Unprecedented
Performance
and Scalability
Performance of HyperAccel LPU
OPT 66B
OPT 30B
OPT 6.7B
OPT 1.3B
23.6
46.5
175.8
520.9
tokens/sec
8x LPU
8:2040
Efficiency Analysis of
HyperAccel LPU vs GPU Platform
Power Efficiency
38.8

8x HyperAccel LPU

29.5

2x NVIDIA H100

1.32x
tokens/sec/kW
OPT 60B
8:2040
Cost Efficiency
1841

16x HyperAccel LPU

783

NVIDIA DGX A100

2.35x
tokens/sec/mil$
GPT 20B
8:2040
Scalability Analysis of
HyperAccel LPU vs. GPU Platform
1.8X
1.74X
1.72X
1.8X
1.72X
1.74X