"System Memory per GPU" filter doesn't work for more than a single GPU
Select 100gb per GPU, 16 vCPUs per GPU. Select 5090 GPU. Note Pod Summary fulfilling criteria. Increase GPU Count to 2. Note that the amount of RAM in the Pod Summary decreases from 116GB RAM to 108GB RAM. Change region (to SE in this case). Note that the pod Summary now shows 232GB RAM - proving that a Pod with 100Gb+ per GPU actually exists.
4 Replies
Each data center has slight variations of hardware. The filter just make sure the machine is meet your requirements
But if I select 100Gb System memory per GPU, and select 2 GPUs I expect to get a pod with 2 * 100 = 200Gb RAM at least. Not 108Gb RAM.
It does not meet the 100Gb RAM per GPU requirement I mean.
Checking the GraphQL payload, I can see that the number of vCPUs (field
minVcpuCount ) are correctly multiplied with the number of GPUs, but the minMemoryInGb is not. Pasting another example with 4 GPUs and min 12 vCPUs per GPU and min 100Gb RAM per GPU. The minMemoryInGb field should be 400 in this case, and not 100.
@yhlong00000 , are you with me?