Discussion RDNA4 + CDNA3 Architectures Thread

DisEnchantment · Mar 23, 2022

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

Timorous · Apr 10, 2024

branch_suggestion said:
Transistor count is just a fun metric at this point, total die area and what perf it has are the relevant metrics.
N31 has a bunch of things that N33 does not, basically a lot of little things that all add up, and bigger chips have certain things that scale disproportionate vs smaller chips.

That is not true for Ada.

AD102 is 76.3M xtors with 144SMs, 96MB L2 and a 384 bit bus with 125.3M xtors/mm
AD103 is 45.9M Xtors with 80SMs, 64MB L2 and a 256 bit bus with 121.1M xtors/mm
AD104 is 35.8M Xtors with 60SMs, 48MB L2 and a 192 bit bus with 121.8M xtors/mm
AD106 is 22.9M Xtors with 36SMs, 32MB L2 and a 128 bit bus with 121.8M xtors/mm

AD104 is almost exactly 75% of AD103
AD106 is almost exactly 50% of AD103
AD102 is almost double AD104 but leans a bit more compute heavy hence the higher density due to the logic/cache/io ratio being a bit more skewed to logic which shrinks better than cache or IO.

N31's use of transistors is baffling when N33 can give you 33% of the spec at 23% of the transistor cost. Sure I expect it to be slightly worse than Ada because of the GCD - MCD links bit not to the tune of 18B transistors.

If you scale up N33 then you can get 64CUs, 256bit, 64MB L3 in 26.6B transistors which would need a density of 110M xtors/mm to work in a 240mm die. AMD have done 140M xtors / mm in Hawk Point and while the logic/cache/io split leans more towards logic in that config I don't see why matching NV at the minimum or hitting around 130 at the high end is out of the question so even in 240mm you have a bit of room to spare. This also says nothing about trimming any fat there might be in RDNA 3 although I tend to assume roughly the same number of transistors per CU when doing this roughing out because while they might trim the fat in places they usually add or improve other features else where and it often comes out in the wash give or take.

The other reason to scale up from N33 rather than down from N31/N32 is because N33 is monolithic like N44 and N48 are going to be so it just removes a variable and is more like for like.

branch_suggestion · Apr 10, 2024

Timorous said:
That is not true for Ada.

Yes, all monolithic and all have the same exact units to play with.

Timorous said:
N31's use of transistors is baffling when N33 can give you 33% of the spec at 23% of the transistor cost. Sure I expect it to be slightly worse than Ada because of the GCD - MCD links bit not to the tune of 18B transistors.

N33 does have 1 cutdown unit outside of uncore, vGPR is 128KB instead of 192KB with the chiplet parts, that adds up fast.

Timorous said:
If you scale up N33 then you can get 64CUs, 256bit, 64MB L3 in 26.6B transistors which would need a density of 110M xtors/mm to work in a 240mm die. AMD have done 140M xtors / mm in Hawk Point and while the logic/cache/io split leans more towards logic in that config I don't see why matching NV at the minimum or hitting around 130 at the high end is out of the question so even in 240mm you have a bit of room to spare. This also says nothing about trimming any fat there might be in RDNA 3 although I tend to assume roughly the same number of transistors per CU when doing this roughing out because while they might trim the fat in places they usually add or improve other features else where and it often comes out in the wash give or take.

Add the beefy parts and you go from 110 to 130 real quick.

Timorous said:
The other reason to scale up from N33 rather than down from N31/N32 is because N33 is monolithic like N44 and N48 are going to be so it just removes a variable and is more like for like.

Yes, but you really can't scale from either properly, N33 being N6 means different rules apply e.g. dark Si spam is more prevalent in N5 class nodes than N7.

Timorous · Apr 10, 2024

branch_suggestion said:
Yes, but you really can't scale from either properly, N33 being N6 means different rules apply e.g. dark Si spam is more prevalent in N5 class nodes than N7.

It just gives you a starting point. IE if you can't get your spec from simple scaling it is likely BS. In this case 64CUs, 64MB, 256bit seems entirely doable within a 240mm die area for a monolithic part.

Another thing to look at is that It is very very similar spec wise to the 4070Ti Super which is around 82.5% of full AD103 so is around 38B transistors. Given AMD don't dedicate much of any die area to tensor cores It also flows that getting 4070Ti Super performance from around 30B transistors is probably possible.

So it seems at a rather cursory and simplified glance a 64CU 64MB 256 bit design should fit in the 240mm mentioned and that kind of design should be capable of 4070Ti Super / 7900XT performance provided the clocks are high enough.

Of course something being possible and AMD actually executing on it are not the same at all so we still wait and see if they actually deliver. Does not seem beyond the possible though which I guess is good.

ToTTenTranz · Apr 10, 2024

Mahboi said:
I don't think the PS5 Pro will change that all that much either with their higher price, even with RDNA 3/4 in it.

If the SoC is indeed on N6, I doubt it will cost that much more than the original PS5.
My guess is we're looking at a 320-350mm^2 SoC, up from the current 260mm^2, and that's about the only significant difference in production cost for Sony. GDDR6 18Gbps are probably not much more expensive than 14Gbps nowadays, perhaps they're getting some $2 worth of extra 2-4GB DDR4 on the SSD controller (which is now a cheap off-the-shelf 8-channel controller instead of their custom 12-channel one), and perhaps a slightly beefier power supply and voltage regulators.

Don't be surprised if the new console costs $500 while they pull the original PS5 to $400 or lower, or +$60 for each if you want the bluray reader in it.

Ghostsonplanets · Apr 10, 2024

ToTTenTranz said:
If the SoC is indeed on N6, I doubt it will cost that much more than the original PS5.
My guess is we're looking at a 320-350mm^2 SoC, up from the current 260mm^2, and that's about the only significant difference in production cost for Sony. GDDR6 18Gbps are probably not much more expensive than 14Gbps nowadays, perhaps they're getting some $2 worth of extra 2-4GB DDR4 on the SSD controller (which is now a cheap off-the-shelf 8-channel controller instead of their custom 12-channel one), and perhaps a slightly beefier power supply and voltage regulators.

Don't be surprised if the new console costs $500 while they pull the original PS5 to $400 or lower, or +$60 for each if you want the bluray reader in it.

SIE Chairman Totoki already said that the hardware business costs haven't go down like it used to and that they don’t have much margins on hardware. And that he expects to improve margins this FY. So I doubt PS5 will drop in price. I expect PS5 at $499, maybe $449. And PS5 Pro at $599/$549.

Ultimately, Nintendo has shown to others console manufacturers that you don't need to drop console price anymore throughout the generation.

ToTTenTranz · Apr 10, 2024

Ghostsonplanets said:
Ultimately, Nintendo has shown to others console manufacturers that you don't need to drop console price anymore throughout the generation.

"Ultimately" means "during COVID" where people were stuck at home with money to spare during a semiconductor crisis that made every electronic equipment more expensive worldwide.
That era is gone. Discrete GPUs are now going down in price (except China, because of sanctions) and Sony missed their 2023 PS5 sales target by 4 million units. Sony isn't going to be bullish on the PS5 Pro's price.

As for Sony wanting to improve margins this FY, they were talking about 1st party game development costs ballooning out of control. Which they clearly did as we saw from Insomniac's leaked documents.

Ajay · Apr 10, 2024

branch_suggestion said:
Transistor count is just a fun metric at this point, total die area and what perf it has are the relevant metrics.
N31 has a bunch of things that N33 does not, basically a lot of little things that all add up, and bigger chips have certain things that scale disproportionate vs smaller chips.

Not just fun - generally speaking, performance is proportionate to the # of xtors. From a product viability standpoint (profitability), PPA is what matters most - with area being the big one wrt cost. As enthusiasts, we are interested in more than profitability, and legitimately so. Companies are strictly focused on the later, and legitimately so. Hence the often repeated complaints about companies not giving us what we want [ed] performance wise at the prices we want (rinse and repeat).

Mopetar · Apr 10, 2024

eek2121 said:
Me vs. Kepler, I'm gonna let this comment stand and double down. This will be fun. 😍

Cat Food Bet!

Cat Food Bet!

RnR_au · Apr 12, 2024

Not sure, but seeing some Sapphire price drops in Australia, maybe price drops as well in Europe and USA?

So are the RX 8000 series already in production? Or just the usual slow downwards price trends on existing product lines?

jpiniero · Apr 12, 2024

RnR_au said:
Not sure, but seeing some Sapphire price drops in Australia, maybe price drops as well in Europe and USA?

I looked at Newegg, and only the 7700 XT and 7900 XT are basically below MSRP... and the 7900 XT doesn't have a lot of options. Although I guess, how many options do you really need.

Mahboi · Apr 14, 2024

I just thought, we're going to have a 130mm² die covering the 6-class cards/1080p gaming with RDNA 4.
Which means that the fight for 6 months is going to be between AD107, a 4050 Ti's die, marketed as a 4060, vs Navi 44, a 8500 xt's die.
6 class cards are now a 130mm² vs 159mm² duel.

Says so much about the state of the GPU market and the margins they're making.

And the complaints about "yet another 8Go card" aren't going to calm down...

Glo. · Apr 15, 2024

Mahboi said:
I just thought, we're going to have a 130mm² die covering the 6-class cards/1080p gaming with RDNA 4.
Which means that the fight for 6 months is going to be between AD107, a 4050 Ti's die, marketed as a 4060, vs Navi 44, a 8500 xt's die.
6 class cards are now a 130mm² vs 159mm² duel.

Says so much about the state of the GPU market and the margins they're making.

And the complaints about "yet another 8Go card" aren't going to calm down...

Those dies should have been integrated into APUs/SOCs years ago.

TESKATLIPOKA · Apr 15, 2024

Glo. said:
Those dies should have been integrated into APUs/SOCs years ago.

Without enough BW It would be meaningless.
And who would pay for such a big APU?

Glo. · Apr 15, 2024

TESKATLIPOKA said:
And who would pay for such a big APU?

ME!

And Dell, Asus, MSI, etc.

MrTeal · Apr 16, 2024

A 12400 is only 157mm², and that includes the existing iGPU. Integrating a 130-160mm² 4060 class GPU onto that die not only doubles the size of the die, but even at DDR5-8000 the 128GB/s bandwidth is less than half the 272GB/s a 4060 has on its own. It's definitely a cool product, but unless you're investing in a dedicated platform for it similar to a console, it's super niche.

You'd be better off just joining the hacking scene and trying to jailbreak a PS5 to load up Windows on it.

Mopetar · Apr 16, 2024

Mahboi said:
And the complaints about "yet another 8Go card" aren't going to calm down...

A GPU that small should be $200 at most and is probably targeting laptops as much as it is desktops. I think they could have made it slightly larger to give it either 10 or 12 GB capabilities, but perhaps they plan on using 3 Gb memory chips to enable a 12 GB model if they want to try making a slightly more premium model.

If they keep the price to $200 and under, I don't think people will complain too much. APUs are getting good enough to where there's not really a market at the extreme low end for discreet cards. 8 GB is actually enough for e-sport titles and that's all some people really care about or that's all their kids are going to play for the most part. Then again, an APU is generally good enough for those as well.

jpiniero · Apr 16, 2024

Mopetar said:
A GPU that small should be $200 at most

See people seem to be stuck in 2005 when it comes to die size versus board prices.

I would say "Ideally" N44 would be Mobile/OEM only. And AMD really needs to get back in mobile.

Mahboi · Apr 16, 2024

Mopetar said:
A GPU that small should be $200 at most and is probably targeting laptops as much as it is desktops. I think they could have made it slightly larger to give it either 10 or 12 GB capabilities, but perhaps they plan on using 3 Gb memory chips to enable a 12 GB model if they want to try making a slightly more premium model.

They're not using GDDR7, so no 3Go memory chips for us. But that'll definitely be the case for later generations.
Also, the card is expected to run between a 6700xt and 7700xt's performance. Which means that they'll definitely sell it for $300.

Sad state of affairs. 7600 is already insultingly bad at $270 for 2024, I assumed they would lower prices once 6600/6600 xts are gone, but no, it's still that high up despite being a $200 card in every way.

Aapje · Apr 17, 2024

I expect that N44 will be sold as a clamshell 16 GB version and a cut down 8 GB version.

Timorous · Apr 17, 2024

Aapje said:
I expect that N44 will be sold as a clamshell 16 GB version and a cut down 8 GB version.

If it performs as claimed (faster than 7600XT, slower than 7700XT) then the top model should have 16GB just from an ASP perspective. $350 for it would be a good perf/$ bump over the 7600XT and it would probably be ballpark 4060Ti/3070 performance if the >> 7600XT and < cut N32 are at all accurate. That would certainly warrant more than 8GB of VRAM. A cut model with fewer CUs, lower clocks and a 96 bit bus with 12GB of VRAM would be a better alternative if performance is worthy of a $280-300 price point. With 8GB I think the realistic maximum price is $250 otherwise Steve and Steve will absolute slaughter it in reviews and even $250 is on the higher side, especially if some VRAM heavy games come out in the interim.

phoenix21 · Apr 17, 2024

jpiniero said:
See people seem to be stuck in 2005 when it comes to die size versus board prices.

I would say "Ideally" N44 would be Mobile/OEM only. And AMD really needs to get back in mobile.

First have a look at this

Then correct me if I'm wrong.
If a 122mm² 6nm io die cost just $21 to manufacture and even if 4nm wafer price is 75% more than 6nm, a ~130mm² 4nm die would only cost ~$40. That's very similar cost to a 204mm² 6nm N33.

Now given that performance of N48 with 64CU is more than 7900XT, which is 33% faster than 7800XT(60CU) in 4k( hardware unboxed/techspot data) means 7900XT is equivalent to 80CU with 7800XT clocks, leads to conclusions that RDNA 4 is 25+% faster than than RDNA 3.

Now N44 with half the CU of N48, with slightly higher clocks can be over 33% faster than RX 7600 yet 19% slower than RX 7700XT, any MSRP between $270-$300 is acceptable.

It's so unsettling that AMD again leaving a huge gap between N44 and cut down N48, only if N44 would have 40CU and 192bit, it would make a perfect product stack and also leaving room for another chip for sub $200 cards. It could have been like this -
Navi 48 RX 8800 XT 256 bit 16GB 64CU 3050mhz $500
Navi 48 RX 8700 XT 14GB 224 bit 52CU 3050mhz $400
Navi 44 RX 8600 XT 12GB 192 bit 40CU 3250mhz $330
Navi 44 RX 8600 10GB 160 bit 32CU 3250mhz $270
Navi 42 RX 8500 XT 08GB 128bit 24CU 3450mhz $220
Navi 42 RX 8400 06GB 96 bit 20CU 3250mhz $170

Mopetar · Apr 17, 2024

I wouldn't pay more than $200 for an 8 GB card no matter how it stacks up. There are already plenty of titles where 8 GB isn't enough for 1080p and almost any major title that recently entered production is going to target the current console generation where there's at least 10 GB of fast VRAM available.

Like I said, they should have made it slightly larger to enable up to 12 GB. Not every card has to have that much, but 8 GB is going to really limit the card.

leoneazzurro · Apr 17, 2024

I hate these videos about "manufacturing costs" because they are largely misleading. These are not even covering the full BOM of the product, and they are missing a lot of other activities (including testing and binning) and additional costs (i.e. packaging factory being in another site respect to the diffusion site adds a transfer cost). And these are only the direct costs. Then there are a lot of indirect costs that a company must cover. Granted, AMD makes a margin on the top of these costs. But saying that a 7950X "costs 69$ to make" is a big fat lie, and no one who has even the lowest knowledge of cost accounting will believe to it.

Aapje · Apr 17, 2024

leoneazzurro said:
But saying that a 7950X "costs 69$ to make" is a big fat lie, and no one who has even the lowest knowledge of cost accounting will believe to it.

Neither Ian or phoenix21 have claimed this, and your usage of quotes to make it seem like you are quoting someone, when you are not, is actually the misleading part here.

leoneazzurro · Apr 17, 2024

Aapje said:
Neither Ian or phoenix21 have claimed this, and your usage of quotes to make it seem like you are quoting someone, when you are not, is actually the misleading part here.

A video called "AMD Ryzen 9 7950X: cost to manufacture?" with a 69$ in the thumbnail, when in the first seconds (0:06 ) we heaar "we are going to calculate how much it costs to amd to manufacture this..." and with the conclusion "Your new shiny CPU cost 69$ to make" (13:30) is not misleading?
Did you even see the video?
Yes, after he says this he puts a written "Bill of Material" under the 69$ but before that he says twice that he was to calculate "the cost to manufacture".
These are both WRONG.
First, what he is listing is not the full BOM: sure, these are probably the most expensive parts of the BOM but BOM includes also auxiliary/interconnect materials, external packaging (blisters, carton boxes, down to instructions fliers).
Second, BOM is NOT cost to manufacture. Cost to manufacture includes other things than direct costs of the materials. It includes labour, testing, binning, direct overheads, shipping costs between production sites. And these are only the direct costs of the production, because in the total cost of manufacture there are many other expenses as well, R&D and tooling are among the first.
Yes, there is a disclaimer "I can be wrong ant this may not be the full BOM" at the end. Well, then at least try to estimate and speak about what it is missing. Presenting a flat 69$ figgure is outright misleading.
And no,I am not quoting anyone. These are cost accounting basics.

EDIT: not surprisingly, there are similar criticism to that video in the comments below it.

Discussion RDNA4 + CDNA3 Architectures Thread

Golden Member

Golden Member

Member

Golden Member

Member

Senior member

Member

Lifer

Diamond Member

Golden Member

Lifer

Senior member

Diamond Member

Platinum Member

Diamond Member

Diamond Member

Diamond Member

Lifer

Senior member

Golden Member

Golden Member

Junior Member

Diamond Member

Senior member

Golden Member

Senior member