Discussion RDNA4 + CDNA3 Architectures Thread

[DisEnchantment]

With the GFX940 patches in full swing since first week of March, it is looking like MI300 is not far in the distant future!
Usually AMD takes around 3Qs to get the support in LLVM and amdgpu. Lately, since RDNA2 the window they push to add support for new devices is much reduced to prevent leaks.
But looking at the flurry of code in LLVM, it is a lot of commits. Maybe because US Govt is starting to prepare the SW environment for El Capitan (Maybe to avoid slow bring up situation like Frontier for example)

See here for the GFX940 specific commits

History for llvm/lib/Target/AMDGPU - llvm/llvm-project

The LLVM Project is a collection of modular and reusable compiler and toolchain technologies. - History for llvm/lib/Target/AMDGPU - llvm/llvm-project

github.com

Or Phoronix

More AMD "GFX940" Enablement Work Landing In LLVM - Phoronix

www.phoronix.com

There is a lot more if you know whom to follow in LLVM review chains (before getting merged to github), but I am not going to link AMD employees.

I am starting to think MI300 will launch around the same time like Hopper probably only a couple of months later!
Although I believe Hopper had problems not having a host CPU capable of doing PCIe 5 in the very near future therefore it might have gotten pushed back a bit until SPR and Genoa arrives later in 2022.
If PVC slips again I believe MI300 could launch before it

This is nuts, MI100/200/300 cadence is impressive.



Previous thread on CDNA2 and RDNA3 here

Question - Speculation: RDNA3 + CDNA2 Architectures Thread

Man I have been dying to make this one for a while now. First rumours for RDNA3 are here so new thread time! Just going to start off with this one for now: kopite7kimi on Twitter: "@VideoCardz Ah, I mean a simple mcm design with 10240 cores is not enough. Because the lift from RDNA2 to RDNA3...

forums.anandtech.com

 

[SolidQ]

AMD can't do Nvidia -$50 with these 2 cards

Azor was saying, they not going minus 50$ strategy

 

[inquiss]

Assuming equal RT performance, and FSR4=DLSS3 image quality, I would start by taking ~20% off Nvidia price to account for the massive improvement in DLSS4. That would be "MSFP" $599, street price ~$700-750 assuming they want to keep their ~10% market share and not kill Radeon off. If they want to get back to >30% market share like they claim, they need to knock another ~20% off that price to convince RTX owners to switch. If RT is worse than 5070 Ti, or if FSR4 is more like DLSS2, then adjust price accordingly. Or, if they pulled off a miracle and have caught up to DLSS4 transformer model upscaling, then no 20% discount.

This mindset is the problem with PC gaming. I don't mean that as a personal attack. That said, I don't think I could have written a better post explaining how effective Nvidia marketing is. "The next great thing" from Nvidia is always a reason why AMD is inferior and therefore deserve to have cheaper cards.

 

[inquiss]

If they do that, I'll buy it. Otherwise, it's not worth switching to AMD.

No you won't. You'll just buy nvidia. Or Nvidia drops the price and you buy nvidia. Ergo, higher prices for everyone. It's a good thing. Gamers get fleeced and stonks go up.

 

[inquiss]

That is historically accurate. And I said the same thing about CPUs for years and years. Yet that worm eventually turned. Nvidia dropped the ball hard this time. And the 💩 show keeps expanding day after day. Who knows, maybe this will be the straw that breaks the camel's back? Enough jilted gamers might finally give Nvidia the 🖕They are certainly talking a good game about jumping ship on the reddit subs. We can dream...

I hope so, but the brand isn't really there and Nvidia always has the "new thing" that makes cards incomparable in the eyes of the gamer. CPUs are CPUs. Faster better. Many more metrics to compare GPUs with, especially now. More ways for Nvidia to be better.

 

[inquiss]

IMO AMD can't do Nvidia -$50 with these 2 cards. That might work when you're looking at prices between $300-$500. But at $800-$1000 people aren't saving enough to go AMD. There has to be more incentive.

Right now the incentive is that Nvidia is unavailable and the connector melts. Might be enough

 

[adroc_thurston]

If RDNA4 really does give 50% uplift over RDNA3 especially if it does so without huge increases in clocks

It's like 15-20% higher fmax.
Rest yeah, microarchitecture.
But your fmax is also a function of microarchitecture.

 

[Heartbreaker]

Well not sure how they did it but they managed somehow. Not exactly sure how that would affect Navi 48 density but here’s some napkin math.

GB203 = 120 MTr/mm2
120 + 27% = 152 MTr/mm2

Which is pretty close to what that leak claims:

I think it's much more likely that poster on a Chinese forum references AnandTech forums, guesstimate of the 350 mm2 die size is just wrong. It's more likely that the people estimating ~390 mm2 where closer.

 

[adroc_thurston]

guestimate of the 350 mm2 die size is just wrong.

Wdym wrong.

 

[In2Photos]

Right now the incentive is that Nvidia is unavailable and the connector melts. Might be enough

The connector isn't an issue on the 5070 series cards. Nobody considering buying a 5080 or 5090 is getting a 9070.

 

[Mopetar]

Wdym wrong.

40 mm^2 of the die is interpolated. AMD has really caught up with Nvidia tech!

 

[Ghostsonplanets]

I think it's much more likely that poster on a Chinese forum references AnandTech forums, guesstimate of the 350 mm2 die size is just wrong. It's more likely that the people estimating ~390 mm2 where closer.

A reviewer with the silicon in hands said it's smaller than GB203.

 

[HurleyBird]

If the die is 350mm^2 and the 9070 XT performs like a 5070 ti, RDNA4 would actually have better PPA than Blackwell.

AMD has come a long way from when the 232mm^2, 150W, 256bit, RX 480 was losing to the 200mm^2, 120W, 196bit GTX 1060.

PPA gets sort of messed up as a metric when you decide to go with slower memory and compensate with big globs of cache. Apples and oranges versus parts that make the opposite decision.

Which of course only makes RDNA4 more impressive than naive PPA would suggest.

 

[Hitman928]

Zen 5 CCD is ~27% denser than Zen 4 on a similar node. I think most of that came from cutting the L3 cache size in half thanks to TSMC FinFlex?

A big part of it is they switched to 6t cache cells.

 

[adroc_thurston]

A big part of it is they switched to 6t cache cells.

That's for L1 SRAM.
L3 went to 6T HD all the way back when (Zen3).

 

[Hitman928]

That's for L1 SRAM.
L3 went to 6T HD all the way back when (Zen3).

First reference I could find to 6T cells being used is for Zen 4c but I'm also not going to take the time to dive deep for confirmation, lol.

 

[Elfear]

The connector isn't an issue on the 5070 series cards. Nobody considering buying a 5080 or 5090 is getting a 9070.

*Raises hand*

Before the 5000 series benchmarks came out and prices went into the stratosphere, I was definitely planning on a 5090 or 5080. If the 9070XT can be purchased for $600-700, it would be a lot more compelling card than a $1,500 5080 or $3,000 5090 . I would plunk down for that no problem.

 

[marees]

I believe MSRP will be higher than $550. This is because AMD launched the 7900gre as 1440p max card but now they are comparing the 9070xt to it at 4K. Their targets seem to have changed

& I agree with below take:

I won't be a bit surprised if the AIBs launch the 9070XT $750+. There's no BBA version this time. The competition is OOS and probably will be for a while, so they'll sell well even at $800 to people who can't get an NV card. I think it'll be the usual mix of MSRP and AIB "custom" models with the MSRP versions being undersupplied and sold out until supply catches up with demand.​

​

As for AMD's "official" MSRP, I think they need to price them based on raytracing performance if they want to take a chunk of market share from NV. Games are slowly starting to require RT, and I think we're going to see more and more of that as the non-RT capable cards age. We're already looking at more than 4 years old for models that can't do RT. Implementing rasterized lighting, shadows and reflections in addition to the RT versions means additional work, which will become less and less worthwhile as older GPUs are retired.​

​

Leaks/rumors suggest big improvements in RT but not enough to catch NV. 9070XT seems likely to match a 4080 or 5070Ti in raster, but leaks suggest it'll be a bit behind in RT, more like a 4070Ti Super. Ballpark 10-15% slower than a 5070Ti in RT. If that's the case RT price parity would be around $650, $599 if they really want to move some volume. Then the 9070 is looking to be all around faster than a 5070, so both cards are looking likely to land between the 5070 and 5070Ti. Given that situation I think $500-550 would be good for the 9070. That won't happen at launch unless you're really lucky and manage to score one of the "MSRP" models that will be available in minimal numbers. They're all going to be AIB cards and most of them will be "enhanced", OCed a wee little bit, and they'll charge a bunch extra until supply catches up with demand.​

 

[Jan Olšan]

That's encouraging that AMD is asking for feedback on pricing. Hopefully they listen to Steve from HWUB - aggressive undercutting with high supply would win them a LOT of mindshare right now. Especially because anything over RTX 4060 level is basically unobtanium or stupidly priced.

Asking for opinion doesn't mean the opinions may change much. If you ask customer (and reviewer, too) there is always going to be an "I want a pony" element to the answer, and there's always going to be a conflict bteween the vendor's and buyer's ideas of prices, inevitably.

Right... cause selling 1milion units for 699$ would be better than 3-4mil at 549 or 6-7mil at 499$

It can easily be better. Imagine a situation where the cost to make is such that ingoring the AIB's margin, AMD gets $40 op. income per card at 499$ and 240$ op. income at 699$.

You gain marketshare, true. In theory. But look how far will people go to justify getting Nvidia for more. So the big issue is that the firesale prices won't give you 6x the sales if you cut your margin by 6. You may not even double the sales". FFS, making the card 200$ cheaper won't even stop people from repeating that "it's always Nvidia - $50" meme!

If the die is 350mm^2 and the 9070 XT performs like a 5070 ti, RDNA4 would actually have better PPA than Blackwell.

AMD has come a long way from when the 232mm^2, 150W, 256bit, RX 480 was losing to the 200mm^2, 120W, 196bit GTX 1060.

It seems RX 9070 XT will end up at under 5070 Ti performance, overall (some games may be faster, average not). It's tricky to compare since 5070 Ti has extra external cost (GDDR7) but also is harvested chip so you can't really point at its die size to prove anything. Would have to compatere with full fat version of the die (RTX 5080).

That said, this may be something that is favorable to AMD - Nvidia has to use cut-down more expensive silicon to compete at this price or performance part of the ladder. It is reversed for 9070 non-XT and and RTX 4070 though. Which means AMD may again price the non-XT less agressively than the XT part, same as it was in RX 7900 line. To get more sales of the better-margin XT part. Hopefully the lower TDP of 9070 non-XT can provide enough savings on VRM and PCB costs so that the cost factor is a bit offset.

 

[basix]

Reality since 2017 says otherwise.

And what happened? Many people are still using older cards, reduce game quality settings etc.
If you, for example, could get a 3080 for MSRP (some people were lucky) or could buy one 2nd hand for a good price you just sit there and wait, even if only 10GB of VRAM gets tight today. There is nothing on the market, which is tempting. You only buy because 8/10GB get just too less for decent quality settings in certain titles.
Many people do not buy the new things because it is just not reasonable.

Sure, quite some people did bite the bullet and did buy new cards (including myself). And the performance king will always be somewhat out of competition, because there are many people who just want to have the very best. But also those begin to think deeper about buying new cards.

For example just looking at myself (sorry if that are only Nvidia cards in the RDNA4 thread, but AMD had not the performance king since ages):

• 980 Ti was a nice card
• 1080 Ti was nice as well, but did already cost quite more than I liked to have it back then. So I skipped that generation.
• 2080 Ti was expensive but delivered cool new tech (I like that, I am a geek). That tech was only really usable from ~2020 and onwards, when DLSS2 and more RT Games were released but in hindsight the 2080 Ti delivered well. Peformance jump over 980 Ti was >2x for <2x price. I primarily accepted the higher price because 980 Ti was as its limits and because of the cool new Turing tech (but that is a personal choice)
• 3080/3090 was OK regarding performance, but delivered no enhancement in power efficiency, 10GB was a VRAM degradation and cards were generally overpriced (Mining, Covid). RDNA2 cards were nice but struggled at RT and had no DLSS2. So I skipped that generation.
• 4090 was very expensive as well (I paid ~1.7x more than the 2080 Ti for it), but delivers >2.5x performance, much better energy efficiency, new features, 24GB. So I bought one at launch. I also use it for local ML/AI stuff so 24GB, CUDA and Tensor-TFLOPS come in handy. RDNA3 sadly underdelivered and therefore was no choice for me.
• 5090 even at regular prices would deliver no performance per dollar increase compared to the 4090. No power efficiency gains. And currently you could not even buy one at reasonable prices. So I will skip that generationa as well.
• AMD 9070? Hey, if it performs well and does not cost too much I might consider it as an option in my 2nd PC and I will 100% recommend it to anyone who asks me for GPU buying advices (online or in real life). Definitely not any of the Nvidia cards right now (which I also not have done for many Nvidia cards since Turing because of the high prices or e.g. just 8GB VRAM). I always recommend performance per dollar cards with enough VRAM (=lives longer) unless someone specifically asks for some features. And the "features" point might look very good for RDNA4 as well (FSR4, better ML/AI support and hopefully also applications like Blender etc.), unless CUDA is required

So even I, who has thought of buying new cards each generation, likes new tech and also buys at the upper end of the stack, just do not buy everything which gets thrown onto the market. Everything else is simply stupid.
And people with a more limited budget are likely to sit the current situation out as well, if their GPU is not already pushed over its limits. They just buy if they "must". They probably want to but do not because not affordable.

 

[marees]

I agree with this take too (despite how Jack Hyunh spinned it)

https://twitter.com/x/status/1893216826588381579

https://twitter.com/x/status/1893297350170255464

 

[Jan Olšan]

I agree with this take too (despite how Jack Hyunh spinned it)

https://twitter.com/x/status/1893216826588381579

https://twitter.com/x/status/1893297350170255464

Yes, well put, agreed.

I would like to underscore the bit about Nvidia being sure to react.
Currently many people think or claim that Nvidia's prices are force majeure and we should be glad that Nvidia is still making GPUs at all, because AI market would take them and pay their weight in gold, only "papa Huang" preventing it out of his good heart. Apparently even some Nvidia people spread that take when they talk with media people, informally. (It's possible they feed it to influencers and other people that will listen in a wider pattern - no idea, I only have anecdotal experience)

People forget that the gaming market is still a multi-billion business for Nvidia that is income and revenue accretive, and what's more, it's a safeguard and hedging against any bubble bursts or other unforeseen or even foreseen issues in the AI market. No sane company will abandon it. Also, leaving it gives AMD stronger position to then challenge them in their (newly) core market.

So no, Nvidia only wants you to think they don't want your money so that you are fine with high prices, but if the business was at risk of going to AMD, the would react quickly.

You just have to look at the prices of GeForce cards few months ago - they were fine and they even gave you the 4080super at better price than 4080. The current pricing situation has mostly single reason: Nvidia not preparing enough stock for launch and having stopped RTX 4000 production too soon so that stock of that doesn't kill the demand for the new generation. It's price inflation caused by artificial scarcity, and thus most likely temporary price inflation.

I don't know how long will it last last and how long will the RTX 5000 prices stay up. 2 months? 3 months? Dunno. 6 months? That already sounds way too long. In any case, they won't let AMD eat away at their gaming business much.

(edit: if those "gaming market is not important to Nvidia anymore" cope stories were a real concern, do you think they would take Nintendo's Switch 2 business, which likely has ever worse margins than GeForce?)

 

[DAPUNISHER]

So the consensus is Nvidia MSRP -10% again?

 

[marees]

So the consensus is Nvidia MSRP -10% again?

$630 msrp probably with a few models in stock at that price

Rest will be $700 plus

Between $550 & $630 no one will be happy

And I don't see them pricing as low as $550

 

[gaav87]

9070xt first party soon

 

[coercitiv]

I agree with this take too (despite how Jack Hyunh spinned it)

100%.

AMD has been playing it safe for the last couple of years with both CPUs and GPUs. Even with a winning product like the 9800X3D (and the healthy demand for 7X3D and 5X3D) they still lacked the confidence to push hard on the production pedal. They definitely played it safe with RDNA4, so IMHO production has at best ramped up reactionary, in the light of recent developments.

This is why I said that AMD should price 9070 XT like the 6800XT (+/-) if it performs well vs. competition, and more like 7800XT if not. All info we have is the chip is quite competitive, assuming the software stack is also up to snuff. This means that 6800XT pricing is likely the floor for 9070 XT, the final price may be higher. My problem with AMD is I doubt they have decent supply even relative to their standards, and I also doubt they will resist the urge to push the price higher. I'm afraid their own FOMO will win, they will rationalize higher prices just like members on this forum did, and the product will fall flat on it's face.