Re: Comment on NVIDIA GPU Versus AMD GPU
Here are two responses to grxbstrd‘s comment From the “AMD fans” and grxbstrd’s rejoinder:
Joe NYC says:
I agree with the analysis to a large extent, that AMD targeted scientific areas of FP64, where AMD has the largest lead (> 4x), which NVidia de-prioritized. AI was a secondary priority (but I bet there may be areas of AI where higher precision compute may be useful). Because if you want to do all at once, you end up with Ponte Vecchio, years delayed, and still questionable if it is even manufacturable...
But where I differ is when he knocks Mi200 as 2 chips competing against one. It's an MCM, just like Zen 1. But AMD already has a roadmap, paved by Zen.
Zen 2 went from just MCM to chiplets (8 of them)
Zen 3 is stacking on top of chiplets (1 x IO, 8x logic, 8x SRAM dies).
Which pretty much overwhelmed Intel. If NVidia does not do the something drastic with Hopper, AMD will overwhelm NVidia in raw performance the same way it overwhelmed Intel in data center. The performance may be untamed, hard to manage, but it does not have to do everything at once, it does not have to have all the answers to AI at launch. It could just be doing 1 thing for 1 customer, and doing it 2x better than NVidia.
That's how Epyc started 4-5 years ago, and now it has 10 out of top 10 customers
So I am betting that AMD will surpass NVidia in hardware, while still lagging in software, and from there, it will start gaining share, in a growing market.
CDNA 3 is likely going to have all of the hardware technologies (chiplets, stacking), and additionally, it will (based on some rumors) will be an APU. Meaning it will be (can be) a computer, not just an accelerator. I have no idea how this is to be implemented, if there will be, say Zen 4 CCD together with all of the graphical chiplets, but it is a possibility.
Additionally, it seems that AMD may have some special sauce at TSMC, some optimized N5 process, that may not be available to all (some people speculate). AMD is promising 2x efficiency vs. current 7nm, roughly translating to 2x performance from the same power envelope.
The conclusion of this, IMO, is that AMD sets goals for itself, and then hits the ball out of the park on those goals. And, you can bet that the goal for CDNA3 and RDNA3 is to beat Nvidia. But unlike CDNA2 and RDNA2 where AMD just maintains token presence, I think AMD will really push into these markets with volume. With RDNA3, into a bigger base, with CDNA3 starting from a small base.
Neolib says:
Well, I and others pointed out the obvious too at the time: AMD targeted FP64 while Nvidia is chasing AI/ML math, so yes they are two different markets, and my comment was that AI/ML is the hot sector in datacenter, not so much HPC, although that too has some traction. Note that the big players are in fact trying to entice supercomputer users to instead turn to the Cloud, rather than their own iron, so its not correct to claim AMD won't find a market outside DOE/DOD. There is a market outside of that, and to the degree the Cloud providers succeed with their dreams, they will greatly expand that market because even small biz that need some HPC simulation can afford it when they only pay for a few days of compute time now and again. But there will be far more demand IMHO for AI/ML from those same Cloud providers.
WRT to Intel, its far from clear how well they get their act back together, because I seriously doubt they will get process tech leadership back from TSMC, so at best, if they turn mostly to outside fabbing at TSMC, they can match AMD, but they are unlikely to ever have a process lead again, something that was very useful to them in the past. Sapphire Rapids is a horrible design compared to Milan. If this guy was complaining about MI200 being two large chips, what about SR being four large chips?
But also, the A100 is 54B transistors in 7nm as a single die, while the MI200 is 58B in two die on 6nm. Which one do you think yields better and which one is actually less silicon? So his chatter about it being two large die against one isn't in fact quite straight. They are about comparable for area, and the AMD design being two half-sized die will definitely yield better. In that sense AMD is in a nice position to the run the tables on Nvidia just like they are doing unto Intel, until both the latter figure out chiplets and get a competitive chiplet architecture. That is the one big thing that AMD has done an outstanding job on, while Nvidia and Intel, both with very deep pockets, screwed the pooch instead.
I do agree with him however on the software side and creating new markets. Nvidia has the goods there, but IMHO, the real problem there is Nvidia likely needs to somehow start owning the end users, and thus they must start competing with their current customers. I think the trend of the big players developing the full stack of hardware and software is well in motion, and that is bad news for Nvidia, AMD, and Intel, unless they can somehow compete with those (former??) customers. Nvidia might be positioned to do that with Omniverse, but if they think they can, they better make it "free" and very user friendly so that they do succeed in getting millions of users. Otherwise it will be MS, Meta, Tencent, etc who develop it and own the end users, and although it might start out using lots of Nvidia (and AMD and Intel) hardware, in the end it won't.
grxbstrd Responds:
“AMD replies"
appreciate your postings, I'm not going to debate them here with you as the intermediary. I'd be happy to debate them in public here on SA though.
Nearly every one of these comments have already been debated on SA both on NVDA articles and on AMD articles. These are the same tired AMD talking points that basically boil down to, just wait until next time. Anyone can look at my history to review the details if they wanted to. |
