Submitted by I had a chance to speak to Jack Huynh, AMD’s senior vice president and general manager of the Computing and Graphics Business Group, during IFA 2024 in a question and answer session. Due to speculation that AMD won’t launch flagship GPUs for its next-gen lineup, I pressed Huynh for information regarding the company’s plans for the high-end GPU market with the RDNA 4-powered Radeon RX 8000-series. His comments sketch out a plan focused specifically on gaining market share in the GPU market above all else, and this strategy deprioritizes chasing Nvidia’s highest-end gaming cards — at least for now.
↫ Paul Alcorn at Tom’s Hardware
Reading through the actual comments, it seems that AMD is not going to chase the very, extreme high-end that NVIDIA serves, like the 4090 level of GPUs. Honestly, I’m completely okay with that – those high-end GPUs are insanely expensive, and unlike what YouTube and tech websites might suggest, nobody buys these GPUs. Consistently, for more than a decade now, it’s the xx60-xx70 levels of cards that dominate the market, and it’s smart of AMD (and Intel) to focus on that segment if you want to sell as many GPUs as possible.
The very top of the GPU market just doesn’t make a lot price/performance sense. You pay considerably more for a 4090 compared to a 4080, but the price increase does not correspond to a similar increase in performance. It simply makes a lot more sense to save that money and spend it elsewhere, such as on a better CPU, more RAM, more storage, or a new display. I’d rather AMD not waste time and energy on making these high-end GPUs nobody buys, and instead focus on improving the GPUs people actually buy.
And of course, AMD just hasn’t been able to match NVIDIA at the top end, and that’s probably not going to change any time soon. Releasing a high-end, expensive GPU, only to be trounced by your one competitor every single time is not a good look, so why even try?
Aiming for the high end is not only about selling quantity, it’s about R+D and tech. If you can build a top GPU, you can scale it down and cost save. What AMD are saying (like intel) is “the best we can make is midrange”. Which, yes, is where the market is, But not where it will be. Effectively they have decided to always operate a generation behind Nvidia in the GPU market. How many of us seriously look at Intel for our GPU needs? AMD will soon be in the same bracket.
This is very similar to what VIA did in the x86 CPU market back in the 2000s. In the end they simply fell so far behind they ceased to be competitive. AMD chose to go toe to toe with Intel and at least kept up and at various times pulled ahead.
Adurbe,
Theoretically AMD could make compelling high end hardware, they aren’t the fab, but IMHO software/cuda are the main barrier to entry at the high end. AMD struggles to displace it.
But he has a point. SiS Xabre could not defeat nvidia even though the Xabre is much stronger than the geforce 3 and 4, bu then everyone was buying a 9700 ATi card. The Xabre 800 was both faster and stronger than the nvidia and ATI card by a good margin, hell even the SiS Xabre 600 was on par with the geforce 3.
What brought down the Xabre cards was the drivers. Some games tested by SiS labs was running amazingly well, and the 2d chip was way beyond what ATI and nvidia had at the time. So it was relegated to “medical equipment” instead of performance market, even though it could have easily competed (and the 600 sometimes outperformed all others at 1280×1224). The matrox parhelia also was a competitor that out performed the GT4600/4800 but it was slower than the 9700pro at a much higher price. In supported games, the Xabre 400 beats the snot out of the parhelia and even the 1999 Voodoo5 5500 PCI can beat the score of a parhelia if you turn on fsaa.
But then again, if not all software works. It is a no go.
Best example of Xabre failiures is despite working perfectly with other UT tiltles: Rune. It was insanely popular in some markets. But it would not run on a Xabre at that time. (it does now, and frankly quite well), another one is Shogo MAD, the fog is gone. Nvidia did shitty fog, but at least it was there. Using a V5 5500 it looked like magic never seen before. nglide sgets it pretty close if you want to try it on a modern machine, but it is not the same… good but not the same. ultim@te race pro is another title that nvidia and ati sucked at that looked tremendous on the Xabre and the Voodoo, but voodoo was unbeatable in terms of fog.
It’s also the strategy AMD used for a long time with the K6. They were never faster than Intel – they were fast enough, and much cheaper. It’s a solid strategy. If AMD can make the case for better performance per dollar, why would anyone buy a more expensive, less capable nVidia GPU, to be laughed at by their friends?
It’s also why bulldozer arch sucked…
I don’t agree with that at all. They were trying something different, and it just didn’t pan out. There were interesting things about Bulldozer – it shared a lot of cache, between cores. If programs were written differently, or if compilers were optimized differently (software side – AMD sucks at that), they could have performed quite amazingly. Some folks did some demos to prove that. It just wasn’t the right architecture for the way folks were using CPUs. They took a risk on that one, and it didn’t pay off – the problem is, risks in CPU land take 10 years to correct from.
It absolutely is the reason… they were not “trying something different” They literally gave up on high perf CPUs and tried to make high throughput server CPU… very similar to the already then failing Sparc which had taken the exact same path of failure.
You can’t live in a dream land where programs are not written the way they are and expect to win…. you will fail EVERY TIME.
Both AMD and Sparc CPUS developed with this ideology sucked…
No it sucked. They increased the pipeline. This instantly decreases the pefromance unless you can back it up with massive mhz increases. The northwood was better than Prescot (24 vs 31 pipeline stages) at almost any task. Yeah the prescot core ruled when using winrar. The lowly 2500+ was better at everything than the 3.067 prescot.
And perhaps you are remembering it wrong, or you are ver localized. In sweden all gamers bought the athlon and the athlon XP after they moved to the socket model instead of SECC.
Gamers also gravitated towards ATi in sweden, forcing the warcraft 3 patch after lots of refunds and complaints. Nvidia had their first battle with a EU country back then (and they were a minor player at the time, ATi and 3dfx was the main players in sweden at the time, 3dfx even had a licensee make cards in sweden for the european market)
Nothing either of you said negates anything I said. Yes, AMD did give up on single core high end performance after they realized they couldn’t solve the software problem – only after. They then immediately pivoted – which as I said, takes 10 years to realize. Pivots in CPU designs are not fast.
There are probably still some very useful deep-dive articles up on anandtech if you care to find them.
As far as sales for gamers – it was a no brainer. Nothing at the time was more constrained by single thread performance than games. AMD had to sell their CPUs VERY cheap in order to move product. But I maintain, it was very possible to get high end performance out of those – it’s just that no developers really put in the effort (for good reason).
3dfx? They stopped mattering in the late 90s. Bulldozer didn’t arrive until 2011. I’m not sure how that’s relevant here.
… but my Piledriver disagrees :^D
CaptainN–,
I wouldn’t say “never”.
A few years AMD were beating intel at both ST and MT loads. Of course, this was because intel got stuck on ~14nm for what seemed like forever. It was a case of TSMC fabs beating intel and benefiting from intel’s missteps.
Hell the duron spitfire is the highest overclocking x86_32 cpu ever made. The ceramic 650 spitfire ran fine at 3500mhz on Tomsharware videos. It was nitrous cooled, but it ran 3dmark 2001 just fine. 5.4 times stock speed.
I meant specifically K6. K7 (Athlon and Athlon XP) was an entirely different story.
I still have an Athlon XP in a tiny Shuttle case somewhere, with the underrated ati 8500LE.
Talking bullplop as usual “captain”. AMD had the performance crown from the athlon was launched until the core2 era. Year by year amd has had more pwerformance crowns that people can actually buy and not just paper releases. than intel. It also had the fastest 486 market before intel launched the pentium, and you also forget that AMD made in the US the NEC 20mhz intel compatible 8086 cpu and the 8088 cpus. Rough history: AMD has the faster and better product from 1980 to 2006, regains the crown with amd_64 and crushes the netburst architecture plans. Then ragains the performance crown with Zen and has not really lost it since….. Yeah that is some real intel dominance right there, and perhaps that is why they resorted to cheating, paying retailers not to stock amd?
I like my quad intel 44 core. But now it is in storage, since the amd uses less power and has better ipc.
I say “AMD’s K6 wasn’t faster than Intel at the time” – and you say K7 was faster than Intel, and what I said is bullplop… I thought this blog had smarter commenters…
It’s not a solid strategy at all. Every time AMD has targeted the midrange they always end almost going out of business. Only when AMD have been bold and have been able to match or surpass Intel have they ever thrived.
I worked in hardware building and reparing computers from 1998 to like 2002 and had the opportunity to test a lot of weird CPUs like AMD K6, Cyrix and VIA C3. Everything was fine and dandy until it came to FPU performance. None of these alternative brand CPUs could run the then mostly software rendered shooters like Quake. That is a major reason they failed.
It got much better with the Athlon XP and Duron though, and by then Cyrix/VIA was largely out of the game.
JTN,
I remember hearing this conventional wisdom too. But other benchmarks didn’t have AMD FPU as bad as some remember. I cite this comment just to show an opposing view.
https://www.vogons.org/viewtopic.php?t=34666
The authors of quake hand optimized the assembly code, interweaving FPU instructions around intel pipelines specifically, giving them an advantage over AMD CPUs which didn’t get to benefit from a similar quake optimization. The popularity of the quake engine probably cemented AMD’s reputation for being slow though.
Nvidia doesn’t make RTX 4090 GPUs to make money, they make them for the halo effect that trickles down to the rest of the product line. It’s the same reason VW produced the Bugatti Veyron despite taking a loss for every one sold.
And this is one of AMD’s two biggest problems: They don’t understand GPU marketing: Halo cards sell the mid-tier cards. Their other biggest problem is that they don’t understand the importance of drivers and software-defined features (it’s why Nvidia was first to release a Direct3D 9-capable multi-GPU with SLI, first to do video decoding in the GPU, first to integrate PhysX in a GPU, and first to GPGPU with Cuda).
I’m totally a victim of halo effect. When I’m buying a new anything haha
I Want the top one…who’s the best…oh, not in my budget, what’s their model down… maybe one more down… bingo.
Why does my laptop has an intel Arc graphics? They ran out of steps down haha
Another point what you said made me think about. Game development. They buy the latest cards to target, knowing that by the time the game is done, it’ll be fairly midrange. So you’ll probably see even more game companies targeting Nvidia as a result. If you target today’s midrange, your not going to be AAA in 2 years time when the thing is finished.
Top end GPUs don’t really turn mid-range any more. I’m still running an RX 580, and it still performs stellar. Even newer **60 range cards can’t compete. The only problem now is that some games are starting to actively disable support for that card, which really sucks. So, obsolete before mid-range. It used to be that way – it used to be that if you buy a **80 level card, it’ll last you a few generations, then the **60s and **70s would beat it (though they’d never last as long as an **80 level card). It doesn’t seem to work that way any more.
kurkosdr,
I would think they’re profitable, even if it’s just binning and pairing them with more memory. Of course to AMD’s point, building up marketshare may be much more critical for the company right now and they have better chances in the mid-market
AMD could have built out their own proprietary cuda, who knows how that could have turned out. They would have had trouble beating nvidia’s cuda to market after purchasing ATI. For better or worse they went with apple’s opencl and at the time it looked like it could become the future standard. Many hardware companies were signing on including nvidia themselves. I think cuda had the edge because they focused on high end markets.
Binning can be pretty expensive when you have a huge die with “risk” yields and want to produce marketable quantities of flawless dies. So, in a sense, it’s the midrange RTX 4000 series cards that subsidize the RTX 4090. It’s also the reason why the GeForce 400 series had zero cards in the lineup with a perfect die, even the top-of-the-line GTX 480 had 32 of the 512 stream processors fused-off. You see, yields were atrocious for the first Fermi generation, so a card with no fused-off stream processors would be either very expensive or the other cards in the lineup would have to be very expensive (note that in Fermi stream processors are fused off in clusters of 32). I don’t know if Nvidia makes money on the RTX 4090, but there is a chance they don’t.
AMD did have their own CUDA, they called it Close To Metal, which was released as the first-generation AMD Stream SDK.
https://en.wikipedia.org/wiki/Close_to_Metal
But it was a proprietary GPGPU API in a market already dominated by another proprietary GPGPU API (Nvidia CUDA), and predictably everyone who was locked into Nvidia CUDA saw no reason to switch. At least OpenCL is a vendor-neutral standard, so there is that incentive to switch.
kurkosdr,
I suspect it is more expensive to have more production lines, but if they use fewer lines with binning it’s ok.
It’s likely they share the same production line with lower end skus, The best become 4090 while imperfect dies can be sold for lower end cards. Depending on supply and demand, they might even be selling 4090 grade dies as lower end skus, but we don’t know this when we buy them. I’m sure they’d rather sell them as higher end skus.
Interesting, I never heard of it. Wikipedia is a bit unclear. .
“CTM’s commercial successor, AMD Stream SDK”
“AMD subsequently switched from CTM to OpenCL”
CTM was ATI’s stack, but AMD replaced it with AMD Stream and/or OpenCL. From the sounds of it, CTM wasn’t production ready.
I seriously contemplated amd+opencl myself, but nvidia had a hardware advantage. I don’t like the fact that nvidia’s stack is proprietary, but I find cuda works well and I do like working with it. I wish it were open source though.
According to Wikipedia, the first version of OpenCL was released at August 28, 2009, and the relevant working group didn’t even exist before June 16, 2008
https://en.wikipedia.org/wiki/OpenCL#History
so, we can safely conclude that AMD could not have access to OpenCL before June 16, 2008 (Apple maybe, AMD not).
So, I think that what Wikipedia means in the following page:
https://en.wikipedia.org/wiki/Close_to_Metal
is that Close To Metal was the API the first version of the AMD Stream SDK used. Remember, the first AMD Stream SDK was released on December 2007, and AMD could not have access to OpenCL by then:
https://web.archive.org/web/20071223181400/http://ati.amd.com/technology/streamcomputing/register.html
So it had to use Close To Metal as the API.
So yes, Close to Metal was a real thing, but nobody used it, so later they switched the SDK to use OpenCL.
More history that corroborates this:
https://en.wikipedia.org/wiki/AMD_FireStream#Overview
So, yes, AMD had their own CUDA.
kurkosdr,
I understand that AMD could have gone with their own proprietary solution, but wikipedia suggests that CTM was only a beta and that AMD Stream needed a complete rewrite. AMD likely started it right after the acquisition. But I’m not sure if wikipedia’s summary is entirely correct.
This article from that time period may reveal more about what they were thinking on OpenCL…
https://www.tomshardware.com/news/AMD-stream-processor-GPGPU,6072.html
Partnering with apple on opencl may have been a mistake in hindsight. It’s clear AMD were expecting apple to stay the course.
Apple uses CTM widely, even though they have moved to in house gpu. Vulkan is not officially supported for the users but works just fine for devs in MacOS. They even named the metal API after the AMD close to metal API and is VERY close in terms cof compatibility. Some cases identical.
Both AMD and Nvidia halo GPU have some of the higest margins of all the GPUs…they absolutely make sense to make.
All rtx does is making games loog shitt, kinda like post processing and most shader usage. I want pure raster performance and even the lowly 6700 is on par with the 4090 in that regard. I also want free drivers, nvidia requier a hughe blob. Not acceptable.
Nvidia has untold powers in the HDMI consotrium and banned AMD to ban open source drivers for linux can not have 120+hz at 4k or 8k resolutions (this can not be legal, right?)
FWIW NVIDIA makes a very nice margin out of the 4090s.
Reminds me of S3 graphics, and all of the others that died in the the late 90’s early 2000’s. Nvidia /AMD were too good but wer’e going to make a GPU thats for everyone else that wants ok performance at a reasonable cost. And they worked well with oems for a bit, but with out the R&D that was aimed at the best, years old Nvidia AMD was soon as cheap and better than they were. I think that may well happen again if they aren’t careful. There saving grace might be that Nvidia may choose to produce in lower quantity and may demand a premium based upon their role as a leader.
I have a S3 delta chrome V8 dual GPU… yeah it performs at best on the level of a XGI 500 or a SiS Xabre 400. The XGI volari v8 at least beats the radeon 9800 in warcraft 3, but so does the Xabre 600 and 800. Ask me how well the XGI or xabre cards run with a non UT or ID tech game like Shogo, Blood2 and many others… no they dont, no they just does not run at all. There are some patches for the Xabre and some for DeltaChrome 8. yeah they still crash for me. Drivers in linux suck for deltachrome, but xgi and xabre cards is decent and works.
Quite a few comments are missing an important part of the story: fabrication capacity and the AI/server market.
Per surface of silicon wafer, the server market pays more than the “client” market. and AI pays WAY more. So AMD would rather focus on selling the Instinct lineup than spend fab capacity on low volume high-end desktop GPUs. AFAIR there was a fabrication capacity bottleneck with some AMD products (not sure if GPUs) and they decided to lean more towards the data center side.
Everybody was having capacity issues during the supply chain collapse of COVID.
All that stuff has long been resolved. Capacity is not the problem that is limiting AMD’s GPU strategy. The RADEON group has been on life support for almost a decade by this point. They simply can’t compete against NVIDIA which has almost double the amount of design resources for consumer GPUs alone.
NVIDIA is just too good when it comes to execution, and they also have a stablished software ecosystem that is pretty hard to overcome. Whereas AMD has always dealt with software as a second or third class citizen.