5 ms·

>Our first offering is the AMD MI300x instead of Nvidia or Intel products. But why? Is there even a market for this? People want GPUs for DNN training and infe

>Our first offering is the AMD MI300x instead of Nvidia or Intel products. But why? Is there even a market for this? People want GPUs for DNN training and inference and the AMD ecosystem for that is absolute trash. Basically unusable at scale.

latchkey 2y ago
I'd like you to back up that statement with some sort of basis. Given that there are zero reliable and detailed public benchmarks on MI300x, how would you know it is unusable at scale? On the other hand, we've got two hyperscalers offering these things, at scale. https://techcommunity.microsoft.com/t5/azure-high-performance-computing/azure-announces-new-ai-optimized-vm-series-featuring-amd-s/ba-p/3980770 https://techcommunity.microsoft.com/t5/azure-high-performanc... https://www.tomshardware.com/news/amd-scores-two-big-wins-oracle-opts-for-mi300x-ibm-asks-for-fpgas https://www.tomshardware.com/news/amd-scores-two-big-wins-or... Oh and I have 3 other competitors in the smaller CSP realm offering MI300x, that I know of today.
- dumpsterdiver 2y ago
  > the AMD ecosystem for that is absolute trash > there are zero reliable and detailed public benchmarks on MI300x I not very familiar with the space, but as an outsider looking in your datapoint appears to support the parent poster's opinion that the space is not mature. Wouldn't a complete lack of public benchmarks suggest that the ecosystem just isn't there yet?
  - latchkey 2y ago
    They didn't say it wasn't mature, they said it is absolute trash. My point is to agree that it isn't mature. Why isn't it mature? Well... nobody has access to these things. Today, they are only available behind HPC / hyperscaler walls. In order to gain some maturity, someone needs to offer them out in the wild. That's where I come in.
    - dumpsterdiver 2y ago
      Fair enough.
    - humansareok1 2y ago
      I challenge you to try to get it working and train a basic LLM yourself on multiple AMD gpus. It is far worse than the situation with Nvidia GPUs/Drivers and Linux circa the mid 2010s. Which is saying a lot... But don't take my word for it. Geohot, maybe one of the best hackers of our generation can't even it working: https://wccftech.com/tinycorp-ditches-amd-tinybox-ai-package-opt-for-nvidia-or-intel-options/ https://wccftech.com/tinycorp-ditches-amd-tinybox-ai-package... It IS trash and it's clearly NOT mature. These are not mutually exclusive. There have been attempts to get AMD to open source their firmware to let people work on it since they themselves seem disinterested in doing so. Until then AMD is not a real choice for serious ML work.
      - latchkey 2y ago
        Quoting George is a bad look and instantly discredits you as a parrot. Did you know... George has never used a MI300x and refused my multiple offers to let him try mine, calling me an "AMD shill." If you know me personally, that is hilariously the furthest thing from the truth one can get. Again, before you can challenge me, I have already challenged you to back up your claim that "it is trash." Quoting George certainly is not that.

latchkey 2y ago

I'd like you to back up that statement with some sort of basis. Given that there are zero reliable and detailed public benchmarks on MI300x, how would you know it is unusable at scale? On the other hand, we've got two hyperscalers offering these things, at scale. https://techcommunity.microsoft.com/t5/azure-high-performance-computing/azure-announces-new-ai-optimized-vm-series-featuring-amd-s/ba-p/3980770 https://techcommunity.microsoft.com/t5/azure-high-performanc... https://www.tomshardware.com/news/amd-scores-two-big-wins-oracle-opts-for-mi300x-ibm-asks-for-fpgas https://www.tomshardware.com/news/amd-scores-two-big-wins-or... Oh and I have 3 other competitors in the smaller CSP realm offering MI300x, that I know of today.

dumpsterdiver 2y ago
> the AMD ecosystem for that is absolute trash > there are zero reliable and detailed public benchmarks on MI300x I not very familiar with the space, but as an outsider looking in your datapoint appears to support the parent poster's opinion that the space is not mature. Wouldn't a complete lack of public benchmarks suggest that the ecosystem just isn't there yet?
- latchkey 2y ago
  They didn't say it wasn't mature, they said it is absolute trash. My point is to agree that it isn't mature. Why isn't it mature? Well... nobody has access to these things. Today, they are only available behind HPC / hyperscaler walls. In order to gain some maturity, someone needs to offer them out in the wild. That's where I come in.
  - dumpsterdiver 2y ago
    Fair enough.
  - humansareok1 2y ago
    I challenge you to try to get it working and train a basic LLM yourself on multiple AMD gpus. It is far worse than the situation with Nvidia GPUs/Drivers and Linux circa the mid 2010s. Which is saying a lot... But don't take my word for it. Geohot, maybe one of the best hackers of our generation can't even it working: https://wccftech.com/tinycorp-ditches-amd-tinybox-ai-package-opt-for-nvidia-or-intel-options/ https://wccftech.com/tinycorp-ditches-amd-tinybox-ai-package... It IS trash and it's clearly NOT mature. These are not mutually exclusive. There have been attempts to get AMD to open source their firmware to let people work on it since they themselves seem disinterested in doing so. Until then AMD is not a real choice for serious ML work.
    - latchkey 2y ago
      Quoting George is a bad look and instantly discredits you as a parrot. Did you know... George has never used a MI300x and refused my multiple offers to let him try mine, calling me an "AMD shill." If you know me personally, that is hilariously the furthest thing from the truth one can get. Again, before you can challenge me, I have already challenged you to back up your claim that "it is trash." Quoting George certainly is not that.