Question Speculation: RDNA2 + CDNA Architectures thread

[uzzi38]

All die sizes are within 5mm^2. The poster here has been right on some things in the past afaik, and to his credit was the first to saying 505mm^2 for Navi21, which other people have backed up. Even still though, take the following with a pich of salt.

Navi21 - 505mm^2

Navi22 - 340mm^2

Navi23 - 240mm^2

Source is the following post: https://www.ptt.cc/bbs/PC_Shopping/M.1588075782.A.C1E.html

 

[PhoBoChai]

You guys with your back & forth about leakers and tech tubers make this tech forum have the same quality (LOW) as reddit. Basically if you don't believe or believe a leaker, leave it be, ppl can view and decide for themselves, there's no need to get so side tracked and it becomes a discussion over personalities instead of what we're all here for: GPU tech.

 

[Vixis Rei]

Coreteks's leak has been declared fake by leaker Kittyyyuuko

https://twitter.com/x/status/1312585634951778304

EDIT: Maybe it's referred to the picture itself, but that was already declared as a photoshop edit by the author of the video

I wonder if kitty knows it's modified / photoshopped? In this case just saying it's fake isn't helpful. Doubt kitty would tell how similar it is to the real thing if they do know.

 

[ModEl4]

In Anandtech's coverage of Hot Chips Series X event, in the interview section there was a reply from the Microsoft team in a question, and they said that the ray tracing portion is only 5% of the GPU transistor budget. The thing is that I cannot find it anymore, does anyone else remembers this?

 

[xilli_fiberbit]

I wonder if kitty knows it's modified / photoshopped? In this case just saying it's fake isn't helpful. Doubt kitty would tell how similar it is to the real thing if they do know.

The part with caps has been taken from NAVI10 but it has modified too the QR code on the corner, it's weird and i don't know why

Why it not showing the board by hiding some elements ?

 

[moinmoin]

In Anandtech's coverage of Hot Chips Series X event, in the interview section there was a reply from the Microsoft team in a question, and they said that the ray tracing portion is only 5% of the GPU transistor budget. The thing is that I cannot find it anymore, does anyone else remembers this?

You mean this?

I don't remember any more specific transistor budget being mentioned.

 

[kurosaki]

Wonder how they implement the RTRT. Is one ray needed per pixel? That grows to quite many gigarays pretty fast. 1440x2560x60fps is like 220 something gigapixles/second. If light bounces three times before it is rendered as a pixel, well that is 660 giga-rays needed to be calculated per sec. Must be something that does not need a setup that isn't one ray per pixel anyways. Anyone knows how it kind of works behind the scenes?

 

[Konan]

Wonder how they implement the RTRT. Is one ray needed per pixel? That grows to quite many gigarays pretty fast. 1440x2560x60fps is like 220 something gigapixles/second. If light bounces three times before it is rendered as a pixel, well that is 660 giga-rays needed to be calculated per sec. Must be something that does not need a setup that isn't one ray per pixel anyways. Anyone knows how it kind of works behind the scenes?

Got this to go off... The ratio for texture unit v RT is 1:1 (edit)

There have also been a few rumours that say not all CU throughput is there when RT is happening. I.e. around 20% of CUs focused on RT efforts.
Similar to how doubling of CUDA cores with ampere doesn’t give a doubling of perf. So like 80CU is not 80CU of throughput when RT happening. That said who knows just a few rumors on the grapevine...I’d take those with some salt but not much else to go on

 

[Asterox]

You guys with your back & forth about leakers and tech tubers make this tech forum have the same quality (LOW) as reddit. Basically if you don't believe or believe a leaker, leave it be, ppl can view and decide for themselves, there's no need to get so side tracked and it becomes a discussion over personalities instead of what we're all here for: GPU tech.

My favorite leaker is David Wang, but he is mostly ignored by everyone.

 

[maddie]

Wonder how they implement the RTRT. Is one ray needed per pixel? That grows to quite many gigarays pretty fast. 1440x2560x60fps is like 220 something gigapixles/second. If light bounces three times before it is rendered as a pixel, well that is 660 giga-rays needed to be calculated per sec. Must be something that does not need a setup that isn't one ray per pixel anyways. Anyone knows how it kind of works behind the scenes?

It's 220 megapixels/sec for that resolution & framerate.

 

[kurosaki]

Sorry, I must have drunken something bad! 😅 Very true!

 

[Glo.]

Got this to go off... The ratio for texture unit v RT is 4:1
View attachment 31014
There have also been a few rumours that say not all CU throughput is there when RT is happening. I.e. around 20% of CUs focused on RT efforts.

Those rumors are complete BS.

 

[Konan]

Those rumors are complete BS.

No one can prove either way. Accept, they are persistent. Maybe where there is smoke there is fire for a hybrid architectural RT approach...

 

[moinmoin]

My favorite leaker is David Wang, but he is mostly ignored by everyone.

Finally we know the source of all the fishy code names!

 

[ModEl4]

You mean this?

I don't remember any more specific transistor budget being mentioned.

Based on the slide, there was a question and MS team replied with the 5% figure, but maybe I remember it wrong.

 

[uzzi38]

No one can prove either way. Accept, they are persistent. Maybe where there is smoke there is fire for a hybrid architectural RT approach...

Yeah they can, they can just go rewatch the Series X HotChips presentation, the Road to PS5 presentation, or alternatively go patent digging.

It's not difficult to filter that rumour out and laugh about it.

In the first place, how did nobody realise that dynamically adjusting CUs for performing BVH traversal and intersection is NOT HARDWARE BASED RAYTRACING.

 

[dzoni2k2]

Are those the same people who say 'RTX games' wont work on AMD GPUs 'cause only Nvidia supports RTX! Nvidia brainwashed entire generation.

 

[ModEl4]

Texture/Ray-box ratio is 1:1
Texture/Ray-triangle ratio is 4:1
It's in the MS slides if you do the calculations.
The fun thing would be, if there was a sci-fi (lol) deterministic correlation between actual in game Ray-triangle performance vs theoretical max like we have for actual real triangle output vs theoretical max like in PS5 unreal engine 5 demo for example. 19-20 million triangles per frame (30fps) at 2230MHz so around 1/15 - 1/16 of the theoretical max (4 primitives) , this would give for PS5 Ray-triangle (if we calculate with 1/16) around 5Grays/s actual performance (RTX 2060 level?) and for Xbox SeriesX around 6Grays/s actual performance. Now that is an analysis worthy of the Sheldon award🤪

 

[Konan]

Hey I'm not arguing, just playing devils advocate. Ok here is a patent (already had read it last week when it was published)
Ray Intersection Unit inside of the CU, patent describes the functions of RT. When I posted it was that there have been persistent rumors regarding CU throughput when RT is involved. Patent is inconclusive on performance but I still feel the questions is open.

 

[leoneazzurro]

In the slide it's clearly written 4 Ray ops/CU per clock OR 4 TMU ops per clock. Then there is ray/triangle which is 1/4 ratio. A ray tracing unit may resove more than one OP per clock, just as a SIMD unit resolves more than one operation per clock...

 

[Ajay]

Yeah they can, they can just go rewatch the Series X HotChips presentation, the Road to PS5 presentation, or alternatively go patent digging.

It's not difficult to filter that rumour out and laugh about it.

In the first place, how did nobody realise that dynamically adjusting CUs for performing BVH traversal and intersection is NOT HARDWARE BASED RAYTRACING.

Yes it is. It’s just that the hardware ray tracing isn’t done by a fixed function unit.

 

[soresu]

around 5Grays/s actual performance (RTX 2060 level?)

I'm high suspicious over the figures nVidia supplied for gigarays per second on their RTX 2000 cards - considering they are fine with giving tensor sparse calculation numbers of equivalent to n TFLOPS rather than exact figures I do not put it past them to fudge the numbers on gigarays /s to push 'RTX' in its infancy.

Considering the backlash against the lackluster performance of RTX early on it seems like a likely strategy to stave off especially bad PR until they got chunkier cards and better DLSS out the door.

 

[Hans Gruber]

Here is the problem AMD will face in November. If Big Navi is not as good as the 3080 and you have to start factoring in the power savings over Nvidia. Then you combine the release of the two new consoles from Sony and Microsoft. People will question buying any AMD or Nvidia cards because of cost compared to the new consoles. I realize that most if not all here are PC gamers. But there is a large percentage of kids who would put off building a new PC or upgrading their graphics cards when the new consoles are released.

With that said I am either getting the 3070 or one of the big Navi variants.

 

[ModEl4]

I'm high suspicious over the figures nVidia supplied for gigarays per second on their RTX 2000 cards - considering they are fine with giving tensor sparse calculation numbers of equivalent to n TFLOPS rather than exact figures I do not put it past them to fudge the numbers on gigarays /s to push 'RTX' in its infancy.

Considering the backlash against the lackluster performance of RTX early on it seems like a likely strategy to stave off especially bad PR until they got chunkier cards and better DLSS out the door.

I was just messing around, no need to analyze it too much😉 Maybe Nvidia numbers are overstated, probably, I don't know, but you have to consider also that the 5Grays/s actual number I came with, it has the necessary condition that each and every CU will do only Ray ops and zero Texture ops, each CU cannot perform at the same time 4 texture and 4 ray ops if I'm not mistaken.

 

[PhoBoChai]

I was just messing around, no need to analyze it too much😉 Maybe Nvidia numbers are overstated, probably, I don't know, but you have to consider also that the 5Grays/s actual number I came with, it has the necessary condition that each and every CU will do only Ray ops and zero Texture ops, each CU cannot perform at the same time 4 texture and 4 ray ops if I'm not mistaken.

You never go by NV's claims, they have proven over the many years, it's a lot of marketing fluff.

As an example, there's an UE RT demo, just a simple small puddle of water reflection, and a 2060S is crippled in perf running it. There's not a chance any of the Turing GPUs can do the UE5 PS5 demo at 30 FPS.

RDNA 2 TMUs do 4 RT ops or texture ops however, texturing steps don't necessarily overlap lighting steps.

ps. When you calculate triangles in the scene and derive the gigaray/s, you have to factor in multiple bounces.

 

[PhoBoChai]

Hey I'm not arguing, just playing devils advocate. Ok here is a patent (already had read it last week when it was published)
Ray Intersection Unit inside of the CU, patent describes the functions of RT. When I posted it was that there have been persistent rumors regarding CU throughput when RT is involved. Patent is inconclusive on performance but I still feel the questions is open.

View attachment 31100 View attachment 31102

RT always initiates first on the SIMD units. Devs call for a regular shader, which contains the ray/bvh data, that gets forwarded to the fixed function units to handle the bvh traversal & intersection testing.

You can even read RTX Ampere's whitepaper, NV does the same RT initialization in their cuda cores.

Not sure why ppl are even taking the 20/80 CU dedicated to RT seriously given both MS & Sony info reveal have dedicated RT hw units in RDNA 2.