5 ms·

I think the quote [1] is technically on point: > A GPU is really built for a very specific class of workloads. Generally those have been called thoughput workl

I think the quote [1] is technically on point: > A GPU is really built for a very specific class of workloads. Generally those have been called thoughput workloads. So it does lots of floating point processing and matrix operations. So it's very dedicated for things like graphics, matrix, and now has worked out to be uniquely good at things like AI. This is all very true. He does seem to underplay the usefulness of GPU in HPC style workloads, but that is inferable from the throughput and floating point and matrix. [1]: https://youtu.be/d07wy5AK72E?si=XfjMotKH5xf0ti9m&t=35 https://youtu.be/d07wy5AK72E?si=XfjMotKH5xf0ti9m&t=35

ArtTimeInvestor 2y ago
Parallelizable workloads are going to be the vast majority of computational workloads in the future. Therefore, calling them a "very specific class of workloads" misses the mark in my opinion.
- MobiusHorizons 2y ago
  Unless you are talking about ai workloads I would absolutely not agree. GPU style compute is throughput optimized to the point that makes it nearly useless for latency sensitive compute except in very narrow cases like graphics that can be made embarrassingly parallel. But you could never run an os scheduler or a file server or a database on a gpu with anything like competitive performance. The memory latency and branch penalty are just too high for anything where the sequence of operations isn’t highly predictable and repeated for a large number of items. Something as simple as searching for an item in a data structure is hard to do efficiently on a gpu. Basically anything with an if statement that depends on data is going to be incredibly slow compared to a cpu.
  - ArtTimeInvestor 2y ago
    Parallelizable workloads include AI workloads. And AI workloads alone will far exceed the amount of CPU workloads mankind uses. If the CEO of Intel failed to realize that, it is no wonder that the company ran into problems.
    - MobiusHorizons 2y ago
      > Parallelizable workloads include AI workloads Certainly! I don’t think anyone is arguing against this. Pat even specifically called it out. > And AI workloads alone will far exceed the amount of CPU workloads mankind uses. Maybe, maybe not. As I see it, AI is still not obviously useful enough that people are willing to pay the compute cost for what it produces. The industry is currently in a hype cycle which allows funding to be no object we try to find the best uses and drive down costs. But when that hype cycle ends and companies try to make everything profitable, a lot of use cases will no longer make economic sense. This always happens even with successful hype cycles. It is far from obvious what the size of the AI market will be at that time. And even if it is wildly successful and parallelizeable compute dominates other forms of compute, you will still need CPUs to run the servers and keep the accelerators (GPU or otherwise) busy. > If the CEO of Intel failed to realize that, it is no wonder that the company ran into problems. This has nothing to do with it. There is plenty of demand for x86 cpus. Intel’s problems have to do with not being able to respond to competition from TSMC and AMD or Apple. The biggest monetary threat is AMD winning in the server market, as this is both lucrative and moves slowly. But they also have big problems from both AMD and Apple on the consumer side. These problems are what Pat was brought on to solve, and Intel has made a lot of progress under his leadership. Intel is finally competitive on core counts and has made a lot of progress on efficiency. I think 18a could have really put them over the edge of competitiveness, but the board is not willing to wait for that investment to bear fruit, so here we are.
      - ArtTimeInvestor 2y ago
        AI is still not obviously useful enough that people are willing to pay the compute cost for what it produces. I witness the exact opposite. Example: Programmers with $100/hour salary who send a query for $0.01 to an LLM which saves them an hour of work to figure out how to use an API and integrate it into their application.
        MobiusHorizons 2y ago
        I’m not saying no one will be willing to pay for it, but rather that the number of people willing to pay for it will decrease when the prices start to reflect all the costs necessary to become profitable. A lot of investor money is currently keeping those costs artificially low for growth, and to find product market fit. In my opinion the number of use cases people (and especially businesses) will be willing to pay full price for are not yet obviously large enough to dominate global compute.

ArtTimeInvestor 2y ago

Parallelizable workloads are going to be the vast majority of computational workloads in the future. Therefore, calling them a "very specific class of workloads" misses the mark in my opinion.

MobiusHorizons 2y ago
Unless you are talking about ai workloads I would absolutely not agree. GPU style compute is throughput optimized to the point that makes it nearly useless for latency sensitive compute except in very narrow cases like graphics that can be made embarrassingly parallel. But you could never run an os scheduler or a file server or a database on a gpu with anything like competitive performance. The memory latency and branch penalty are just too high for anything where the sequence of operations isn’t highly predictable and repeated for a large number of items. Something as simple as searching for an item in a data structure is hard to do efficiently on a gpu. Basically anything with an if statement that depends on data is going to be incredibly slow compared to a cpu.
- ArtTimeInvestor 2y ago
  Parallelizable workloads include AI workloads. And AI workloads alone will far exceed the amount of CPU workloads mankind uses. If the CEO of Intel failed to realize that, it is no wonder that the company ran into problems.
  - MobiusHorizons 2y ago
    > Parallelizable workloads include AI workloads Certainly! I don’t think anyone is arguing against this. Pat even specifically called it out. > And AI workloads alone will far exceed the amount of CPU workloads mankind uses. Maybe, maybe not. As I see it, AI is still not obviously useful enough that people are willing to pay the compute cost for what it produces. The industry is currently in a hype cycle which allows funding to be no object we try to find the best uses and drive down costs. But when that hype cycle ends and companies try to make everything profitable, a lot of use cases will no longer make economic sense. This always happens even with successful hype cycles. It is far from obvious what the size of the AI market will be at that time. And even if it is wildly successful and parallelizeable compute dominates other forms of compute, you will still need CPUs to run the servers and keep the accelerators (GPU or otherwise) busy. > If the CEO of Intel failed to realize that, it is no wonder that the company ran into problems. This has nothing to do with it. There is plenty of demand for x86 cpus. Intel’s problems have to do with not being able to respond to competition from TSMC and AMD or Apple. The biggest monetary threat is AMD winning in the server market, as this is both lucrative and moves slowly. But they also have big problems from both AMD and Apple on the consumer side. These problems are what Pat was brought on to solve, and Intel has made a lot of progress under his leadership. Intel is finally competitive on core counts and has made a lot of progress on efficiency. I think 18a could have really put them over the edge of competitiveness, but the board is not willing to wait for that investment to bear fruit, so here we are.
    - ArtTimeInvestor 2y ago
      AI is still not obviously useful enough that people are willing to pay the compute cost for what it produces. I witness the exact opposite. Example: Programmers with $100/hour salary who send a query for $0.01 to an LLM which saves them an hour of work to figure out how to use an API and integrate it into their application.
      - MobiusHorizons 2y ago
        I’m not saying no one will be willing to pay for it, but rather that the number of people willing to pay for it will decrease when the prices start to reflect all the costs necessary to become profitable. A lot of investor money is currently keeping those costs artificially low for growth, and to find product market fit. In my opinion the number of use cases people (and especially businesses) will be willing to pay full price for are not yet obviously large enough to dominate global compute.