4 ms·

It is already a highly speculated bubble, the share price is swinging drastically within the same day. That being said, Nvidia will have a good 2-3 years in t

It is already a highly speculated bubble, the share price is swinging drastically within the same day. That being said, Nvidia will have a good 2-3 years in the future, until the cloud providers start mass replace their chips with in-house ones. For money, unfortunately, it is as good as anyone's guess, but if you join now, double your share, it will be close to 4T. I personally didn't think that is sustainable amount of money if market/competition does their duty.

chatmasta 3y ago
> Nvidia will have a good 2-3 years in the future, until the cloud providers start mass replace their chips with in-house ones How do you imagine that happening on such a rapid timeline? IMO, the biggest threat to Nvidia is radical innovation on the software side of AI that allows common use cases (training and inference) to run on clusters of commodity non-GPU hardware for cheaper price than GPU. A similar threat is posed from ASIC chips that are dedicated to LLM tasks, but at least in that case Nvidia can still compete for a specific hardware design rather than overcoming an entire category of commodity hardware.
- karmasimida 3y ago
  Simply because of there is too much money on the table, news from today: https://qz.com/intel-microsoft-new-chip-semiconductor-nvidia-1851275354 https://qz.com/intel-microsoft-new-chip-semiconductor-nvidia... NVDA's margin is too high, it attracts competitors to eat into that.
  - chatmasta 3y ago
    But how are competitors going to magically start creating market leading specialized hardware in 2-3 years? Manufacturing chips is not a core competency of any of the cloud providers, and to the extent they've been trying to make it one, they're focused on CPU.
    - karmasimida 3y ago
      Well there is TPU, and it is at least as good as GPU for training, all Google's models are trained TPU, it is the best class of training chips in industry. Second, there is no need to be market leading, it just needs to be cost efficiently to be successful, even if it is slower, as long as it is cheaper. All cloud providers have huge chip design/manufacturing history, AWS/Azure and Google. I would argue it IS indeed the core competency for those cloud providers, because of price. The inhouse chips exist because they can reduce the cost. I know for a fact, there is huge momentum in some big clouds to replace Nvidia GPUs with their own chips, at least for their internal compute needs at first.
    - HDThoreaun 3y ago
      All the cloud megasclaers already have in house chip design teams that have released successful products. Designing a GPU is easier than a cpu, they just need to figure out CUDA.
      - chatmasta 3y ago
        They have some small manufacturing operations that are producing a minority of the chips they actually use in their servers. Maybe they'll be able to produce enough GPUs for their use cases eventually, but there is no way they'll get there in "2-3 years" as OP suggests. Also, if GPUs are easier to make than CPUs, then why isn't Intel leading the market in GPUs by now? Surely it's been clear they should compete with Nvidia there for at least the past half a decade. Or do they need more than 2-3 years to catch up?
        HDThoreaun 3y ago
        They dont manufacture themselves so ramping up is as simple as "here's another billion build 10 times as many." If they design GPUs that they think have significant cost advantages without downsides they will absolutely make that call.
        chatmasta 3y ago
  - seec 3y ago
    Considering how long it took apple to design a competent chip I think Nvidia will be fine for quite a while. Really their current hardware design is unparalleled and they even have some software lock-in with CUDA…
- godelski 3y ago
  > A similar threat is posed from ASIC chips that are dedicated to LLM tasks, but at least in that case Nvidia can still compete for a specific hardware design rather than overcoming an entire category of commodity hardware. But also consider the reason Nvidia is doing so well in the first place. The reason is because they were already developing hardware that specialized at performing tensor operations. Maybe not at the size or always the same types we see in ML, but not very far either. There's reasons why they aren't filling cards with tensor cores. But they are clearly developing specialized hardware for ML tasks. I'm not sure betting on LLMs sticking around is the best bet though. I'll be very impressed if scale is all you need.

chatmasta 3y ago

> Nvidia will have a good 2-3 years in the future, until the cloud providers start mass replace their chips with in-house ones How do you imagine that happening on such a rapid timeline? IMO, the biggest threat to Nvidia is radical innovation on the software side of AI that allows common use cases (training and inference) to run on clusters of commodity non-GPU hardware for cheaper price than GPU. A similar threat is posed from ASIC chips that are dedicated to LLM tasks, but at least in that case Nvidia can still compete for a specific hardware design rather than overcoming an entire category of commodity hardware.

karmasimida 3y ago
Simply because of there is too much money on the table, news from today: https://qz.com/intel-microsoft-new-chip-semiconductor-nvidia-1851275354 https://qz.com/intel-microsoft-new-chip-semiconductor-nvidia... NVDA's margin is too high, it attracts competitors to eat into that.
- chatmasta 3y ago
  But how are competitors going to magically start creating market leading specialized hardware in 2-3 years? Manufacturing chips is not a core competency of any of the cloud providers, and to the extent they've been trying to make it one, they're focused on CPU.
  - karmasimida 3y ago
    Well there is TPU, and it is at least as good as GPU for training, all Google's models are trained TPU, it is the best class of training chips in industry. Second, there is no need to be market leading, it just needs to be cost efficiently to be successful, even if it is slower, as long as it is cheaper. All cloud providers have huge chip design/manufacturing history, AWS/Azure and Google. I would argue it IS indeed the core competency for those cloud providers, because of price. The inhouse chips exist because they can reduce the cost. I know for a fact, there is huge momentum in some big clouds to replace Nvidia GPUs with their own chips, at least for their internal compute needs at first.
  - HDThoreaun 3y ago
    All the cloud megasclaers already have in house chip design teams that have released successful products. Designing a GPU is easier than a cpu, they just need to figure out CUDA.
    - chatmasta 3y ago
      They have some small manufacturing operations that are producing a minority of the chips they actually use in their servers. Maybe they'll be able to produce enough GPUs for their use cases eventually, but there is no way they'll get there in "2-3 years" as OP suggests. Also, if GPUs are easier to make than CPUs, then why isn't Intel leading the market in GPUs by now? Surely it's been clear they should compete with Nvidia there for at least the past half a decade. Or do they need more than 2-3 years to catch up?
      - HDThoreaun 3y ago
        They dont manufacture themselves so ramping up is as simple as "here's another billion build 10 times as many." If they design GPUs that they think have significant cost advantages without downsides they will absolutely make that call.
        chatmasta 3y ago
- seec 3y ago
  Considering how long it took apple to design a competent chip I think Nvidia will be fine for quite a while. Really their current hardware design is unparalleled and they even have some software lock-in with CUDA…
godelski 3y ago
> A similar threat is posed from ASIC chips that are dedicated to LLM tasks, but at least in that case Nvidia can still compete for a specific hardware design rather than overcoming an entire category of commodity hardware. But also consider the reason Nvidia is doing so well in the first place. The reason is because they were already developing hardware that specialized at performing tensor operations. Maybe not at the size or always the same types we see in ML, but not very far either. There's reasons why they aren't filling cards with tensor cores. But they are clearly developing specialized hardware for ML tasks. I'm not sure betting on LLMs sticking around is the best bet though. I'll be very impressed if scale is all you need.