5 ms·

What moat does Nvidia have. AMD could have ROCm perfected if they really want to. Also most of pytorch, specially those relevant to transformers runs perfectly

What moat does Nvidia have. AMD could have ROCm perfected if they really want to. Also most of pytorch, specially those relevant to transformers runs perfectly on Apple Silicon and TPUs and probably other hardware as well. If anyone has moat related to Gen AI, I would say it is the data(Google, Meta).

klelatti 2y ago
> AMD could have ROCm perfected if they really want to. It's not an act of will or CEO dictat. It's about hiring and incentivising the right people, putting the right structures in place etc all in the face of competing demands. Nvidia have a huge head start and by the time AMD have 'caught up' Nvidia with it's greater resources will have moved further ahead.
- YetAnotherNick 2y ago
  If head start is a moat, why wouldn't you count OpenAI's headstart as moat?
  - echelon 2y ago
    Anyone can make an LLM. There are hundreds of choices in the market today. Many of them are even open source. OpenAI brings absolutely nothing unique to the table.
  - klelatti 2y ago
    Because we already see firms competing effectively with OpenAI. There is as yet no indication that AMD can match Nvidia's execution for the very good reason that doing so is extremely difficult. The head start is just the icing on the cake.
    - YetAnotherNick 2y ago
      I see pre AGI AI to be superset of search and its hard to argue that google isn't ahead technically for decades and what they are doing is extremely difficult.
  - PKop 2y ago
    Not all industries or product segments are equal is the obvious answer. The point here whether one agrees or not is models are easier to catch up to than GPUs
  - dragonwriter 2y ago
    A head start is a moat iff you can't move easily from the leader to a competitor that catches up qualitatively; nvidia’s headstart against AMD is a moat to the extent that you can't just take the software written against NVidia GPUs and run it on AMD if AMD catches up. (That is, being currently ahead isn't a moat, but it can impose switching costs which are.) Taking code that runs against one hosted LLM and running it against a different backend LLM is... not generally a big deal. So OpenAI being ahead—in the core model, at least—is just being ahead, its not a moat.
    - YetAnotherNick 2y ago
      ROCM is close to being perfected and I think in few years and some investment you can use it directly to run 99% software written in CUDA with similar performance.
      - klelatti 2y ago
        Since 2007 Nvidia has built with CUDA and compatible hardware - a large and growing ecosystem - a massive installed base of backwards compatible hardware - the ability to massively scale delivery of new systems - and lots lots more They now have scale that enables them to continue to invest at a level that no competition can do. None of these are easily reproduced. As per SemiAnalysis AMD in late 2024 can’t get essential software working reliably out of the box. It’s easy to say AMD ‘is close to perfecting ROCM’ the reality of competing with Nvidia is much harder.
        YetAnotherNick 2y ago
        It's hard but not $3T hard. AMD has consistently underinvested in their software. If they invest double digit million to make installation and usage of drivers and ROCM smoother I highly doubt they can't acheive it. There are open source projects volunteer run projects[1] that are better than official AMD implementation in many ways. [1]: https://github.com/CHIP-SPV/chipStar/ https://github.com/CHIP-SPV/chipStar/
        2y ago

klelatti 2y ago

> AMD could have ROCm perfected if they really want to. It's not an act of will or CEO dictat. It's about hiring and incentivising the right people, putting the right structures in place etc all in the face of competing demands. Nvidia have a huge head start and by the time AMD have 'caught up' Nvidia with it's greater resources will have moved further ahead.

YetAnotherNick 2y ago
If head start is a moat, why wouldn't you count OpenAI's headstart as moat?
- echelon 2y ago
  Anyone can make an LLM. There are hundreds of choices in the market today. Many of them are even open source. OpenAI brings absolutely nothing unique to the table.
- klelatti 2y ago
  Because we already see firms competing effectively with OpenAI. There is as yet no indication that AMD can match Nvidia's execution for the very good reason that doing so is extremely difficult. The head start is just the icing on the cake.
  - YetAnotherNick 2y ago
    I see pre AGI AI to be superset of search and its hard to argue that google isn't ahead technically for decades and what they are doing is extremely difficult.
- PKop 2y ago
  Not all industries or product segments are equal is the obvious answer. The point here whether one agrees or not is models are easier to catch up to than GPUs
- dragonwriter 2y ago
  A head start is a moat iff you can't move easily from the leader to a competitor that catches up qualitatively; nvidia’s headstart against AMD is a moat to the extent that you can't just take the software written against NVidia GPUs and run it on AMD if AMD catches up. (That is, being currently ahead isn't a moat, but it can impose switching costs which are.) Taking code that runs against one hosted LLM and running it against a different backend LLM is... not generally a big deal. So OpenAI being ahead—in the core model, at least—is just being ahead, its not a moat.
  - YetAnotherNick 2y ago
    ROCM is close to being perfected and I think in few years and some investment you can use it directly to run 99% software written in CUDA with similar performance.
    - klelatti 2y ago
      Since 2007 Nvidia has built with CUDA and compatible hardware - a large and growing ecosystem - a massive installed base of backwards compatible hardware - the ability to massively scale delivery of new systems - and lots lots more They now have scale that enables them to continue to invest at a level that no competition can do. None of these are easily reproduced. As per SemiAnalysis AMD in late 2024 can’t get essential software working reliably out of the box. It’s easy to say AMD ‘is close to perfecting ROCM’ the reality of competing with Nvidia is much harder.
      - YetAnotherNick 2y ago
        It's hard but not $3T hard. AMD has consistently underinvested in their software. If they invest double digit million to make installation and usage of drivers and ROCM smoother I highly doubt they can't acheive it. There are open source projects volunteer run projects[1] that are better than official AMD implementation in many ways. [1]: https://github.com/CHIP-SPV/chipStar/ https://github.com/CHIP-SPV/chipStar/
        2y ago