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

 

[DownTheSky]

Those are GPU scores. You see in the picture 9070XT gets 30xxx and 9070 26xxx

 

[Tup3x]

My bad, i based my claim on the Guru3d link added below. Probably they use different settings as there XTX gets 29000 and XT gets 26000.

But great, then the perf is right where expected!

It's IMO higher than expected. Not sure what score you are looking at but it's the graphics score that you should be using when making comparisons. That's what Guru3D uses.

 

[Tuna-Fish]

Cut down N44 to 28CU, 96bit and reduce TGP aka lower the GPU clock rates. There you have your 9060 12GB. If using 20Gbps instead of 18Gbps GDDR6, you could even compensate for the narrower bus width.

But even if the chip does not use faster GDDR6 and loses 5-10% performance, 12GB would be the better deal for most people.

It would lose more performance than that. Cutting a quarter of the memory bus also cuts a quarter of the IC (it is a memory-side cache, directly attached to memory controllers), and that's probably already low enough that any cuts are really painful. I think the loss would be more like 20-25%.

 

[adroc_thurston]

It's IMO higher than expected.

what were the expectations anyway

 

[gaav87]

what were the expectations anyway

That is the 304w card
I know 340W goes brrrrrrrrrr

 

[Timorous]

It would lose more performance than that. Cutting a quarter of the memory bus also cuts a quarter of the IC (it is a memory-side cache, directly attached to memory controllers), and that's probably already low enough that any cuts are really painful. I think the loss would be more like 20-25%.

And when you run out of Vram the performance loss can be 90%.

A cheap card having compromises is fine, but lowering a few bandwidth heavy effects is by far the preferred option than lowering texture quality.

With more and more games baking in RT effects as well some games simply won't run with 8GB of VRAM without massive IQ sacrifices.

 

[Tup3x]

what were the expectations anyway

That XT would be somewhere between XT and XTX, in rasterisation at least.

 

[coercitiv]

And when you run out of Vram the performance loss can be 90%.

The point was that cutting the the bus to 96 bit incurs too much of a performance penalty, not that 8GB is fine. If you're aiming to make a 12GB card, might as well go the extra mile to 16GB and keep bandwidth and cache intact.

 

[adroc_thurston]

That XT would be somewhere between XT and XTX, in rasterisation at least.

Ah. Well new microarchitectures are good.

The point was that cutting the the bus to 96 bit incurs too much of a performance penalty, not that 8GB is fine. If you're aiming to make a 12GB card, might as well go the extra mile to 16GB and keep bandwidth and cache intact.

Bingo.
12GB config@128b only makes sense with 24Gbit GDDR7.

 

[Gideon]

It will depend on the game. TPU, for example, recently changed their power consumption methodology. Coincidence, possibly.
But if they can't reduce power consumption in typical games it may explain why it doesn't appear in laptops.

I'm stil quite certain we will see some form of RDNA4 at the very least in the 16" Framework. They specifically mentioned that they have long plans with the platform.

 

[Ghostsonplanets]

8GB indeed would be a hard sell at >$200. Specially when upcoming games such as Doom, AC Shadows and others already ask for 8GB VRAM as the minimum specification. It's not a future-proof configuration at all.

Could AMD do clamshell like Xbox did with Series S? i.e Clamshell only on one memory channel, allowing for 10GB on a 128-bit bus. Granted, this solution has a heavy penalty on bandwidth.

 

[adroc_thurston]

I'm stil quite certain we will see some form of RDNA4 at the very least in the 16" Framework. They specifically mentioned that they have long plans with the platform.

They should stop doing dumb stuff and stick stxH there instead.

 

[Timorous]

The point was that cutting the the bus to 96 bit incurs too much of a performance penalty, not that 8GB is fine. If you're aiming to make a 12GB card, might as well go the extra mile to 16GB and keep bandwidth and cache intact.

That is what the 9060XT should be, and if yields are great then maybe don't even bother with a cut part.

If there is going to be a cut part though then it makes far more sense for it to be 96bit + 12GB and 28CUs or whatever CU count makes sense than to make an 8GB version simply because the drawbacks of 8GB are vastly greater than the drawbacks of being a bit bandwidth bound over the next 5 years.

I also do not think that 240GB/s of bandwidth (20gbps ram) with 24MB IC is insufficient for the price tier and performance tier it would land in ($250 or less). Performance should be in the region of 6700/6700XT even with the constraints and for an entry level 1080p card that seems perfectly fine to me. On top of that RDNA 4 is supposed to be a lot more bandwidth efficient than RDNA 3 so why would going from 288GB/s to 240GB/s make such a huge difference at the low end, it won't

The other factor is that it will be an 8x PCIe bus and we know that when you spill over to system ram PCIe bandwidth starts to matter more, especially with PCIe 3 on an older board which is a more likely pairing with an entry level card. (Do AMD want a repeat of the B580 retests when that was tested with more likely CPU pairings and performance tanked, because they will get people testing a 9060 8GB with a B350 or B450 motherboard and you will see the 8GB buffer hurt it even more.)

EDIT: Another factor to think about is that the 9060 should be mildly capable of handling the games with baked in RT in a way 7000 series parts are not, RT has a VRAM footprint that will make 8GB look even worse. Just go look at Indiana Jones for example in the 8GB thread, more games will be released in the next 5 years like that and 8GB parts will suffer horribly where as a slightly bandwidth bound 12GB part is far more likely to be okay).

 

[GTracing]

That is what the 9060XT should be, and if yields are great then maybe don't even bother with a cut part.

If there is going to be a cut part though then it makes far more sense for it to be 96bit + 12GB and 28CUs or whatever CU count makes sense than to make an 8GB version simply because the drawbacks of 8GB are vastly greater than the drawbacks of being a bit bandwidth bound over the next 5 years.

I also do not think that 240GB/s of bandwidth (20gbps ram) with 24MB IC is insufficient for the price tier and performance tier it would land in ($250 or less). Performance should be in the region of 6700/6700XT even with the constraints and for an entry level 1080p card that seems perfectly fine to me. On top of that RDNA 4 is supposed to be a lot more bandwidth efficient than RDNA 3 so why would going from 288GB/s to 240GB/s make such a huge difference at the low end, it won't

The other factor is that it will be an 8x PCIe bus and we know that when you spill over to system ram PCIe bandwidth starts to matter more, especially with PCIe 3 on an older board which is a more likely pairing with an entry level card. (Do AMD want a repeat of the B580 retests when that was tested with more likely CPU pairings and performance tanked, because they will get people testing a 9060 8GB with a B350 or B450 motherboard and you will see the 8GB buffer hurt it even more.)

EDIT: Another factor to think about is that the 9060 should be mildly capable of handling the games with baked in RT in a way 7000 series parts are not, RT has a VRAM footprint that will make 8GB look even worse. Just go look at Indiana Jones for example in the 8GB thread, more games will be released in the next 5 years like that and 8GB parts will suffer horribly where as a slightly bandwidth bound 12GB part is far more likely to be okay).

First off, the B580 performance tanks when paired with older CPUs because it has super high CPU overhead. It has nothing to do with VRAM limitations.

For the few games that have higher VRAM requirements, you can turn down settings. And I'm not saying that 8GB is fine, but a 96bit memory bus feels like a "the cure is worse than the disease" situation.

 

[SolidQ]

Sapphire and Asrock RX 9070 series

 

[jpiniero]

8GB indeed would be a hard sell at >$200. Specially when upcoming games such as Doom, AC Shadows and others already ask for 8GB VRAM as the minimum specification. It's not a future-proof configuration at all.

Which is fine since the price range you'd be talking about is sub-entry level.

 

[amenx]

NDA lifts today?

 

[Timorous]

First off, the B580 performance tanks when paired with older CPUs because it has super high CPU overhead. It has nothing to do with VRAM limitations.

For the few games that have higher VRAM requirements, you can turn down settings. And I'm not saying that 8GB is fine, but a 96bit memory bus feels like a "the cure is worse than the disease" situation.

I know why B580 performance tanks, read what I said. B580 was tested with the typical test system to remove CPU bottlenecks but then it was re-tested with CPUs that were far more likely to be paired with such a cheap GPU and performance tanked which as you say is due to the overhead.

A cheap 9060 GPU is far more likely to be paired with an existing Zen 3 system given the lack of decent budget platform options when you factor in CPU, motherboard and RAM costs. As such if you give it just 8GB of VRAM you will have the double whammy of insufficient VRAM and insufficient PCIe bandwidth to even somewhat alleviate that flaw.

I can guarantee if there were 2 9060 configs one was 8GB 128 bit and 28CU and the other was 12GB 96bit and 28CU both at the same clocks the 12GB card will be substantially better in enough titles to make the cases where games run better on the 8GB version due to bandwidth moot, and that is today, give it 5 years and the 8GB card will be unusable where as the 12GB card will not suffer the same fate quite as much. You can see cases now where the 3060 12GB is ahead of many 8GB cards despite having a massive compute and bandwidth deficit.

The other big advantage of doing the 96bit / 12GB variant is that you can use more trash dies because you can use parts where there is an issue in a block of L3 Cache, you can use dies with duff CUs and you can use dies with duff memory controllers so you can have slightly higher supply without needing to castrate as many perfectly working dies.

For AMD the other thing is market acceptance. The market would accept a $250 card with 12GB of VRAM, even on a 96bit bus, if performance was 6700/6700XT level. The market will not accept an 8GB card at that price point even if performance is better in some titles because there will be other cases where performance or IQ is absolute trash.

 

[ToTTenTranz]

Nah it was even more of a paper launch than Blackwell

It looks like the B580 @ $250 was more of a marketing op than anything.

They needed to tell the world they can have a decent offer at a decent price so they shouldn't be dismissed yet. And they got glowing reviews for it.


Though I guess if N5 wafers get less expensive, they could still ramp up production at some point.

 

[exquisitechar]

NDA lifts today?

Yes, probably very soon.

 

[DAPUNISHER]

Bryan slipped a score in?

 

[jpiniero]

For AMD the other thing is market acceptance. The market would accept a $250 card with 12GB of VRAM, even on a 96bit bus, if performance was 6700/6700XT level. The market will not accept an 8GB card at that price point even if performance is better in some titles because there will be other cases where performance or IQ is absolute trash.

Case in point - I suspect the reason AMD did two 16 GB 9070 models is because of the memory bandwidth.

But really, you are talking about sub-entry level.

 

[Ghostsonplanets]

Which is fine since the price range you'd be talking about is sub-entry level.

If they do a 8GB SKU at $199 or below, I agree. But >$200 it needs 12GB minimum in my opinion.

Bryan slipped a score in?

Extremely strong performance. Nearly matching a 96 CU GPU with only 2/3 of the silicon. RDNA 4 is really a generational leap.

 

[jpiniero]

If they do a 8GB SKU at $199 or below, I agree. But >$200 it needs 12GB minimum in my opinion.

I assume the MSRP will be higher than the 7600 / XT was (which was $269 for the 8 and $329 for 16).

 

[Timorous]

If they do a 8GB SKU at $199 or below, I agree. But >$200 it needs 12GB minimum in my opinion.

I doubt they will do an 8GB SKU at $199 though. The 16GB SKU if it performs around the 7700XT / 6800 / 4070 would make a great card at $330 and it would be a direct replacement for the 7600XT.

The next step down below $330 is around $250 and as you say at that price point 12GB is a requirement.

Given the size of N44 I don't see yields being so bad that there is room for more than 2 SKUs.

So yea. If I was AMD 12GB 96bit @ $250 and it should roughly match the B580 which was well received and not have the driver overhead issue. That would be a winner.

An 8GB 128bit card at that price point though will be absolutely destroyed because Aussie Steve will throw out an Indiana Jones at 1080p benchmark where the 12GB B580 is playable and the 8GB 9060 either does not load the game at all or is in the single digits.