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

 

[soresu]

For higher resolutions the 6800 has a big cache and bandwidth advantage though

This is my thinking on RT, that even with the improvements to memory management and compression in RDNA4 that N48's RT performance may be handicapped by cache.

Especially as AMD was eerily silent about it during the presentation yesterday.

I don't recall anyone mentioning infinity cache at all.

 

[basix]

Does it need to be a different model? Sure, doubling memory requirements (space, bandwidth) and only a quarter of the TFLOPS per CU will hurt. But I did profile DLSS2 back then (still CNN) and the execution window on the tensor cores was very narrow. Maybe 0.25ms with ~50% of peak utilization on a 4090 (4K output). A 7900XTX deilvers roughly 35% of a 4090's Matrix FP16 TFLOPS with roughly the same raw bandwidth. If AMD can get to a litte bit better utilization rate, a DLSS2 style CNN can be executed in maybe 0.5ms, which ist not very much.

 

[Josh128]

The 7900 XT has a 315W TDP

TBP, not TDP according to AMD site, while 9070XT has 304W TBP according to the same site. According to TPU, 7900XT is 19% faster on average than 7900GRE Sapphire Pulse in both 4K raster, and 4K RT. AMD claims +42% vs 7900GRE across a mix of RT and non RT at 4K. Lets assume the typical first party benchmarks BS and use the obvious knowledge that RT has improved more and thus has heavier weight in AMDs comparison than raster. Chop off 10% overall and you have +32% "no BS" average uplift vs 7900GRE in mixed 4K usage, vs 7900XT's +19% from TPU benches. Compared to each other, that makes 9070XT ~11% faster than 7900XT, while using around 4% less power. Keep in mind this should be a pretty realistic comparison as I deducted the 10% "AMD tax" from AMDs perf claims.

How is Charlie seriously claiming worse efficiency? What numbers is he using?

 

[Novacius]

This is my thinking on RT, that even with the improvements to memory management and compression in RDNA4 that N48's RT performance may be handicapped by cache.

Especially as AMD was eerily silent about it during the presentation yesterday.

I don't recall anyone mentioning infinity cache at all.

AMD clearly showed bigger gains with RT than without, so cache doesn't seem to be a huge hindrance. 7800 XT with 64MB was also a bit faster in RT than 6800 XT with 128MB.

 

[DisEnchantment]

This patent below could help you with the box node compression

Bounding volume hierarchy box node compression - Advanced Micro Devices, Inc.

<div p-id="p-0001">A technique for performing ray tracing operations is provided. The technique includes combining one or more common exponent values of a compressed box node with one or more minimum

www.freepatentsonline.com

View attachment 95606

This patent below contains handling both BVH4 and BVH8.

VARIABLE WIDTH BOUNDING VOLUME HIERARCHY NODES - Advanced Micro Devices, Inc.

<div p-id="p-0001">A technique for performing ray tracing operations is provided. The technique includes processing small bounding box nodes in a box intersection test circuit to generate intersection

www.freepatentsonline.com

View attachment 95607

I think for prebuilt BVH trees performance gain won't be much. For runtime generated BVH trees it could be a lot of boost.
Cache and memory subsystem will be key.
For BVH4 it fetches two cache lines, for BVH8, it fetches single cache line.



https://twitter.com/NIV_Anteru , who also presented the amazing workgraphs demo recently is inventor of these patents below for generating new kind of BVH trees





Will only help runtime generated trees it looks like to me




Updated way to perform intersection tests too with rotated boxes


I doubt PS5 will get all these with the weak CPU (if the BVH generation is not done on GPU)

Here’s a closer look at the RT improvements. Looks like it supports BVH compression and shader execution reordering?
View attachment 118359
View attachment 118360
View attachment 118361
View attachment 118362
View attachment 118363

Wow a lot of patents/applications made it to the product, pleasantly surprised. The new CU is indeed very potent.

Looking back at the patents...

Oriented Bounding Box was indeed a big part of the new RT arch

BOUNDING VOLUME HIERARCHY HAVING ORIENTED BOUNDING BOXES WITH QUANTIZED ROTATIONS
From <https://www.freepatentsonline.com/y2023/0099806.html>
TECHNIQUES FOR INTRODUCING ORIENTED BOUNDING BOXES INTO BOUNDING VOLUME HIERARCHY
From <https://www.freepatentsonline.com/y2023/0027725.html>
COMMON CIRCUITRY FOR TRIANGLE INTERSECTION AND INSTANCE TRANSFORMATION FOR RAY TRACING
From <https://www.freepatentsonline.com/y2023/0206541.html>
VOLUME INTERSECTION USING ROTATED BOUNDING VOLUMES
https://www.freepatentsonline.com/y2023/0410426.html
TECHNIQUE FOR TESTING RAY FOR INTERSECTION WITH ORIENTED BOUNDING BOXES
From <https://www.freepatentsonline.com/y2024/0221284.html>
EMULATING ORIENTED BOUNDING BOXES IN BOUNDING VOLUME HIERARCHIES
From <https://www.freepatentsonline.com/y2024/0221283.html>

BVH Compression

• BOUNDING VOLUME HIERARCHY BOX NODE COMPRESSION
  
  • From <https://www.freepatentsonline.com/y2023/0206540.html>
• BOUNDING VOLUME HIERARCHY LEAF NODE COMPRESSION
  • From <https://www.freepatentsonline.com/y2024/0203032.html>

BVH8

VARIABLE WIDTH BOUNDING VOLUME HIERARCHY NODES
From <https://www.freepatentsonline.com/y2023/0206542.html>

Instance Transform

COMMON CIRCUITRY FOR TRIANGLE INTERSECTION AND INSTANCE TRANSFORMATION FOR RAY TRACING

From <https://www.freepatentsonline.com/y2023/0206541.html>

Listening to Andy on the Arch reveal seems to more or less in line with what the patents have been discussing.

Dynamic VGPR is not specific to RT but seems there in patents and patches been out for a year.


AMD is cooking some more stuffs for upcoming archs
New radical shader reordering and sorting mechanism to minimize divergence described here

STREAMING WAVE COALESCER CIRCUIT - ADVANCED MICRO DEVICES, INC.

<div p-id="p-0001">A Streaming Wave Coalescer (SWC) circuit stores a first set of state values associated with a first subset of threads of a first wave in a bin based on each of the first subset of t

www.freepatentsonline.com

Some patents like BVH nodes with delta instances, BVH overlaying etc. did not make it

 

[SolidQ]

AMD is cooking some more stuffs for upcoming archs
New radical shader reordering and sorting mechanism to minimize divergence described here

for RDNA5 that what Kepler hinted?

 

[Heartbreaker]

for RDNA5 that what Kepler hinted?

So wait for RDNA, RDNA2, RDNA3, RDNA4, UDNA...

 

[Kepler_L2]

for RDNA5 that what Kepler hinted?

SWC is in gfx13 yeah

 

[SolidQ]

Now interesting how they gonna increase RT further, because it's going to console for long time. Games will use it.

 

[eek2121]

Now interesting how they gonna increase RT further, because it's going to console for long time. Games will use it.

There is still plenty of room for improvement.

 

[gaav87]

That form AMD site. Native vs UPS + FG

Thats looks bad slower than 4080 xtx 5070ti slightly above 4070tis

 

[Hail The Brain Slug]

Thats looks bad slower than 4080 xtx 5070ti slightly above 4070tis

Isn't this exactly where we expect it to land in a game with RT?

 

[Kepler_L2]

Thats looks bad slower than 4080 xtx 5070ti slightly above 4070tis

AMD sucks at CryEngine

 

[marees]

Isn't this exactly where we expect it to land in a game with RT?

Yup. The 9070xt > 4070ti super in each & every single game

But mostly slightly falls behind 5070 ti

 

[eek2121]

Something I just now realized is that AMD is now ahead of NVIDIA in terms of performance/watt and performance/area for raster. Also possibly equal in RT. (NVIDIA has more cores so I suspect that is a big reason for the gap)

This is comparing the 9070XT to the 5070 Ti.

They were in the perfect position to drop a halo product. Of course the halo product got canned.

 

[SolidQ]

Spiderman RT results suspicious, almost double compare to RTX 4080
AMD saying it's maxed RT
Fabio was mention he was surprise in SM2 results even with RTX 5080

 

[eek2121]

Spiderman RT results suspicious, almost double compare to RTX 4080
AMD saying it's maxed RT
Fabio was mention he was surprise in SM2 results even with RTX 5080

It is a game designed for AMD hardware. What is suspicious?

 

[Mahboi]

AMD sucks at CryEngine

I felt they did well this generation, typically it's more CringeEngine

Thats looks bad slower than 4080 xtx 5070ti slightly above 4070tis

what's bad about it? My expectation from forever ago was RT at the level of a 4070 Ti with clearly better raster. If we get 4070 Ti Super RT and ~10%> 7900 XT raster, it's a literal 1440p bomb. Weak at 4K, but seriously, past all the marketing, nobody should expect a 70 XT card to be a good 4K card...

 

[SolidQ]

It is a game designed for AMD hardware. What is suspicious?

I understand, but double compared to 4080?
We need more data in what scenario RDNA4 RT is strong

 

[TESKATLIPOKA]

I am pretty impressed about RDNA4 to be honest. AMD improved everything in which they were behind and MRSP looks good, If you actually could get one at that price.

 

[Mahboi]

I wonder if their improved h264 encoder is from Xilinx. It should be about the right time to switch IPs, 20% improvement is pretty massive, and AMD never was good h264. I tried streaming on Twitch with my 7900 XT and the quality is about on par with a Pascal or Turing era GPU. AV1 is excellent though, but that's not possible on Twitch......still......in 2025....

 

[Heartbreaker]

Something I just now realized is that AMD is now ahead of NVIDIA in terms of performance/watt and performance/area for raster. Also possibly equal in RT. (NVIDIA has more cores so I suspect that is a big reason for the gap)

This is comparing the 9070XT to the 5070 Ti.

They were in the perfect position to drop a halo product. Of course the halo product got canned.

Oddly driven by a massive leap in transistor density. I wonder how they did that.

A change of process might put NVidia back in the lead. GB203 is only 45.6B transistors vs 53.9B for N48.

Plus 5070 Ti is the cut down part on GB203, 5080 is the full part.

 

[SolidQ]

More than 10 RTX 5090

 

[gaav87]

Isn't this exactly where we expect it to land in a game with RT?

No i mean its kingdom come 2 no RT in this game. Expected little higher.
But anyway judging by the tech power up game selection for it tests 5070ti will do decent.

 

[gdansk]

A change of process might put NVidia back in the lead. GB203 is only 45.6B transistors vs 53.9B for N48.

I'm not sure transistor counts are as comparable as people would like.
It's 2x the transistor density of Intel's B580 on what is supposed to be a similar density process. Is it really that many less transistors or are they counting differently?