AMD Instinct MI300X Could Become Company's Fastest Product to Rake $1 Billion in Sales

[Avro Arrow]

With Radeon Instinct already powering two of the fastest three supercomputers in the world (and soon to be three of the top-four), it's quite clear that CUDA is more or less inconsequential at the supercomputer-level.

Those are bold claims. Can't say with much confidence that's from a diverse set of clients. More likely down to a few clients of likes of National labs ordering by contract for a SuperComputer.

Sales are sales.

CUDA is really entrenched in Universities especially in the US, it is even part of the curriculum in some Unis for their undergrads/grads. These future professionals once graduated they take that CUDA mindset to their workplaces. If at all they get into steering roles they will push Nvidia ecosystem. Then when those people hire the next group of engineers, these seniors will be looking for CUDA experience. That's not brand loyality its just prefefnece for familarity.

These supercomputers calculate extremely complex atomic and molecular interactions like protein folding and nuclear chain-reactions. Two of the top-three fastest supercomputers in the world are powered by EPYC and Radeon Instinct, those being Frontier (currently the fastest on Earth) and LUMI (3rd-place). When El Capitan comes online, the two fastest supercomputers in the world will be powered by EPYC and Radeon Instinct. CUDA is of no consequence at this level and it shows by the very presence of Frontier and LUMI in the #1 and #3 spots, respectively. The fastest supercomputer with nVidia hardware is Leonardo and it's in 4th-place. These supercomputers aren't used for running blender and the components are chosen by some of the top computer engineers in the world. They don't care what colour the box is when the parts arrive so nVidia doesn't have its "clueless noob" advantage like it does in the gaming spaces. These computers are used by engineers at a level that is far above what you would find in a university. These computers are used by engineers from agencies like NASA. Do you think that NASA engineers give a damn about a bunch of university professors using CUDA in their CGI courses?

Here's a hint...they don't. They want what does the job that they need done and CUDA clearly has nothing to do with these purposes, as we've seen by the part selection made over the past few years. If EPYC and Instinct do the job they need done better than the offerings from Intel and/or nVidia, they're going to use them. This is the US government we're talking about here, you know, the one with the unlimited budget (unless it comes to giving their citizens universal healthcare). If nVidia hardware was better (as you seem to be trying so hard to claim), then that's what they'd be using. Oh, except that they're not...

AMD wont make much head way even with relatively cheaper hardware when their University drives are next to nothing. Professors who asked their postdocs to go with AMD did so only becuase the grants were limited for their work.

I don't know what to say to this because (other than what universities use being completely irrelevant) the very existence of the top-3 supercomputers, Frontier, Fugaku and LUMI proves your statement to be 100% false. None of these supercomputers use nVidia hardware for anything. The Kubaku is the odd one out because it uses Fujitsu-designed ARM8.2 cores and doesn't require separate GPU cores. The other two have been extremely successful in their roles with EPYC and Instinct for the USA and Finland. After their experience with Frontier, if there were any drawbacks to using Radeon Instinct GPUs, there's not a chance that the US government would be investing even more money into an even more expensive and powerful Cray like El Capitan. They can spend limitless amounts of money on whatever they want which means that if they chose Radeon Instinct, then it is either the best for the task or at the very least, nVidia would be no better.

Then there is an imminent reality this AI bubble will be burst by startups and other companies putting out specialized ASICs that handle training and inferencing much quicker and far more efficientlty. Inferenecing on analog designs is already a thing and vastly more efficient.

What does that have to do with the M1200X being the fastest AMD product to reach $1,000,000,000? You must be a real fanboy to bring up things that have nothing to do with the article in your attempt to say "But, but, but, CUDAAAA!!!".

 

[samum]

After their experience with Frontier, if there were any drawbacks to using Radeon Instinct GPUs, there's not a chance that the US government would be investing even more money into an even more expensive and powerful Cray like El Capitan. They can spend limitless amounts of money on whatever they want which means that if they chose Radeon Instinct, then it is either the best for the task or at the very least, nVidia would be no better.

While I think your post is generally right, never underestimate the capacity of the U.S. government to waste money.

 

[Kapone33]

It's all about time frame

 

[HisDivineOrder]

One of these days, the crazy music will finally end and these GPU companies will stop dancing to their daydream of being in another industry. Imagine if they just tried to be the best gaming GPU company they could be.

They chase these fads, swell their stock prices, and then look stupidly swollen when the fad inevitably bursts. It'll be no different this time. You'd think the investors would see through the fads by now. They just got done with another one a year or so ago.

 

[mahirzukic2]

One of these days, the crazy music will finally end and these GPU companies will stop dancing to their daydream of being in another industry. Imagine if they just tried to be the best gaming GPU company they could be.

They chase these fads, swell their stock prices, and then look stupidly swollen when the fad inevitably bursts. It'll be no different this time. You'd think the investors would see through the fads by now. They just got done with another one a year or so ago.

But that is just par for the course in the capitalism.

 

[Avro Arrow]

While I think your post is generally right, never underestimate the capacity of the U.S. government to waste money.

Fair point. I'll give you that.

 

[kingo]

H100 is specialized for AI performance, so it is slower and less power efficient than MI250X for supercomputers.