aschleck

It's been a while since I thought about this but isn't the reason providers advertise only 3.2tbps because that's the limit of a single node's connection to the IB network? DGX is spec'ed to pair each H100 with a Connect-X 7 NIC and those cap out at 400gbps. 8 gpus * 400gbps / gpu = 3.2tbps.

Quiz 2 is confusingly worded but is, iiuc, referring to intranode GPU connections rather than internode networking.

Bazel makes easy things hard and impossible things possible

I had this problem and complained to AT&T support. I know it doesn't make sense to replace the modem because this is a software, not hardware, issue but they very quickly offered to replace the modem and now my DNS isn't getting hijacked. Would recommend trying it!

This is both a good thing and a bad thing, but Pulumi is way more flexible than Terraform. I wanted to have a cloud-provider-specific submodule that created resources (like EKS and GKE) that then exported output values (think kubeconfig), and then I wanted the parent module to pass those in as inputs to a cloud-provider-independent submodule. Terraform couldn't do it without needing to duplicate a ton of code, or without something heinous like Terragrunt (not sure it even would have worked.) Pulumi makes it trivial and in a language I like writing.

Additionally, our applications consume our cloud configuration (eg something that launches pods on heterogenous GPUs needs to know which clusters support which GPUs, our colo cluster has H100s but our Google cluster has A100s etc.) Writing in the same language in the same monorepo makes it very easy to share that state.

I found that using "helm template" to convert every Helm chart into yaml, and then using Pulumi to track changes and update my clusters (with Python transformation functions to get per-cluster configuration) made my life so much better than using Helm. Watching Pulumi or Terraform watch Helm watch Kubernetes update a deployment felt pointlessly complicated.

It's worth noting that just because an H100 has a higher flops number doesn't mean your program is actually hitting that number of flops. Modern TPUs are surprisingly competitive with Nvidia on a perf/$ metric, if you're doing cloud ML they are absolutely worth a look. We have been keeping costs down by racking our own GPUs but TPUs are so cost effective that we need to do some thinking about changing our approach.

I'm not certain but I think part of this is that XLA (for example) is a mountain of chip-specific optimizations between your code and the actual operations. So comparing your throughput between GPU and TPU is not just flops-to-flops.