News Jensen Huang says even free AI chips from his competitors can't beat Nvidia's GPUs

[bit_user]

AI is too big for any company to maintain an unassailable lead. There’s too many startups willing to take technological risks for big gains.

I was thinking this, maybe about 5 years ago. Now, I'm not so sure. Nvidia has more resources to invest into R&D of next gen chips than many startups, put together. Not to mention expertise. There's always the possibility they might make some missteps, but I see no reason why the next revolutionary AI chip can't come out of Nvidia, itself.

Nvidia has split off its AI training + HPC architecture from its client graphics architecture. IMO, they just need to take the next step and separate AI from HPC.

 

[digitalgriffin]

I was thinking this, maybe about 5 years ago. Now, I'm not so sure. Nvidia has more resources to invest into R&D of next gen chips than many startups, put together. Not to mention expertise. There's always the possibility they might make some missteps, but I see no reason why the next revolutionary AI chip can't come out of Nvidia, itself.

Nvidia has split off its AI training + HPC architecture from its client graphics architecture. IMO, they just need to take the next step and separate AI from HPC.

AMD has steadily eroded Intels dominant market share in servers and DIY Home builds. All it takes is a few winning consecutive years. Right now the MI300 is a huge winner. And the ROCm toolkit is a wedge to break CUDA monopoly. Give it a few years. NVIDIA, AMD, and Intel are limited to what techonology is availble for fabbing chips. So one can't do X Y Z that's more than 2 generations ahead for a similar chip size. And the advantage of starting late is you have a clean sheet where you don't have legacy support which eat up resources. Key examples: ARM, and Tesla.

 

[dosmastrify]

Jensen Huang has taken to the stage to claim that Nvidia's GPUs are "so good that even when the competitor's chips are free it's not cheap enough." Huang explains that Nvidia GPU pricing isn't really significant in terms of the all-important data center TCO.

Jensen Huang says even free AI chips from his competitors can't beat Nvidia's GPUs : Read more

If the free chips are more power hungry he may well be entirely correct

 

[dosmastrify]

Anything that rises too fast, falls much faster.

Lucid and weed stocks have entered the chat

 

[bit_user]

AMD has steadily eroded Intels dominant market share in servers and DIY Home builds.

But they got a massive aid by Intel's 10 nm process node getting stuck. So, when you combined some good engineering and outside-the-box thinking with AMD's good fortune of TSMC firing on all cylinders and Intel's 10 nm node stalling out, AMD was able to climb out of its hole and actually mount real competition for Intel.

Nvidia, on the contrary, only had some slight disadvantages from using a 12 nm node for Turing and Samsung 8 nm for Ampere (if we're talking about client GPUs). On the design front, they were slightly disadvantaged by the amount of silicon they spent on Tensor cores and RT cores. Other than that, Nvidia was almost as competitive as ever. Now that they're on effectively the same node as AMD and wised up to the amount of cache they needed, they've effectively eliminated any self-imposed disadvantages.

AMD made smart moves on reducing their wavefront size, in RDNA, and with RDNA2's Infinity Cache. However, instead of doubling down on Infinity Cache, they tripped over their own feet with the chiplet thing. If it would've let them increase the amount of cache, then heck yeah! But, it instead resulted in less cache, possibly at higher-latency, and a more expensive assembly process that probably offset the silicon savings from using N6 chiplets for the MCDs. So, right now, it's really AMD who's on the back foot.

All it takes is a few winning consecutive years.

I don't expect AMD to get a more favorable opportunity than they had with Turing and Ampere. Nvidia probably isn't going to take the risk of using a cheaper node, after seeing how it briefly cost them the uncontested crown to the RX 6950XT.

Right now the MI300 is a huge winner.

Yeah, but it's only going to have about 6 months of shipping in volume, before Blackwell ramps, if that? That's better than nothing, but it's not enough to turn around AMD's fortunes.

And the ROCm toolkit is a wedge to break CUDA monopoly.

Hardly. The best ROCm can hope for is matching Intel's driver & userspace stack.

Give it a few years. NVIDIA, AMD, and Intel are limited to what techonology is availble for fabbing chips.

Back in 2019(?), Nvidia bought Mellanox, which has been crucial to enabling them to scale their training solutions to rack-scale and beyond.

So one can't do X Y Z that's more than 2 generations ahead for a similar chip size.

What Nvidia has is much more than just chips. They have the entire stack, and it's been tuned and optimized over an entire decade!

What Intel saw in Habana Labs was not just the silicon and software, but also a scaling solution. If I were Nvidia, I'd be more worried about Intel than AMD.

 

[shalako33]

Didn't they recently put a stop to others being able to use their own hardware to run Cuda not long after AMD released software that can run Cuda on AMD hardware ?
If their hardware's so good why stop that ?
I thought there was some kind of law against doing this.