJan 1, 20232×10¹²
40Intel Habana Gaudi3IntelSep 24, 20241.9×10¹²
41Google TPU v5eGoogleAug 29, 20231.7×10¹²
42Cambricon MLU270-F4CambriconJun 20, 20191.7×10¹²
43Baidu Kunlun R200 (150W)Kunlunxin BaiduJan 1, 20221.7×10¹²
44Baidu Kunlun IKunlunxin BaiduJan 1, 20191.7×10¹²
45Baidu Kunlun R100 (75W)Kunlunxin BaiduJan 1, 20221.7×10¹²
46Baidu Kunlun IIKunlunxin BaiduAug 18, 20211.7×10¹²
47Cambricon MLU370-X4CambriconNov 3, 20211.7×10¹²
48Baidu Kunlun K200Kunlunxin BaiduJan 1, 20201.7×10¹²
49Baidu Kunlun R100 (100W)Kunlunxin BaiduJan 1, 20221.7×10¹²
50Google TPU v5pGoogleDec 6, 20231.7×10¹²
SourceEpoch AI, 'AI Models'. Published online at epoch.ai. Retrieved 2026-07-29 from https://epoch.ai/data/ai-models. Licensed under CC BY 4.0.
Browse all hardware →
Other rankings of hardware
Die schnellsten KI-Chips