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

 

[Arctic Islands]

Interesting why NV making 5070ti faster release, isn't that 8800XT faster than RTX 5070?

It depends on how fast the 5070 is. Some Reliable source (not MLID) also pointed out that 8800XT will match 4080 super in rasterization while being very close in RT.

 

[jpiniero]

It depends on how fast the 5070 is. Some Reliable source (not MLID) also pointed out that 8800XT will match 4080 super in rasterization while being very close in RT.

Totally unrealistic. That would be over 40% faster in raster than the 7800 XT with basically no CUs or bandwidth upgrades. If it had GDDR7, maybe it could if it was also hitting high 3's in clocks, closer to 4 Ghz.

 

[Tuna-Fish]

Totally unrealistic. That would be over 40% faster in raster than the 7800 XT with basically no CUs or bandwidth upgrades. If it had GDDR7, maybe it could if it was also hitting high 3's in clocks, closer to 4 Ghz.

+6.7% CU. And I don't think the +30% clocks needed to hit ~+40% raster is that much of a tall order. But I agree that the complete flatlining of the dram interface would make that hard.

 

[Win2012R2]

7800 XT is listed at TPup as having boost at 2430 Mhz, 30% up on that is 3159 Mhz - seems doable with fixed RDNA3 issues and improved process, no?

Memory bandwidth is a problem though unless they improved compression and/or caching.

Typical AMD cheapening out on GDDR7

 

[GTracing]

Totally unrealistic. That would be over 40% faster in raster than the 7800 XT with basically no CUs or bandwidth upgrades. If it had GDDR7, maybe it could if it was also hitting high 3's in clocks, closer to 4 Ghz.

The 5700 XT has 8% lower performance than Radeon VII with 2/3rds the CUs. That's 39% higher performance per CU.

The 6700 XT had 31% higher performance than the 5700 XT and they have the same number of CUs. The 6700 XT also only has 0.86x the memory bandwidth. That's 50% higher performance per GB/s of memory bandwidth.

40% higher performance per CU seems within the realm of possibility for me.

ETA: I used TechPowerUp's 1440p performance numbers from their 5700 XT and 6700 XT reviews.

 

[igor_kavinski]

Typical AMD cheapening out on GDDR7

Not their fault. The Geforce crowd is unwilling to give them a chance, always coming up with some excuse.

 

[Win2012R2]

Not their fault. The Geforce crowd is unwilling to give them a chance, always coming up with some excuse.

RDNA2 did very well because it was a good product, despite being total crap at ray tracing, people loved it.
RDNA3 was buggy so turned to be a crock - they should have priced it lower to start with.

It's pathetic they've given up (for 4th gen at least) chiplets, but I get the commercial angle - they'll make more from selling MI300, like A LOT MORE, however they'd better come up with good killer RDNA5 top card in 2025 - I'd buy it IF it beats NVIDIA, proper beats top end card by at least 25-30%.

 

[SolidQ]

however they'd better come up with good killer RDNA5 top card in 2025

2026

Totally unrealistic.

Different arch and clocks/optimizations

 

[igor_kavinski]

RDNA3 was buggy so turned to be a crock - they should have priced it lower to start with.

Who knows how many bugs other release silicon has. AMD shouldn't have let it leak that it had bugs because a lot of people (including myself) kept waiting for fixed silicon.

 

[Win2012R2]

Who knows how many bugs other release silicon has

Every silicon got bugs, but RDNA3's were bad enough to hit the main perf goal.

Also weakness in ray tracing becomes as untenable as 8 GB cards - AMD can't allow Intel being better in that area!

AMD shouldn't have let it leak that it had bugs because a lot of people (including myself) kept waiting for fixed silicon

And that's why they leaked it...

 

[adroc_thurston]

AMD can't allow Intel being better in that area!

Intel isn't relevant at all.

 

[marees]

N48 incoming

No news yet of (8gb) N44

Version History​

v2.61.0 (December 16th, 2024)​

• Added support for Intel Arc B570 & B580 (Battlemage), Core Ultra 200 iGPU (Arrow Lake)
• Added support for AMD Navi 48, Ryzen 9800X3D iGPU
• Added support for NVIDIA RTX 2080 Ti ES, H100 80GB HBM3, A4000H, A800 40 GB Active, RTX 5880 Ada, Tesla K40st
• Added support for Qualcomm Adreno 540, 630, 640, 642L
• Fixed crash on some AMD Ryzen systems with older drivers when dGPU installed and iGPU device disabled
• Added PCI vendors Shangke and ONIX

TechPowerUp GPU-Z v2.61.0 Download

GPU-Z is a lightweight utility designed to give you all information about your video card and GPU.

www.techpowerup.com

Loading…

videocardz.com

 

[RnR_au]

however they'd better come up with good killer RDNA5 top card in 2025 - I'd buy it IF it beats NVIDIA, proper beats top end card by at least 25-30%.

This is ofcause in bang-for-buck right? Not just fps with RT enabled in an nvidia sponsored title?

Right? Right?!?!?

 

[SolidQ]

Also weakness in ray tracing becomes as untenable as 8 GB cards - AMD can't allow Intel being better in that area!

What if AMD will surprise us, and there only 16GB N44 version

 

[Kepler_L2]

What if AMD will surprise us, and there only 16GB N44 version

No chance, they even have a 96-bit SKU planned.

 

[jpiniero]

No chance, they even have a 96-bit SKU planned.

Probably mobile? nVidia is keeping the 4050 6 GB Laptop on the market so a N44 that was 96 bit would compete with that.

 

[Win2012R2]

What if AMD will surprise us, and there only 16GB N44 version

Nah, zero change - they'd better have 16 GB option launched with max +50 bucks for it.

 

[Win2012R2]

This is ofcause in bang-for-buck right? Not just fps with RT enabled in an nvidia sponsored title?

AMD's flagship must be flat out better - I am less bothered about not having DLSS because proper flagship should be able to provide very good (100+) FPS in 4k with everything maxed out, yes it will cost $1500-2000-2500, but there is a fairly big market for top end gaming GPUs.

 

[Arctic Islands]

Totally unrealistic. That would be over 40% faster in raster than the 7800 XT with basically no CUs or bandwidth upgrades. If it had GDDR7, maybe it could if it was also hitting high 3's in clocks, closer to 4 Ghz.

The same source also pointed out that Navi 48 has a GCD and MCDs. The total die size is over 400 mm^2. Trust it or not.

 

[beginner99]

The same source also pointed out that Navi 48 has a GCD and MCDs. The total die size is over 400 mm^2. Trust it or not.

Yeah as previously already stated, either the original rumors of relatively small monolithic die are true and it will probably be less than 7900 xt in raster or these rumors were wrong on indeed the small die was only the GCD. But small monolith at 7900 xtx makes zero sense. chiplet design is believable. they could reuse the exact same MCD from RDNA3, eg no r&d costs for that part and maybe even easier to validate the chip?

 

[Kepler_L2]

they could reuse the exact same MCD from RDNA3

It's a new version of MALL so they can't.

 

[inquiss]

It's a new version of MALL so they can't.

New in what way?

 

[Kepler_L2]

New in what way?

 

[SolidQ]

Level 0 – Legacy Solutions
Level 1 – Software on Traditional GPUs
Level 2 – Ray/Box and Ray/Tri Testers in Hardware
Level 3 – Bounding Volume Hierarchy (BVH) Processing in Hardware
Level 4 – BVH Processing with Coherency Sorting in Hardware
Level 5 – Coherent BVH Processing with Scene Hierarchy Generator in Hardware

Interesting what level is RDNA4 lvl3?

 

[beginner99]

It's a new version of MALL so they can't.

I'm pretty clueless about this low level stuff. So my question might sound stupid. i would have assume the mcd is some controller + memory and the controller runs a firmware so that update could happen on that side only?

else with a new MCD design needed just for N48, doesn't make a whole lot of sense to me.