*this* is why there is no high end RDNA 4. and for a very good reason. i also find it amusing that even in the hpc/ai market AMD is giving more vram than Nvidia... and at this point one has to wonder why (especially at these margins at these elevated prices). all in all extremely competitive and as has been proven, more effective in supercomputing. as long as AMD is producing this level of product i can deal with skipping a halo gpu every now and then.

This looks good and appears to be a competitive alternative to Nvidia. I suspect the benchmark is written using OpenCL since it's one of the few multi-platform APIs for this kind of stuff. The big question for those who use these kind of devices is the API - if they use CUDA then AMD's chip is not an alternative for them. If there was a viable way to use CUDA with AMD devices that would really, really help them. They would likely come in at a lower cost than Nvidia so they could potentially take a fair amount of business away from Nvidia if that were to happen.

AMD has usually offered pretty good server and workstation hardware for reasonable prices. The problem always comes down to drivers.

schmidtbag:

AMD has usually offered pretty good server and workstation hardware for reasonable prices. The problem always comes down to drivers.

more right than wrong for sure. not really on point here though because MI300 (both versions) really is cutting edge hpc that has extremely high relevance for ai. as for the hpc applications they're all bespoke and lean into known AMD strengths. this is pretty much the reverse of the H100 that leans into the cuda library for ai, but is relevant for hpc.

Wonder why AMD used H100 for comparison instead of H200 in their official marketing? Looking at indepedent ML Perf reporting those results you can see the comparisons against H100, H200 and B200. In this case the extra memory advantage against H100 did not result in a large difference in performance.

I feel vibes of "AMD marketing" again to make their competitors look worse by putting them in unfavorable scenarios: 1. Why not testing against B200? - maybe because of the power draw of B200? - it could be the only valid reason. 2. Why using a different CPU for Nvidia? No CUDA support - then who cares? Almost all AI works today revolves around CUDA because of AMD fault - they did not foreseen this AI fast development and they did not cared enough to have a competitive comparable solution. Now they are too little too late, an year late to precise since H100 was launched in march last year.

pharma:

Wonder why AMD used H100 for comparison instead of H200 in their official marketing? Looking at indepedent ML Perf reporting those results you can see the comparisons against H100, H200 and B200. In this case the extra memory advantage against H100 did not result in a large difference in performance.

Shaddy marketing, as usual. Every company loves to do this... But it doesn`t matter, this GPU is going to be a sales success, mainly because Nvidia can`t serve everyone who wants/needs their GPUs...

Maybe becouse B200 price is 30,000$-40,000$ and MI300x is 15,000$.

pharma:

Wonder why AMD used H100 for comparison instead of H200 in their official marketing? Looking at indepedent ML Perf reporting those results you can see the comparisons against H100, H200 and B200. In this case the extra memory advantage against H100 did not result in a large difference in performance.

until you build bleeding edge supercompute where the MI300's (both versions) are the best for scaling. there's a reason the DoE chose AMD for El Capitan for "run of the mill" AI use you're absolutely right. but a point also can be made that that the supply issue that Nvidia is facing with the crap Blackwell yield (postponed orders, long time frame for delivery) that if one is kitting out "skynet" and has a firm timetable one can get a MI300 order faster than an H100/200 (both are selling as fast as they're made) depending on the number of units needed. but Nvidia has already told its best clients B100 isn't shipping in volume any time soon.

tunejunky:

until you build bleeding edge supercompute where the MI300's (both versions) are the best for scaling. there's a reason the DoE chose AMD for El Capitan for "run of the mill" AI use you're absolutely right. but a point also can be made that that the supply issue that Nvidia is facing with the crap Blackwell yield (postponed orders, long time frame for delivery) that if one is kitting out "skynet" and has a firm timetable one can get a MI300 order faster than an H100/200 (both are selling as fast as they're made) depending on the number of units needed. but Nvidia has already told its best clients B100 isn't shipping in volume any time soon.

I really don't know where you get your info but select clients already have B200 and will start ramping up in Qtr 4 this year. Volume production will be during Qtr 1 2025. You might want to read the linked below for more info on AMD's AI offering. AMD's AI Plan: The Nvidia Killer or a Wasted Effort? (hpcwire.com)

pharma:

I really don't know where you get your info but select clients already have B200 and will start ramping up in Qtr 4 this year. Volume production will be during Qtr 1 2025. You might want to read the linked below for more info on AMD's AI offering. AMD's AI Plan: The Nvidia Killer or a Wasted Effort? (hpcwire.com)

They seem to have relatively solid plan. And prolly will gain some marketshare

Fun fact: AMD/ Radeon give 0 shits about the plebeian "gamer" GPU market. All their focus is on feeding the AI boomers these overpriced MI300X chips, who think they're cashing in, they only want to keep the Radeon brand functional enough to not lose the console contracts and to have a presence in the home market. Not that nGreedia aren't the same shit.

Neo Cyrus:

Fun fact: AMD/ Radeon give 0 shits about the plebeian "gamer" GPU market. All their focus is on feeding the AI boomers these overpriced MI300X chips, who think they're cashing in, they only want to keep the Radeon brand functional enough to not lose the console contracts and to have a presence in the home market. Not that nGreedia aren't the same crap.

AI accelerators and CPUs comes first for AMD. They plan to introduce a new GPU's every year while their gaming market is shrinking. Who gives a crap about those, right? Nvidia is doing the same but their brand recognition and market share is alot higher.

Neo Cyrus:

Fun fact: AMD/ Radeon give 0 shits about the plebeian "gamer" GPU market. All their focus is on feeding the AI boomers these overpriced MI300X chips, who think they're cashing in, they only want to keep the Radeon brand functional enough to not lose the console contracts and to have a presence in the home market. Not that nGreedia aren't the same crap.

MI300X is actually very competitively priced. If you had a company and could sell products to individuals for upto $1000 for ONE unit, or you could sell many multiple units to corporations for $15k-$20k per unit? Its a no brainer for ANY business, it allows for increased profits and therefore increased RnD for future products and that server tech always ends up inside our desktop chips later down the line. Its how its worked for decades. Only now they have a seriously lucrative reason to push everything they can into the server market due to the AI boom, hence no high end RDNA4 where all those chips will go into server units. Why bother with gamers when 80%-/+ just follows suite and buys Nvidia anyway (me included), why have big chips sitting on store shelves when they can charge 10-20X and have a gaurenteed sale. This is another reason why Zen 5's biggest increases is inside server workloads, AVX512, branch predictions, etc. Some benchmarks show upwards of 50% performance increase over Zen 4, once again they chose this route because its where the real money is. While we get 5-10% if we are lucky.

CPC_RedDawn:

MI300X is actually very competitively priced.

That's another way of saying "as expensive as they could get away with (while within their optimum formula)". I don't need an explanation of their reasoning. I just think they, & especially nGreedia, could have made their giant fat profits without having to shit & laugh in the faces of the customers who got them to where they are. As for Zen 5... the Zen 5% gaming memes need to stop... because it's more like 0.5% if you get rid of extreme outliers no one plays like Asseto Corsa. Even including those, HUB got 3% & TPU got 2%. Zen 5% makes it sound like it was an actual step forward for gaming, as you said it was made entirely with workstations in mind. Their only gaming goal was not to regress in too many games/engines vs Zen 4. "Who gives a shit?" - Engineers as AMD

Neo Cyrus:

That's another way of saying "as expensive as they could get away with (while within their optimum formula)". I don't need an explanation of their reasoning. I just think they, & especially nGreedia, could have made their giant fat profits without having to crap & laugh in the faces of the customers who got them to where they are. As for Zen 5... the Zen 5% gaming memes need to stop... because it's more like 0.5% if you get rid of extreme outliers no one plays like Asseto Corsa. Even including those, HUB got 3% & TPU got 2%. Zen 5% makes it sound like it was an actual step forward for gaming, as you said it was made entirely with workstations in mind. Their only gaming goal was not to regress in too many games/engines vs Zen 4. "Who gives a crap?" - Engineers as AMD

the mi300x costs litterally less than half of what nvidia charges for the h100. and its a pretty big chip. they are not selling at what its worth, they are trying to gain market share from nvidia, much like what they did with epyc, also amd's HPC accelerators aren't even derived from the recent consumer gpus, they are still gfx9(vega), all of the money amd is making off of them come from corporate clients, its been 3 generations of hpc accelerators with no consumer variant, last one being the mi50, (radeon VII), if anything its the HPC stuff that funds radeon. amd's consumer gpus dont bring in the same kind of cash. radeon flounders on its own unfortunately. zen is much the same way, corporate funds the development, not consumer, thats why the consumer stuff is "second rate", its litterally reject epyc dies. It is the double edged sword of the chiplet designs, you get cpus that aren't solely designed for consumer products and as such aren't as good as they could be for the task, but you are also directly getting the benefit of the corporate r&d, this juggling act is actually rather impressive, because amd managed to cut alot of corners, and still put out silicon that directly competes with intel monolithic consumer chips and work to extend their lead in the server space.

LOL i still remember all that hype about MI250x beating Nvidia H100, but now the MI300x that came out a year later just barely tie with H100 in MLperf benchmark? Nvidia H200 is around 40% faster than H100 in MLperf AMD hype train sure get derailed every time

@user1 The key phrase in what I said was "while within their optimum formula".

Neo Cyrus:

@user1 The key phrase in what I said was "while within their optimum formula".

you said in the context of the mi300x being " competitively priced"

I just think they, & especially nGreedia, could have made their giant fat profits without having to crap & laugh in the faces of the customers who got them to where they are.

how is a company crapping on the faces of their customers if they are building the products that their customers(corporate) want and doing it at a lower price than their competition ? the fact is that abysmal state consumer gpu market is in, and the lack of consumer focused performance orientated cpus is because people don't actually care about it much at all, they just want to play their games or run their renders and do their spreadsheets, the introduction of " filthy casuals" is what has made the market what it is today, its why the rtx 4060 is the best selling gpu, despite being such a poor value in relation to its predecessor. the market was much healthier when people actually cared about what they are buying.

Undying:

AI accelerators and CPUs comes first for AMD. They plan to introduce a new GPU's every year while their gaming market is shrinking. Who gives a crap about those, right? Nvidia is doing the same but their brand recognition and market share is alot higher.

In the AI market the situation today is more than "brand recognition" and market share. Nvidia invested a lot in software support also, AMD did not. Simple as that since hardware without software is nothing. We use a lot of GPU accelerated simulations and the question than always comes when we need to invest more is who has the best software support for our simulations and AI pattern recognition and not "let's go with Nvidia because this is all that we know" - we talk here about a lot of money invested and I don't think that brand recognition or market share plays too big of a role here. CUDA is easy to use, even fun to work with and AMD has nothing against it.