undefined | Better HN

0 pointsdoctorpangloss1y ago0 comments

Okay, but your GPUs are in ECS. Don't I just want this feature from Amazon, not you, and natively via Nitro? Or even Google has TPU attachments.

> 1. If you're actively developing and need a GPU [for fractional amounts of time]...

Why would I need a GPU for a short amount of time during development? For testing?

I don't get it - what would testing an H100 over a TCP connection tell me? It's like, yeah, I can do that, but it doesn't represent an environment I am going to use for real. Nobody runs applications to GPUs on buses virtualized over TCP connections, so what exactly would I be validating?

0 comments

5 comments · 2 top-level

teaearlgraycold1y ago· 2 in thread

I develop GPU accelerated web apps in an EC2 instance with a remote VSCode session. A lot of the time I’m just doing web dev and don’t need a GPU. I can save thousands per month by switching to this.

amelius1y ago

Sounds like you can save thousands by just buying a simple GPU card.

teaearlgraycold1y ago

Well, for the time being I'm really just burning AWS credits. But you're right! I do however like that my dev machine is the exact same instance type in the same AWS region as my production instances. If I built an equivalent machine it would have different performance characteristics. Often times the AWS VMs have weird behavior that I would otherwise be caught off guard with when deploying to the cloud for the first time.

bmodel1y ago· 1 in thread

I don't believe Nitro would allow you to access a GPU that's not directly connected to the CPU that the VM is running on. So swapping between GPU type or scaling to multiple GPUs is still a problem.

From the developer perspective, you wouldn't know that the H100 is across a network. The experience will be as if your computer is directly attached to an H100. The benefit here is that if you're not actively using the H100 (such as when you're setting up the instance or after the training job completes) you are not paying for the H100.

doctorpanglossOP1y ago

Okay, a mock H100 object would also save me money. I could pretend a 3090 is an A100. “The experience would be that a 3090 is an A100.” Apples to oranges comparison? It’s using a GPU attached to the machine versus a GPU that crosses a VPC boundary. Do you see what I am saying?

I would never run a training job on a GPU virtualized over TCP connection. I would never run a training job that requires 80GB of VRAM on a 24GB VRAM device.

Whom is this for? Who needs to save kopecks on a single GPU who needs H100s?

j / k navigate · click thread line to collapse