ABCDEFGHIJKLMNOPQRSTUVWXYZAA
1
Baseline (GPT-4o)Explanation GPT-4o with long outputSmaller model (e.g. 4o-mini)Pessimistic (Larger model + worse utilization)Pessimistic + long outputs
2
Active parameters (billions)100Using 400 billion total, 100B active10020200200
3
Output tokens per query500Chiang et al found an average response of 269 tokens in Chatbot Arena.
Actual tokens generated can vary dramatically, in both directions, based on prompt and usage 		15005005001500
4
FLOP per query 1.00E+142 * active params * tokens3.00E+142.00E+132.00E+146.00E+14
5
GPU peak FLOP/s9.89E+14
H100 peak FLOP/s (bf16 without sparsity. but inference could use fp8 instead)
9.89E+149.89E+149.89E+149.89E+14
6
GPU FLOP utilization0.10.10.10.050.05
7
GPU time in seconds required to process a query1.013.030.204.0412.13
8
Data center power per GPU (W)1275DGX H100 server with 8 GPUs is rated at 10.2 kW, or 1275 W per GPU1275127512751275
9
Data center power-use effectiveness (PUE)1.2Global multiplier at the data center level due to non-compute equipment. Ranges from 1.1 to 1.3 for AI data centers per SemiAnalysis1.21.21.31.3
10
Power adjustment for partial utilization0.7Average consumption found by Microsoft0.70.70.70.7
11
Average GPU power (W)1071Adjust GPU power by power adjustment, add node-level consumption, multiply by PUE107110711160.251160.25
12
Energy per query (Watt-hours)0.3010.9020.061.3043.911
13
14
15
NOTE: cost of queries with long inputs are calculated separately in this notebook: Gradient Updates Chatgpt query flop cost.ipynb
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100