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

 

[TESKATLIPOKA]

Yeah it could easily be an ES at close to 400w though.
If there is a big Navi that is 2x 5700XT that’s still less than, what 24B transistors and at the largest rumored die size of 505mm2 makes it not much more density than Navi 10 (24B/505).
Just don’t see a high end part coming within reach of the 3090.
I expect some competitive SKUs though!

We don't know how much transistors Big Navi really has.
On the other hand I don't believe Ampere has 35B transistors.
It's almost 2x more compared to RTX 2080Ti (18.6B), yet It has only 5248Cuda(84SM)? Most likely a cut down version, but even the full die won't have much more, maybe 5760Cuda(90SM), that's only 25% more, what is the rest for? I don't believe Tensor and RTX need so much space.
Big Navi will most likely have comparable or a bit higher clocks than Navi 10, Ampere probably will stay at the same clocks, number of units is higher on Ampere, so the only question is how much better is the IPC for both of them.
I won't speculate which one will be faster, because I don't know.

 

[TESKATLIPOKA]

Filll with what? Cries?
My feeling is that, if those craze Ampere prices are really true, I suspect that Nvidia is sure that RDNA2 will be a competitor to Turing, not to Ampere. I fear that Big Navi will go as far as between 3060 and 3070...
Pessimistic thoughts, yes, but still, given the retrospect isn't impossible.

So you believe RTX 3060 or RTX 3070 is ~20-30% faster than RTX 2080Ti?
That's impossible when big Ampere has "only" 5248 Cuda and RTX 2080Ti has 4352 Cuda. These chips will have a lot less Cuda cores, so how can they be faster than RTX 2080Ti? Or you actually believe Big Navi won't be faster than RTX 2080Ti?
Seriously I don't understand why some users are so optimistic or so pessimistic about Big Navi.

 

[jpiniero]

We don't know how much transistors Big Navi really has.
On the other hand I don't believe Ampere has 35B transistors.it
It's almost 2x more compared to RTX 2080Ti (18.6B), yet It has only 5248Cuda(84SM)?

Ray Tracing for sure is a lot of the transistor count increase. They might even have overdone it...

 

[TESKATLIPOKA]

Ray Tracing for sure is a lot of the transistor count increase. They might even have overdone it...

If I am not mistaken, then Tensor cores use more space than RT cores and even combined It doesn't cost a massive amount of transistors. Do you believe RT cores or Tensor cores were more than doubled per SM? Currently It's 1 RT core and 8 tensor cores per SM and I have to wonder how many do you need to increase the transistor count by at least 10B, unless Ampere has a significant increase in IPC, which would explain the transistor count.
RDNA1 also used a lot of transistor to increase It's IPC compared to Polaris.

 

[jpiniero]

If I am not mistaken, then Tensor cores use more space than RT cores and even combined It doesn't cost a massive amount of transistors. Do you believe RT cores or Tensor cores were more than doubled per SM? Currently It's 1 RT core and 8 tensor cores per SM.

RT cores yes. Tensor cores maybe. Probably won't know how good AMD's RT implementation is until reviews.

 

[TESKATLIPOKA]

Considering the consoles have AMD GPU, It won't be such a big problem. Nvidia can have the better implementation, but they use a separate unit for them while AMD has RT inside TMU.

 

[A///]

Are the AMD GPUs in the new consoles 70% of their desktop counterparts? I never really understand CUs and how they're calculates for cores. Not that it ever mattered until now.

 

[TESKATLIPOKA]

One CU(Compute unit) consists of 64 stream processors("cores" or unified shaders) and big Navi has 80CU -> 5120 cores.

 

[A///]

One CU(Compute unit) consists of 64 stream processors("cores" or unified shaders) and big Navi has 80CU -> 5120 cores.

Ah that makes sense. I'd watched a few videos recently but couldn't quite figure out how the math was being done. I'm guessing there is no direct comparison to CUDA or Tensor cores, right?

 

[TESKATLIPOKA]

Stream processor and Cuda core are unified shaders. Tensor core or RT core has a different purpose.

 

[Mopetar]

Seriously I don't understand why some users are so optimistic or so pessimistic about Big Navi.

Have you seen an AMD GPU launch the last five years? It's always a massive amount of hype and in many ways a similarly massive amount of let down.

That said there's certainly more reason than ever to be optimistic, but there's a lot of recent history of AMD's top dog barely managing to slot in at the top of NVidia's midrange or the bottom of their high-end. There's a reason people reach back to Evergreen and Fermi to describe a scenario where AMD actually comes out on equal (or near enough so) footing.

 

[maddie]

Have you seen an AMD GPU launch the last five years? It's always a massive amount of hype and in many ways a similarly massive amount of let down.

That said there's certainly more reason than ever to be optimistic, but there's a lot of recent history of AMD's top dog barely managing to slot in at the top of NVidia's midrange or the bottom of their high-end. There's a reason people reach back to Evergreen and Fermi to describe a scenario where AMD actually comes out on equal (or near enough so) footing.

To be fair, Evergreen dominated, but the 290X was battling right up there with Nvidia's best.

 

[DDH]

Have you seen an AMD GPU launch the last five years? It's always a massive amount of hype and in many ways a similarly massive amount of let down.

That said there's certainly more reason than ever to be optimistic, but there's a lot of recent history of AMD's top dog barely managing to slot in at the top of NVidia's midrange or the bottom of their high-end. There's a reason people reach back to Evergreen and Fermi to describe a scenario where AMD actually comes out on equal (or near enough so) footing.

Fury, Vega were the disappointing launches. And rewind 4 years ago the same could have been said for AMD CPU division.

It is hubris to think NVIDIA will be always be untouchable in the high end

Sent from my SM-N975F using Tapatalk

 

[Olikan]

It is hubris to think NVIDIA will be always be untouchable in the high end

While true, Nvidia this time really went "TRY HARD" mode... i mean, can "power of engeenring" really beat Gddr6X, 12 layer pcb, triple slot cooler and ~120mm more die size?

...eh, Nvidia will have to regress in alot of metrics to lose so hard

 

[kurosaki]

While true, Nvidia this time really went "TRY HARD" mode... i mean, can "power of engeenring" really beat Gddr6X, 12 layer pcb, triple slot cooler and ~120mm more die size?

...eh, Nvidia will have to regress in alot of metrics to lose so hard

In the end, for me, it's still all in the pricing. If team N is taking 200 USD more for a 10% performance boost. They are out. But I get it, someone always want the creme de la creme. No matter the cost. Just remember the new consoles are costing approx 500 USD. Anything above that is AMD or nvidia screwing us over. Double the diesize of a console Apu? Still not ok.

 

[Kenmitch]

While true, Nvidia this time really went "TRY HARD" mode... i mean, can "power of engeenring" really beat Gddr6X, 12 layer pcb, triple slot cooler and ~120mm more die size?

...eh, Nvidia will have to regress in alot of metrics to lose so hard

Best to just wait for the reviews and see what happens.

The " power of engineering " comment is silly at best. Wouldn't it be the " power of engineering " that determined the winner either way?

 

[DDH]

While true, Nvidia this time really went "TRY HARD" mode... i mean, can "power of engeenring" really beat Gddr6X, 12 layer pcb, triple slot cooler and ~120mm more die size?

...eh, Nvidia will have to regress in alot of metrics to lose so hard

We don't know AMD die size. we don't know what node NVIDIA is on, we don't know what node rdna2 is on, we dont know what memory rdna2 uses.

Let's put something in to perspective that we do know however, Xbox series X GPU is 172mm2 for 56CU, excluding memory bus/controllers. 172mm2. 56CU. 140w power max.

Sent from my SM-N975F using Tapatalk

 

[kurosaki]

We don't know AMD die size. we don't know what node NVIDIA is on, we don't know what node rdna2 is on, we dont know what memory rdna2 uses.

Let's put something in to perspective that we do know however, Xbox series X GPU is 172mm2 for 56CU, excluding memory bus/controllers. 172mm2. 56CU. 140w power max.

Sent from my SM-N975F using Tapatalk

I can understand why Nvidia went "all in" here. if we (Wrongly) follow linear scaling up to 80 CU's the wattage would still be below 200W. The power of a double 5700XT would be enough to have Nvidia swet a bit. Add to that, there is obviously functioning RT and VRS (Shown by sony and MS). Nvidia is kind of keen keeping it ahead, being the top brand. It's not in their interest to be no2. So they went all out with everything, and managed to cool 350W on air. I'm just happy if I could get hold of a card with 2x 5700XT perf, for under 450 usd. More of a principle, im not to keen on giving away money for air. Remember that the new consoles does not cost more than 500 usd. I'm not a fool. I'm never going to buy a 1500 usd GPU (Inflation counted for), because that's just being an asshole.

 

[jpiniero]

My guess is the top model will be a thousand.

 

[DiogoDX]

My guess is the top model will be a thousand.

I think that will be $800 max for the top one if they want it to sell. They don't have the brand power of nvidia to sell 1000+ cards.

 

[kurosaki]

I think that will be $800 max for the top one if they want it to sell. They don't have the brand power of nvidia to sell 1000+ cards.

Well, that kind of depends on the power of the card and the pricing of the competitor now. Dosen't it? If "Big Navi" performed as good as 3090 that's rumoured to cost 1500 usd+, then it would make sense to price it at 999 usd, partially to show them the finger (Look Nvidia is over pricing!), partially because they still would make insane amounts of money at that range and still come out as the good guy. They would manage good on a price range of 500 usd also, but everyone seemed to have forgotten that, so i'm giving up.

 

[AtenRa]

My estimate, Top model with 16GB of ram at $999 will be 10% faster than 3080 10GB at $999 and 3080 20GB at $1199

 

[Konan]

Chinese/Korean forums saying late October announcement & availability in November.
Also estimating pricing and performance levels. Folks won't want to hear the performance & wattage so I won't post yet as all unverified and nothing else out that yet corresponding/collaborating.

SKU​	Est.USD
6900 XT	$799
6900	$699
6800 XT	$599
6800	$499
6700 XT	$399
6700	$299

 

[Karnak]

Availability in November yet estimated pricing. Just lol. Chinese folks are the worst when it comes to early "leakes".

RDNA2 will launch before November and there won't be a total of 6 (!) SKU's. Not gonna happen.

 

[Konan]

RDNA2 will launch before November and there won't be a total of 6 (!) SKU's. Not gonna happen.

No not all SKUs are for launch at release time no.
Based on other credible leakers saying the PCB design was only completed not long ago and drivers optimization happening I can see that it could be out in October. That fits with the other rumor of October 7.
The main rumor was that AMD didn't release a 5900 XT was due to two main points. They main one was that it had massive heat problems as the voltage was raised past the sweet spot and was therefore cancelled plus RDNA1 still had GCN pedigree so they wanted it to be all RDNA2 with the hopes that the problems seen were brought too a more controllable level (remains to be seen). This change in redesign increased the delivery time (the speculation for November push out). As for launch timing we'll see what comes...