Cloud Cost Analysis

AWS

GCloud GPUs

V100

P100

K80

$3.06 per GPU (p3.2xlarge)

- Spot price: $0.99 USD per GPU (variable)

- 16 GB memory per GPU

- up to 8 GPUs per instance

$0.90 per GPU (p2.xlarge)

- Spot price: $0.27 USD per GPU (variable)

- 12 GB memory per GPU

- up to 16 GPUs per instance

$3.10 per GPU

- an equivalent custom configured Google instance starts at $3.10, but then can get down to $2.19 with "sustained use" discounts

- Preemptible price: $0.74 USD per GPU

- 16 GB HBM2 memory per GPU

- up to 8 GPUs per instance

$1.46 per GPU

- Preemptible price: $0.43 USD per GPU

- 16 GB HBM2 memory per GPU

- up to 4 GPUs per instance

$0.62 per GPU

- an equivalent custom configured Google instance starts at $0.62, but then can get down to $0.45 with "sustained use" discounts

- Preemptible price: $0.135 USD per GPU

- 12 GB HBM2 memory per GPU

- up to 8 GPUs per instance

Pricing

link (calculator)

Pricing

link

Speedup

- Depends on the workload. Larger models perform better on v100.

- Also to note, V100 has more GPU memory (so will see better performance on memory bound models)

- 1.7x for smaller

- 3.1x for larger model (Resnet)

(source)

(not offered)

-

Speedup

- variation of the speedup across applications can be explained by the different application characteristics, in particular the relation of compute instructions to memory access operations.

source

M60

$1.14 per GPU (g3.4xlarge)

Description