AlphaZero

[CheckersGuy]

There would't be much marker for AlphaZero, as you need a TPU to usefully run it.

However, now that NN are proving their worth in various applications, it becomes a good question whether in the future we will see more NNs in common applications, and TPU technology seeping into our PCs and phones to efficiently run those. There has been a time where floatig-poit arithmetic was considered so difficult and outladish that standard CPUs did not support them, and you needed a separate FPU chip (the 8087) to equip your PC with that capability. Now we have them as a standard feature in every CPU core. It is quite possible that TPUs will go the same way.

Note that the Google TPUs are not really much more complex in terms of number of transistors or power consumption than state-of-the-art CPU chips with many x86 cores. (And they are pretty big, as TPUs go.) The benefit of going from 4 cores to 8 cores might be very small for the average user, while having 4 CPU cores plus a TPU on that chip could open a plethora of new applications.

That is all true, but Google doesn't have much saying into this. It is all on NVIDIA. In terms of chip design TPU is totally ordinary thing, 22nm process is already 2-3 years behind cutting-edge, they most probably used old TSMC process.
Once NVIDIA actually concludes that dedicated ML chips are good for business we'll have ML chip of 10x TPU performance in no time.
However, atm I'm pretty sceptical that even 15% of for GV100 would be used for ML. It would still be 60% gaming and 25% mining. So no market yet for dedicated chip.

There a lot more companies than nvidia working on Ai chips that will end up in the market. Most notable Graphcore and Wave which now have the money to go ahead and build their chips

[jhellis3]

It would still be 60% gaming and 25% mining.

Lol, the designed use case for GV100 is ~0% gaming. It is over 800mm^2. Its designed purpose is HPC, ML, and professional graphics, roughly in that order.

The number 1 customer for GV100: https://www.olcf.ornl.gov/summit/

[Milos]

It would still be 60% gaming and 25% mining.

Lol, the designed use case for GV100 is ~0% gaming. It is over 800mm^2. Its designed purpose is HPC, ML, and professional graphics, roughly in that order.

The number 1 customer for GV100: https://www.olcf.ornl.gov/summit/

Lol, you really believe that?
4600x6 = 27600. So you believe 28000 GV100 is majority of GV100 sold???
One thing is what NVIDIA publishes in their press release wanting to popularise it in ML market, and completely other things is where they are actually going to sell the card.
GV100 will be sold in hundreds of thousands unless NVIDIA pushes Series 20 gaming cards in near future (for which there are no indications). The catch is that it is much more powerful than anything existent for the games and with price of 3000$ is more than affordable for enthusiasts. You really don't know much about the world of professional gaming do you?
And I don't even touch the mining market.
Here is an overview of the gaming and mining performance:
https://hothardware.com/reviews/nvidia- ... gpu-review

[CheckersGuy]

Still, the titan V might be pretty good for deepLearning/machine-learning but having something that' completly dedicated to ML instead of a mixture would be much better

[jhellis3]

So you believe 28000 GV100 is majority of GV100 sold?

Reading skills can be helpful. I can understand if English is not your first language, but the phrase number 1 customer only implies nobody else is buying more. And those are full spec Teslas, not the salvage bin Titans. If your misunderstanding was intentional, well you can create strawman fallacies all you like, but imitating ARB level delusions is probably a poor way to live one's life. Anyway, your ignorance is your own. It is certainly not my problem.

You really don't know much about the world of professional gaming do you?

Other than actually playing at a professional level in an FPS?

[Milos]

So you believe 28000 GV100 is majority of GV100 sold?

Reading skills can be helpful. I can understand if English is not your first language, but the phrase number 1 customer only implies nobody else is buying more. And those are full spec Teslas, not the salvage bin Titans. If your misunderstanding was intentional, well you can create strawman fallacies all you like, but imitating ARB level delusions is probably a poor way to live one's life. Anyway, your ignorance is your own. It is certainly not my problem.

Your argument is a straw-man. You shown 28000 GV100 as a main argument that most GV100 would go into HPC. Why else would you mention it?
Largest single customer of 28000 doesn't tell absolutely anything regarding proportions of global sales of the product that will be sold in hundreds of thousands of units.

You really don't know much about the world of professional gaming do you?

Other than actually playing at a professional level in an FPS?

Yea sure, and I'm actually NVIDIA sales VP.

[jhellis3]

You seem to be confused. Like I said, your ignorance is not my problem.

[Milos]

You seem to be confused. Like I said, your ignorance is not my problem.

Sure, while you demonstrated so much knowledge. Your posts are so educative, I learned so much from them. Thanks man, you are a hero!

[jhellis3]

I do what I can. I see incorrect information, I correct it.

But I am sorry if I led you to believe that I care what you personally think, feel, or do; that was never my intention.

[Milos]

I do what I can. I see incorrect information, I correct it.

I don't think you are actually capable of correcting anything. Maybe in your own version of reality. And that's kind of ASD territory.