3 ms·

Great article. I'm in the process of building this business myself, so I'm intimately aware of everything in the article. Just keep in mind that a lot of the ma

Great article. I'm in the process of building this business myself, so I'm intimately aware of everything in the article. Just keep in mind that a lot of the math is back of the napkin guesses. All of this is managed on personal relationships and deals with the vendors and much of the pricing that actually gets set, isn't made public. Our first offering is the AMD MI300x instead of Nvidia or Intel products. But, unlike all of my competitors, we are not picking sides. Our unique value prop is that we eventually plan to offer the best of the best compute of anything our customers want. Effectively building a varied super computer for rent, with bare metal access to everything you need and white glove service and support. In essence, we are the cap/opex for businesses that don't want to take on this risk themselves. What people don't understand is how difficult and capital intensive to deploy and manage large scale compute, especially on the high end cutting edge. Gone are the days of just racking/stacking a few servers. This stuff is way more complicated and involved. It is rife with firmware bugs, limitations, and hardware failures. The end of the article says some nonsense about a glut of GPU capacity. I do have to call that out. It isn't going to happen for a long while at least. The need for compute is growing exponentially. Given the complexities of just deploying this stuff, it isn't physically possible to get enough of it out there, fast enough, to satisfy the demand. I love every challenge of this business. Happy to answer questions where I can.

jdewerd 3y ago
> It is rife with firmware bugs, limitations, and hardware failures. Is it? I've been out of the space for a while but I have only seen firmware bugs/limitations/failures rise to strategic importance with AMD installs, which is why the space has such a strong preference for NVidia despite spec sheet cost disadvantage. > Our first offering is the AMD MI300x instead of Nvidia or Intel products. Oh no. I'm sorry.
- latchkey 3y ago
  > Is it? I've been out of the space for a while but I have only seen firmware bugs/limitations/failures rise to strategic importance with AMD installs, which is why the space has such a strong preference for NVidia despite spec sheet cost disadvantage. You're right. I'm not doing this the easy way, at all! Is there any easy way though? Without a direct connection to Jensen, there is a 50+ week lead time on Infiniband, so what other options do I have? Who wants to compete with $1.1b worth of orders from CoreWeave? I'm fine with the bet on Lisa Su and I'm not trying to beat CoreWeave. I'd just like to carve out my own niche and I think there are enough people out there that want compute, that they are willing to look beyond just a single provider for everything. What fortune 500 puts all their eggs in one basket anyway? Oh and one more thing... this business isn't just about AMD/Nvidia. There are 1000 other components in the system, all with their own issues. We just discovered VRF is not fully working on our very expensive 400G switches running Sonic. Cheers.
  - jdewerd 3y ago
    > Without a direct connection to Jensen, there is a 50+ week lead time on Infiniband Oof, that's new to me, sounds like you're between a rock and a hard place. Here's hoping you get the wrinkles ironed out and make it to the other side, the ecosystem will be certainly be better for it. Cheers.
humansareok1 3y ago
>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 3y 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 3y 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 3y 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 3y ago

jdewerd 3y ago

> It is rife with firmware bugs, limitations, and hardware failures. Is it? I've been out of the space for a while but I have only seen firmware bugs/limitations/failures rise to strategic importance with AMD installs, which is why the space has such a strong preference for NVidia despite spec sheet cost disadvantage. > Our first offering is the AMD MI300x instead of Nvidia or Intel products. Oh no. I'm sorry.

latchkey 3y ago
> Is it? I've been out of the space for a while but I have only seen firmware bugs/limitations/failures rise to strategic importance with AMD installs, which is why the space has such a strong preference for NVidia despite spec sheet cost disadvantage. You're right. I'm not doing this the easy way, at all! Is there any easy way though? Without a direct connection to Jensen, there is a 50+ week lead time on Infiniband, so what other options do I have? Who wants to compete with $1.1b worth of orders from CoreWeave? I'm fine with the bet on Lisa Su and I'm not trying to beat CoreWeave. I'd just like to carve out my own niche and I think there are enough people out there that want compute, that they are willing to look beyond just a single provider for everything. What fortune 500 puts all their eggs in one basket anyway? Oh and one more thing... this business isn't just about AMD/Nvidia. There are 1000 other components in the system, all with their own issues. We just discovered VRF is not fully working on our very expensive 400G switches running Sonic. Cheers.
- jdewerd 3y ago
  > Without a direct connection to Jensen, there is a 50+ week lead time on Infiniband Oof, that's new to me, sounds like you're between a rock and a hard place. Here's hoping you get the wrinkles ironed out and make it to the other side, the ecosystem will be certainly be better for it. Cheers.

humansareok1 3y ago

>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 3y 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 3y 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 3y 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 3y ago