3 ms·

If I could ask a single CEO one question, it would be Lisa Su. What is she thinking? She's sitting on what could be the 2nd most valuable company in the world.

If I could ask a single CEO one question, it would be Lisa Su. What is she thinking? She's sitting on what could be the 2nd most valuable company in the world. And she's doing nothing. AMD's GPUs are as good as NVidia's. But their drivers suck. They did all the hard work. Now they need to invest a relative pittance in fixing up their drivers and open sourcing the rest. But they won't. Absolutely nothing about this situation makes any sense. It's a company that's sitting on gold (actually, by weight, the value of GPUs is far higher than that of solid gold bars). But... they're throwing it all away for nothing? Is she getting paid off by NVidia to do this? AMD is the single most confusing entity out there.

JackYoustra 2y ago
The other day I tried to rent a spot MI-300. I couldn't find a way to do that! Whereas fairly easy to rent an H-100. I think they're just behind a bit on their tsmc allocation! Their stock seems to think so, it's priced much more aggressively than Nvidia's.
- latchkey 2y ago
  My company Hot Aisle is working on solving this problem. It has nothing to do with availability or pricing. You can't rent single MI300x today cause ROCm doesn't support virtualization. PCIe passthrough doesn't work yet. AMD knows about the issue and is working on it. We will have 8xMI300x for rent soon. Once AMD fixes their stuff, we will be able to rent out one at a time.
  - JackYoustra 2y ago
    I see! That’s really cool, I’m quite curious about more about this. Any neat technical things you can share, or is it mostly just business challenges?
    - latchkey 2y ago
      Thanks! Oh it has been a wild ride so far and we just have 8 gpus! We just closed our second SAFE and are about to order 128 more, from Dell this time. That is its own whole story that I posted about on reddit. Cluster size of 16 boxes works well for 128 port 400g switches. Exciting times.
latchkey 2y ago
> And she's doing nothing. This is totally wrong. MI300x was released Dec 2023, just a few months ago. It is a fantastic product, but not into the hands of enough people, yet. Give it time.
- light_hue_1 2y ago
  I can tell you don't work with these cards. The MI300x is bug-ridden junk at the moment. As are all AMD cards. You cannot use them for serious compute.
  - latchkey 2y ago
    Very weird to say something like that. Mine, and all the others who I know who have them, are being used for serious compute.
    - light_hue_1 2y ago
      You use jax and pytorch to develop new models on an MI300x without encountering encountering random bugs? I'm shocked.
      - latchkey 2y ago
        So what I hear you saying in this response is that you do not work with these cards, yet you have some strong opinions. The only bugs we have run into so far are the lack of virtualization and the documented ones in the AMD rocm docs.
        light_hue_1 2y ago
        Before buying H100s we tested MI300s and they were borderline unusable. If you're just running some popular pretrained models out of the box, sure. But you're saying that you develop new model architectures on them? And it works reliably? That's not at all what we saw.

JackYoustra 2y ago

The other day I tried to rent a spot MI-300. I couldn't find a way to do that! Whereas fairly easy to rent an H-100. I think they're just behind a bit on their tsmc allocation! Their stock seems to think so, it's priced much more aggressively than Nvidia's.

latchkey 2y ago
My company Hot Aisle is working on solving this problem. It has nothing to do with availability or pricing. You can't rent single MI300x today cause ROCm doesn't support virtualization. PCIe passthrough doesn't work yet. AMD knows about the issue and is working on it. We will have 8xMI300x for rent soon. Once AMD fixes their stuff, we will be able to rent out one at a time.
- JackYoustra 2y ago
  I see! That’s really cool, I’m quite curious about more about this. Any neat technical things you can share, or is it mostly just business challenges?
  - latchkey 2y ago
    Thanks! Oh it has been a wild ride so far and we just have 8 gpus! We just closed our second SAFE and are about to order 128 more, from Dell this time. That is its own whole story that I posted about on reddit. Cluster size of 16 boxes works well for 128 port 400g switches. Exciting times.

latchkey 2y ago

> And she's doing nothing. This is totally wrong. MI300x was released Dec 2023, just a few months ago. It is a fantastic product, but not into the hands of enough people, yet. Give it time.

light_hue_1 2y ago
I can tell you don't work with these cards. The MI300x is bug-ridden junk at the moment. As are all AMD cards. You cannot use them for serious compute.
- latchkey 2y ago
  Very weird to say something like that. Mine, and all the others who I know who have them, are being used for serious compute.
  - light_hue_1 2y ago
    You use jax and pytorch to develop new models on an MI300x without encountering encountering random bugs? I'm shocked.
    - latchkey 2y ago
      So what I hear you saying in this response is that you do not work with these cards, yet you have some strong opinions. The only bugs we have run into so far are the lack of virtualization and the documented ones in the AMD rocm docs.
      - light_hue_1 2y ago
        Before buying H100s we tested MI300s and they were borderline unusable. If you're just running some popular pretrained models out of the box, sure. But you're saying that you develop new model architectures on them? And it works reliably? That's not at all what we saw.