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

 

[jpiniero]

the 7600 would be a good sub $200 card to replace the 6600. 8GB at that price point would be totally acceptable and N33 is a touch cheaper to make than N23 was so there is little reason why AMD can't sell 7600's for the same price they have been selling 6600's for a while.

I assume RDNA2 is completely out of production and has been for some time.

Don't confuse me advocating for an acceptable compromise that is superior to an 8GB part at $250 or more as thinking I consider it a viable long term solution.

The long term solution for sub-$400 is nothing.

 

[coercitiv]

The long term solution for sub-$400 is nothing.

I thought your long term solution was to stop making video cards altogether.

 

[marees]

I assume RDNA2 is completely out of production and has been for some time.

N24 (6500xt & 6400xt ) are on 6nm. So there is nothing stopping their production

The long term solution for sub-$400 is nothing.

The prices of 4nm+ fabs have been hiked repeatedly

Before latest tariffs 7nm/6nm was being discounted but not sure now. The future does indeed look bleak for budget GPUs

 

[Mopetar]

There were already announcements that tariffs would not be applied to GPUs. Just as the mining boom was winding down, the AI boom took its place. Neither Nvidia nor AMD make the best money from the gaming market, so don't expect them to cater to it, least of all the bottom of it.

 

[basix]

I'm on board with the idea that 8GB cards are inadequate for many games. I do not agree that cutting down 128bit GPUs to 96bits is a good way to solve it.

I absolutely think that 96bit with 12GByte would be the better card for consumers. We need to work with what we have. And for GDDR6 that are 2 GByte/Package modules and max. 128bit bus width. The reduced SI width can be compensated by cut down CUs and clock rates. And it is better for a consumer to leave 5% performance on the table because of less bandwidth, compared to getting all the problems with 8GByte. 8GByte performance can be OK at 1080p. But depending on the game you get stutters and reduced texture quality. At 1440p its much worse and if you want to use RT, 8GB get smashed even at 1080p.

So yes, I would accept a narrower memory bus and losing a few percent of performance (low single digit) compared to the 8GB problems. You also need to consider, that you buy the card today and you want it to use it for 2-4 years. Good luck with 8GB. And many people forget, that the reselling value of an 8GB card will be much lower compared to a 12GB and 16GB variant. The latter ones might still be usable in 3 years but the 8GB variant definitely not.

Here an exemplary portfolio which I would strongly recommend to AMD:
- Do not release a 9060 (XT) with only 8GB
- 8GB for the 9050 XT is OK

SKU	CU	SI	VRAM	GPU Clock	Memory Clock	TBP	MSRP
9070 XT	64	256bit	16 GByte	2.9 GHz	20 Gbps	304W	599$
9070	56	256bit	16 GByte	2.5 GHz	20 Gbps	220W	549$
9070 GRE	48	192bit	12 GByte	2.7 GHz	20 Gbps	200W	449$
9060 XT	32	128bit	16 GByte	3.2 GHz	20 Gbps	180W	349$
9060	28	96bit	12 GByte	2.8 GHz	20 Gbps	140W	289...299$
9050 XT	24...28	128bit	8 GByte	2.5 GHz	16...18 Gbps	110W	239...249$
7600	32	128bit	8 GByte	2.6 GHz	18 Gbps	165W	199$

 

[TESKATLIPOKA]

The $300 5060 is going to get destroyed in the press as well. I also see no scenario where AMD charge less than $250 for an 8GB part and the whole argument is that at that price 96bit + 12GB would be better overall.

RTX 5060 8GB is priced 50% higher than what you think is a fair price for N44 8GB.

Because it will be B580 performance and that was well received until the driver overhead issue and the lack of stock became apparent.

B580 was praised because It performed better than A770 and had the best perf/$ at every resolution.

You previously compared it to the 7600XT by saying a 96bit part would have less bandwidth. Using 20Gbps Vram would reduce that deficit since the 7600XT has 18Gbps modules.

I did use 20Gbit modules, yet N33 still has 20% higher BW.
7060(XT) 8(16)GB 128-bit 18Gbit -> 288GB/s
9050XT 12GB 96-bit 20Gbit -> 240GB/s
Without It It would be even worse and let's not forget about the smaller IC(32MB vs 24MB).

A few cases? There are dozens and there will be many more by the end of the year. Also the reality is that there is zero chance that an 8GB part would be sold for $200 even though that is the only price point that would make sense given the shortcomings of 8GB or VRAM.

Are those "dozens" of cases actually >30FPS or just some of them? Never mind, It's not important.

If somehow AMD did decide to make a $200 8GB 9060 then sure thing don't bother with a 12GB part. The point is though that the 8GB part if released is going to be $250 to $300 and at that price point being obsolete in the latest AAA titles and having insufficient Vram to use features like RT is not acceptable. That is something a 12GB variant would not suffer from (at 1080p) and despite not having the same peak performance as an 8GB part would it won't have the troughs either and it will be a more consistent experience for the user.

Even that 12GB card would be pretty obsolete in current games let alone the future ones. Black Myth Wukong is not playable even at 1080p without RT, 6700XT managed only 27.3FPS, but the funny thing is the full N44 128-bit 8GB should manage ~35FPS.
If both of them cost the same then I can't recommend either of them over the other unless you are ok with upscaling.
Of course this is only "If both would be released", but we pretty much know AMD will release 128-bit 8GB.

That is what I am proposing....

The 8GB part is DOA unless it costs $200, which it won't. With that in mind the better alternative is 12GB 96bit at the $280ish price point the 8GB part is actually going to be priced at (assuming it is not cancelled)

Edit. And the best alternative would be a 16GB 9060 with slower ram, 28CUs, and lower clocks in that $280-$300 bracket.

There is no info It will cost $280. We don't know the price yet.
And this debate is honestly pointless, because AMD very likely won't release a 96-bit 12GB card, they could but won't.
They will release either a full or cutdown 128-bit 8GB card for >$229.

 

[Timorous]

RTX 5060 8GB is priced 50% higher than what you think is a fair price for N44 8GB.

B580 was praised because It performed better than A770 and had the best perf/$ at every resolution.

I did use 20Gbit modules, yet N33 still has 20% higher BW.
7060(XT) 8(16)GB 128-bit 18Gbit -> 288GB/s
9050XT 12GB 96-bit 20Gbit -> 240GB/s
Without It It would be even worse and let's not forget about the smaller IC(32MB vs 24MB).

Are those "dozens" of cases actually >30FPS or just some of them? Never mind, It's not important.

Even that 12GB card would be pretty obsolete in current games let alone the future ones. Black Myth Wukong is not playable even at 1080p without RT, 6700XT managed only 27.3FPS, but the funny thing is the full N44 128-bit 8GB should manage ~35FPS.
If both of them cost the same then I can't recommend either of them over the other unless you are ok with upscaling.
Of course this is only "If both would be released", but we pretty much know AMD will release 128-bit 8GB.

There is no info It will cost $280. We don't know the price yet.
And this debate is honestly pointless, because AMD very likely won't release a 96-bit 12GB card, they could but won't.
They will release either a full or cutdown 128-bit 8GB card for >$229.

A lot of blablabla to miss the point again.

The 9070XT has less cache and bandwidth than the 7900XTX yet competes quite well with it. Zero reason a 96bit 12GB N44 part won't be faster than the 7600XT despite the far more modest bandwidth deficit. Lots of reasons to think otherwise and 10% more than the 7600XT is B580/6700XT territory.

9060XT 16GB is likely in the $350 region if it performs close to the 5060Ti 16GB which I think it will. It will be akin to the 9070XT Vs 5070Ti

An 8GB variant is going to cost between $270 and $300 and at that price point it is going to be DOA just like the 5060 will be.

If AMD want a non DOA $300 part the best option is a 16GB 28CU 9060. If AMD don't want 2 parts so close together the next best choice is a 12GB part around the $250 price point.

The other option is AMD do a $350 9060XT 16GB and just repurpose the 7600XT as the $250 part. I think at $250 a 12GB N44 based part would be faster than the 7600XT though.

 

[Tuna-Fish]

Why would you imagine that a 12GB model can be priced lower than the 8GB one?

 

[Kepler_L2]

Why would you imagine that a 12GB model can be priced lower than the 8GB one?

The currently planned 9060 XT 8GB has the full die, a 12 GB model can absord defects in the IC/memory controller.

 

[marees]

N48 12gb launched in China ?

the AMD Radeon RX 9070 GRE graphics card is now online on the official website, and all specifications have been officially confirmed.

The e-commerce platform currently has a Radeon RX 9070 GRE graphics card on the shelves, which shows that the actual sales (payment balance) time is 9:30 am on May 8; the lowest-priced model is priced at 4,199 yuan.

AMD RX 9070 GRE 显卡上线官网：48CU + 12GB GDDR6，5 月 8 日发售 - IT之家

RX 9070 GRE 的整体规格大致相当于完整 Navi 48 的 3/4，TBP 与 RX 9070 同为 220W，现有型号中价格最低的售 4199 元。

www.ithome.com

https://twitter.com/x/status/1916699378531856852

 

[SolidQ]

the AMD Radeon RX 9070 GRE g

fps from AMD

 

[Timorous]

fps from AMD

Not bad for less cache, less mem BW and far fewer compute units than the 7900GRE. It is almost as though RDNA4 is a much better architecture than RDNA3 so who would have thought a cut part could perform very well?

Makes me think a 96bit, 12GB, 28CU 9060/9050XT should have been on the roadmap instead of having an 8GB 9060XT variant. Would easily be 6700XT performance and priced in the $250 to $280 window it would be well received and in a head to head vs the 5060 it would be far superior for those who can't stretch to whatever the 16GB 9060XT is going to cost.

 

[marees]

Makes me think a 96bit, 12GB, 28CU 9060/9050XT should have been on the roadmap instead of having an 8GB 9060XT variant. Would easily be 6700XT performance and priced in the $250 to $280 window it would be well received and in a head to head vs the 5060 it would be far superior for those who can't stretch to whatever the 16GB 9060XT is going to cost.

Yeah, if RDNA 4 was this good then they should have pivoted to a 12gb 9050xt. I understand initial plans would have been n41, n42, & n43 but a smart product manager should have pivoted to 12gb for a cut down n44 (only wildcard here is the cut down infinity cache)

 

[sl0519]

fps from AMD

Fewer CUs, less bandwidth, but still the same 220W TDP? Efficiency doesn't look too good to me. Also, how did they manage to cut the bus width from 256-bit to 192-bit while still using the same Navi 48 die?
Does that mean it's also possible for Nvidia to just use defective GB203 dies and make a 12G 192-bit RTX 5070 too?"

 

[adroc_thurston]

Efficiency doesn't look too good to me

Efficiency is not the design point for anything RDNA4.

Also, how did they manage to cut the bus width from 256-bit to 192-bit while still using the same Navi 48 die?

You can just like disable IMC segments you know.
Tonga is like the most notable example of that in AMD lands.
The full die was 384b but what shipped outside of Apple lands was 256b. Too bad!

 

[sl0519]

Efficiency is not the design point for anything RDNA4.

You can just like disable IMC segments you know.

Does that mean it's also possible for Nvidia to just use defective GB203 dies and make a 12G 192-bit RTX 5070 too? Or by same logic, they could've made a 192-bit 5060 on GB206 but they've chosen not to?

 

[adroc_thurston]

Does that mean it's also possible for Nvidia to just use defective GB203 dies and make a 12G 192-bit RTX 5070 too?

Yeah lol they do it all the time later in the product lifecycle.
You won't believe the places you can find AD103 in.
Some NV SKUs become mysterymeat in truest sense of that word.

Or by same logic, they could've made a 192-bit 5060 on GB206 but they've chosen not to?

GB206 is a 128b die so no.

 

[jpiniero]

Not bad for less cache, less mem BW and far fewer compute units than the 7900GRE. It is almost as though RDNA4 is a much better architecture than RDNA3 so who would have thought a cut part could perform very well?

Makes me think a 96bit, 12GB, 28CU 9060/9050XT should have been on the roadmap instead of having an 8GB 9060XT variant. Would easily be 6700XT performance and priced in the $250 to $280 window it would be well received and in a head to head vs the 5060 it would be far superior for those who can't stretch to whatever the 16GB 9060XT is going to cost.

Again, it's not enough of a BOM cost reduction to make sense. If the 8 GB won't sell well, just not ship much.

 

[SolidQ]

Interesting how much price cut will be after NV launch RTX 5070 Super 18GB.

 

[Timorous]

Does that mean it's also possible for Nvidia to just use defective GB203 dies and make a 12G 192-bit RTX 5070 too? Or by same logic, they could've made a 192-bit 5060 on GB206 but they've chosen not to?

As Adroc says GB206 is 128bit so NV could have just done a 16GB 5060Ti with the 128 bit bus and a 96 bit 12GB 5060 which probably would have been fine at $330 and NV would be in a better position. Even more so because compared to the 4060 96bit GDDR7 is still more bandwidth.

 

[Timorous]

Again, it's not enough of a BOM cost reduction to make sense. If the 8 GB won't sell well, just not ship much.

The 8GB parts are a total waste of resources and there will be N44 dies that don't have all the MCs working or have a faulty bit of cache or have a defective CU or two so a 28CU 96bit 24MB part allows AMD to use some of their salvage dies. It also allows them to use dies that might be fully functioning but don't hit the 9060 XT clocks.

Right now their plan is a 16GB full fat card and an 8GB full fat card. The BOM cost advantage of being able to use salvage dies is far more beneficial than saving a few $ on 2 ram chips. Those parts would also sell well at the sort of price AMD is going to want to charge for an 8GB 9060XT so you know, selling stuff > stuff sitting on the shelf.

 

[jpiniero]

The 8GB parts are a total waste of resources and there will be N44 dies that don't have all the MCs working or have a faulty bit of cache or have a defective CU or two so a 28CU 96bit 24MB part allows AMD to use some of their salvage dies. It also allows them to use dies that might be fully functioning but don't hit the 9060 XT clocks.

People will buy the 8 GB model just because it's cheaper. Some at least. Just don't ship much.

As to what to do with the small % that has busted IMC dies, well... Save it for later.

 

[Timorous]

People will buy the 8 GB model just because it's cheaper. Some at least. Just don't ship much.

As to what to do with the small % that has busted IMC dies, well... Save it for later.

I would rather companies did not ship out obsolete (for the price) at launch parts when there are viable alternatives that give the same kind of margin but offer a much better customer experience.

The reception to a $280 8GB 9060XT will be a lot lot lot worse than a $280 12GB 9060 that performs like B580 / 6700XT. That is just a fact. The BOM cost difference is minimal at best but 1 part will sell well and make AMD look good and the other part will be laughed, ridiculed and make AMD look as bad as NV. Long term even if the 12GB part had 1/2% lower margins than the 8GB part it would more than pay for itself in goodwill.

 

[jpiniero]

I would rather companies did not ship out obsolete (for the price) at launch parts when there are viable alternatives that give the same kind of margin but offer a much better customer experience.

The reception to a $280 8GB 9060XT will be a lot lot lot worse than a $280 12GB 9060 that performs like B580 / 6700XT. That is just a fact. The BOM cost difference is minimal at best but 1 part will sell well and make AMD look good and the other part will be laughed, ridiculed and make AMD look as bad as NV. Long term even if the 12GB part had 1/2% lower margins than the 8GB part it would more than pay for itself in goodwill.

If AMD was worried about goodwill (or reviewers hating on it), they wouldn't ship the 8 GB at all.

 

[Timorous]

If AMD was worried about goodwill (or reviewers hating on it), they wouldn't ship the 8 GB at all.

I hope they don't. I hope the rumour they are cancelling it are true.

If they were really smart they would have known this a while ago and would have had an alternative plan for the lower part.

As it is they can pivot and just release it a bit later. A $350 9060XT 16GB basically makes the 5060 range pointless and the 7600XT can fill in below for the time being.