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

 

[Viking Warrior]

I was very sceptical of all that, and assumed that what they showed was 80CU part. But let's call a spade a spade-Paul's sources are AMD themselves, what he received was controlled leaks whether he knows about it or not. There is no way in hell someone that is not an insider would know what he leaked. So if he says this was not the fastest part, I believe it.

This was definitely not the 80 CU Big Navi,it was the 60 CU version of the chip which has a much higher clock speed than the 3080.

 

[CastleBravo]

Each die will likely have 3 different skus, with the biggest and best card being nearly 3X as fast as a 5700XT.

Hype train has gone to plaid.

 

[Glo.]

Each die will likely have 3 different skus...

And how is that going to work out for Navi 22 die with 40CUs/192 bit bus, and Navi 23 with 32 CUs/128 bit bus, hmmm?

...with the biggest and best card being nearly 3X as fast as a 5700XT.

And people call me a hype train conductor...

 

[Kenmitch]

And people call me a hype train conductor...

I thought you were the ticket taker? Guess a congrats on the promotion is due!

 

[Glo.]

I thought you were the ticket taker? Guess a congrats on the promotion is due!

eek2121 appears to have snatched that job from me, by himself.

 

[TESKATLIPOKA]

In my opinion AMD will have some form of upscaling. It's

eek2121 appears to have snatched that job from me, by himself.

Are you sad?

 

[moinmoin]

It's

Ok.

 

[TESKATLIPOKA]

Ok.

It look like I didn't notice an unfinished post and It got mixed with the next one. You finished It nicely, thanks.

 

[alexruiz]

If I read correctly, AMD used a Ryzen 9 5900X to test RDNA2.
This tells me that they already know Zen 3 is faster than core 10 gen, and hence they want to get the best possible result for Navi 21.
This might provide a better picture of the already known numbers for RTX 3080 as those were run on intel cpus. On the other hand, if the tests were run at 4k, CPU should not matter...
We'll see in a few days

Looking forward to build a few Zen 3 + RDNA2 PCs.

 

[moinmoin]

It look like I didn't notice an unfinished post and It got mixed with the next one. You finished It nicely, thanks.

Honestly, after some re-reading I became unsure and thought you actually meant:
"In my opinion AMD will have some form of upscaling. It's eek2121."

I'll see myself out.

 

[ModEl4]

The only thing that is certain regarding the RX6000 mini preview was that by AMD not indicating the model position in the upcoming Big Navi stack, generated a LOT of pessimism judging by the comments in various forums and social media etc.
AMD fanboys can find the lowest results that 3080 had from the meriads of reviews for these games, assume that this was a heavily cut down part in clocks and CUs (2.0GHz 64CU 96RBEs 12GB vs 2.2GHz 80CU 128RBEs 16GB or whatever) and shoot for 3090 stardom.
And Nvidia supporters can also find the best results that 3080 had for these games, assume the settings are the same, add 1% PCI-EXPRESS scaling and another 1% for the magic ZEN3 performance/Samsung 980PRO lol and God knows what else and come with the conclusion that the highest SKU would be +10% than 3070 on average in upcoming reviews.
I'll wait for the reviews, too little info for meaningful speculations, the mini preview did not add anything concrete.
But certainly knowing the 251mm2 1.9GHz 40CU 5700XT performance (around half of 3080 performance at 4K)
a 2.2GHz 80CU part with IPC improvements and with the magic cache at freaking 536mm2 (if accurate) how it cannot be considered a failure if it is only 15% faster than a 3070 (Coretecs's first tier partner info) (Sure AMD can position it in a competitive price point but from a performance/design point it would be a massive disappointment imo and another missed opportunity for AMD)

 

[DeathReborn]

And how is that going to work out for Navi 22 die with 40CUs/192 bit bus, and Navi 23 with 32 CUs/128 bit bus, hmmm?

And people call me a hype train conductor...

Has hype train stoker been filled or...

 

[Mopetar]

Where have they said anything about a 3080ti?

The truth of the matter is the 3090 is the 3080ti in everything but name. By naming it the 3090 no one complained about the ridiculous price hike

The 3090 is basically a Titan in everything but name. Typically that's what the best 102 die has been called in the last several generations. The xx80 Ti is usually a cut down 102.

This time the cut down 102 die is quite a bit more cut down than it has been in prior generations, but it doesn't seem to matter much since nothing manages to be able to scale well enough to utilize all of the extra CUDA cores in the 3090.

Even if NVidia launches a 3080 Ti or a 3080 SUPER (I think this name is more likely) we already know that the performance is bound by the 3090 which isn't that much better than the 3080. This hypothetical card is maybe 10% better than a 3080, but unlikely to be better value per dollar. Why buy it over a 3090 if you want the best performance and why buy it over a 3080 if you want the best value?

It uses the same die as the 3080 an 3090 which are already supply constrained in a bad way. NVidia can't magically make more cards with this hypothetical Ti either.

 

[DDH]

The 3090 is basically a Titan in everything but name. Typically that's what the best 102 die has been called in the last several generations. The xx80 Ti is usually a cut down 102.

This time the cut down 102 die is quite a bit more cut down than it has been in prior generations, but it doesn't seem to matter much since nothing manages to be able to scale well enough to utilize all of the extra CUDA cores in the 3090.

Even if NVidia launches a 3080 Ti or a 3080 SUPER (I think this name is more likely) we already know that the performance is bound by the 3090 which isn't that much better than the 3080. This hypothetical card is maybe 10% better than a 3080, but unlikely to be better value per dollar. Why buy it over a 3090 if you want the best performance and why buy it over a 3080 if you want the best value?

It uses the same die as the 3080 an 3090 which are already supply constrained in a bad way. NVidia can't magically make more cards with this hypothetical Ti either.

Actually it's not, because it doesn't get the pro driver features than actual titans used too.

Now, a prosumer who would have bought a titan for professional work and needed the drivers is forced to buy the Quadro

There is no gap between a 3080 and 3090 performance wise to slot in a 3080ti, and still sell a 3090 for 1500. The 3090 is the 3080ti

 

[Mopetar]

Actually it's not, because it doesn't get the pro driver features than actual titans used too.

Now, a prosumer who would have bought a titan for professional work and needed the drivers is forced to buy the Quadro

There is no gap between a 3080 and 3090 performance wise to slot in a 3080ti, and still sell a 3090 for 1500. The 3090 is the 3080ti

Hobbled Titan then. NVidia could have provided such drivers but chose not to, but it doesn't change that the last several generations have had the best cut 102 die be classified as a Titan.

I still think that they're not using either the Titan or Ti branding at the high end because they realize that these aren't worthy of those designations. There's another clever bit in there because even if AMD does wind up having a better card than the 3080 or even the 3090 they only managed to beat a 3080 or a 3090, not a Ti or a Titan.

Now there's an interesting question. Was it shame or fear that kept NVidia from using those names? It's a clever move, but not one made by anyone confident in their position.

 

[DDH]

Hobbled Titan then. NVidia could have provided such drivers but chose not to, but it doesn't change that the last several generations have had the best cut 102 die be classified as a Titan.

I still think that they're not using either the Titan or Ti branding at the high end because they realize that these aren't worthy of those designations. There's another clever bit in there because even if AMD does wind up having a better card than the 3080 or even the 3090 they only managed to beat a 3080 or a 3090, not a Ti or a Titan.

Now there's an interesting question. Was it shame or fear that kept NVidia from using those names? It's a clever move, but not one made by anyone confident in their position.

I struggle to know where to even begin....

A hobbled titan is the xx80ti...... The titan is all about the pro driver features, thats why they charge so much for them. Its not the full die or the extra ram, its the drivers. Always has been. Thats why a full GA102 quadro costs 5 grand, even with 48gbs ram.

Even amd have their pro workstation line with the exact same chip but costs many times more because it has professional drivers that unlock performance optimisations in workstation programs

Nvidia found that to save money a person or company that needed the pro driver would buy the titan rather than the quadro. So they killed it off and spun it like your getting a titan for cheaper, except only a gamer would accept that because they dont need or even know about the pro drivers that normal go with the titan brand.

They couldnt sell the 3090 as a 3080ti because gamers and tech channels would have made noise about the price increase from 2080ti to 3080ti. So they called it the 3090 instead, as if introducing an entirely new product segment. And people bought it

They released a 3080 10gb model as low as possible to paint amd into a corner. A 3080 20gb will be several hundred more than a 3080 10gb with a higher tdp, yet perform no better.

There will be no 3080ti because there is no room for it in the stack between the 3080 and 3090(real 3080ti).


If amd end up beating the 3090 then they beat the best nvidia had to offer, regardless of what ever name it had.

 

[beginner99]

I still think that they're not using either the Titan or Ti branding at the high end because they realize that these aren't worthy of those designations

Given that NV is also dropping Quadro and Tesla names, it seems a general branding shift is taking place.

 

[eek2121]

Hype train has gone to plaid.

Not really. Big Navi has twice as many CUs as
the 5700XT and 25% higher clocks. I just felt the need to kick some of the trolls here in the teeth. The theoretical improvements of Big Navi point to a 225% increase excluding the IPC improvements. The actual improvements are obviously a bit (and only a bit) less.

 

[Saylick]

Not really. Big Navi has twice as many CUs as
the 5700XT and 25% higher clocks. I just felt the need to kick some of the trolls here in the teeth. The theoretical improvements of Big Navi point to a 225% increase excluding the IPC improvements. The actual improvements are obviously a bit (and only a bit) less.

250%, but yeah. 😉

 

[TESKATLIPOKA]

Not really. Big Navi has twice as many CUs as
the 5700XT and 25% higher clocks. I just felt the need to kick some of the trolls here in the teeth. The theoretical improvements of Big Navi point to a 225% increase excluding the IPC improvements. The actual improvements are obviously a bit (and only a bit) less.

RX 5700 XT has 1887Mhz clockspeed on average so 25% higher clockspeed for Big Navi is actually 2359Mhz, but the info is only about 2.2Ghz, which is only 16.6% better.
Another thing is that doubling the number of CUs won't increase your performance by 100%, but by ~90-95% and you also need double of ROPs.
It should be at least on the level of RTX3080, but reviews will tell us the truth.

 

[PhoBoChai]

There will be no 3080ti because there is no room for it in the stack between the 3080 and 3090(real 3080ti).

If amd end up beating the 3090 then they beat the best nvidia had to offer, regardless of what ever name it had.

There was an ES 3080Ti, it was supposed to be between 3080 and 3090.. but because we now know there's hardly any worthwhile gap, 10% between the two, it's a pointless SKU.

3090 is this gen's 3080Ti, its the fastest GAMING GPU from NV on this architecture. It's no Titan, don't let that bs marketing attempt slide.

 

[Tup3x]

RX 5700 XT has 1887Mhz clockspeed on average so 25% higher clockspeed for Big Navi is actually 2359Mhz, but the info is only about 2.2Ghz, which is only 16.6% better.
Another thing is that doubling the number of CUs won't increase your performance by 100%, but by ~90-95% and you also need double of ROPs.
It should be at least on the level of RTX3080, but reviews will tell us the truth.

You also need more bandwidth and it's not doubling. Personally I think all in all about double the 5700 XT performance is realistic exception. Since manufacturing process didn't change power usage might limit the clock speeds in 80CU version. Increased perf/W is what enables the 80CU version in the first place since 5700 XT wasn't exactly a low power card... I'm assuming that the 80CU version has ~300W TGP.

 

[GaiaHunter]

You also need more bandwidth and it's not doubling. Personally I think all in all about double the 5700 XT performance is realistic exception. Since manufacturing process didn't change power usage might limit the clock speeds in 80CU version. Increased perf/W is what enables the 80CU version in the first place since 5700 XT wasn't exactly a low power card... I'm assuming that the 80CU version has ~300W TGP.

But 7nm is alot more power efficient than the previous node.

That the 5700xt on 7nm wasn't more power efficient than the 20 series on the previous node, only points of a lot of room to improve

 

[Kuiva maa]

But 7nm is alot more power efficient than the previous node.

That the 5700xt on 7nm wasn't more power efficient than the 20 series on the previous node, only points of a lot of room to improve

The 5700XT wasn't particularly efficient because it was clocked outside its efficiency range so AMD could compete with the 2070 in the market, while creating almost two 5700XT chips for every 2070 nvidia made (die area comparison). That's not to say AMD hasn't been improving in that area with RDNA2 but the difference might be a bit misleading.

 

[TESKATLIPOKA]

Let's go back to leak about only 256bit bus for Big Navi. I don't believe It's true, even If It was GDDR6x, which is probably exclusive for Nvidia. The reason being the memory specs of consoles.
PS5: ~10 TFlops 36CU GPU and 256bit 14GHz GDDR6 with 448GB/s bandwidth
Xbox Series S: ~4 TFlops 20CU GPU and 128bit 14GHz GDDR6 224GB/s and another 32bit 14Ghz GDDR6 56GB/s
Xbox Series X: ~12 TFlops 52CU GPU and 320bit 14GHz GDDR6 560GB/s bandwidth and another 192bit 14Ghz GDDR6 336GB/s

For the 8 core Zen2 CPU 51.2-57.6GB/s(3.2-3.6Ghz 128bit DDR4) is enough bandwidth, the rest is for the GPU.
If I subtract the needed bandwidth for the CPU portion, we will end up with:
PS5: ~390GB/s for 10Tflops or 39GB/s per 1Tflop
Xbox series S: ~222GB/s for 4Tflops or 56GB/s per 1Tflop
Xbox series X: ~560GB/s + 278GB/s for 12TFlops or 47+23GB/s per 1Tflop
For comparison:
RX5700 XT: 448GB/s for ~9-9.7TFlops or ~46-50GB/s per 1TFlop

From this It's pretty clear 2GHz 80CU ~20TFlops Big Navi can't have only 256bit GDDR6! Even If I go with the lowest bandwidth per TFlops which I got from PS5, then I end up with ~780GB/s, which is a bit higher than 384bit 16GHz GDDR6(768GB/s). The same could be told about 192bit GDDR6 for 40CU GPU and 128bit GDDR6 for 32CU GPU.

P.S. I think the bandwidth in Xbox can't be added together.