ABCDEFGHIJKLMNOPQRSTUVWXYZ
1
h/t Lukas Finvedden for help with this.
2
Training FLOP (physical FLOP)
3
3.00E+25
4
k
5
1.5
If there are p parameters, a forward pass takes k*p FLOP. Bio Anchors uses k=1.5
6
Train model for mp tokens
7
20Chinchilla
8
Horizon length
9
1.5
Longer horizons increase training FLOP without increasing runtime FLOP/s.
10
p
11
4.71E+11
Training takes 3hkmp^2 FLOP. [Why the factor of 3? Training takes (FLOP for forward pass)+(FLOP for backward pass)~=3x FLOP for forward pass]
12
Runtime FLOP/forward-pass of AI
13
7.07E+11
14
Duration of training (seconds)
15
1.00E+074 months
16
17
How many tokens per second, by reallocating training compute to inference?
18
4.24E+06
19
20
Token per second, per task
21
10
22
Number of tasks in parallel
23
4.24E+05
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