Closed dchaley closed 4 months ago
Test on Lynn's GCP project - get project info from Lynn
Did test (reported results in sheet) w/2 and 4 NIVDIA T4 GPUs. Oddly better perf w/2 GPUs of this type (than 4). Other GPUs to test V100 (2x cost of T4s) and possibly TPU v2 Pod (512 cores) 16GB RAM (costs 1.5x T4 GPU)
Suspect that we need to verify the CUDA drivers and configuration on the runtime used - WIP to test...
Don't think we'll need 4 GPUs - testing with 1 or 2 GPUs max - so closing this one
In GCP / VertexAI, GCP types are associated with GCE instance types, for example A2 instances ALL include 1 or more GPUs. Other instance types allow for configuration of 1 or more of 1 or more GPU types.
Generally there is span of 7x in pricing between lowest and highest end available GPUs so best-fit configuration of instance type, type and quantity of GPU(s) is key to TCO. Also, GPU pricing (for most GPUs) is variable by pricing tier (on-demand, SPOT or reserved) and account type (enterprise accounts often include pricing discounts).
Finally, each GCP project includes a base level of quota for each type of GPU available - this is 0 or 1 GPU of each type per service for new accounts by default. While Google can adjust quotas, it is at their discretion as to when they will do so.