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

 

[leoneazzurro]

Probably the +10% is in reality a -10% (also, the 5700XT is not exactly the better perf/w part in the RDNA1 family). In that case, the numbers could fit more .

 

[blckgrffn]

Probably the +10% is in reality a -10% (also, the 5700XT is not exactly the better perf/w part in the RDNA1 family). In that case, the numbers could fit more .

Well that's just it. The 5600 with it original bios was a really solid Perf/W player right?

AMD hasn't told us they are seeing the 50% gains at TBP of 250W or on their 150W parts. It seems unlikely that it will scale linearly across the entire top to bottom card range.

 

[GodisanAtheist]

Its really shocking the night and day difference between the absolute waterfall of NV rumors and leaks vs the almost total blackout on AMD's side of the equation.

I can only hope we'll get a reveal shortly after the NV Ultimate event.

Nvidia has rained on AMD's parade for years, always having a product or price adjustment or reveal to steal some of their competitor's thunder. Would be interesting to see AMD hit back with the same tactics for once.

 

[Glo.]

Its really shocking the night and day difference between the absolute waterfall of NV rumors and leaks vs the almost total blackout on AMD's side of the equation.

I can only hope we'll get a reveal shortly after the NV Ultimate event.

Nvidia has rained on AMD's parade for years, always having a product or price adjustment or reveal to steal some of their competitor's thunder. Would be interesting to see AMD hit back with the same tactics for once.

"Bad news in this industry spread with the speed of light, while good news always are kept as close to your chest as possible."

Look at Intel woes, and how fast they spread out.

 

[Gideon]

"Bad news in this industry spread with the speed of light, while good news always are kept as close to your chest as possible."

Look at Intel woes, and how fast they spread out.

Well Bulldozer and VEGA were hyped till release

 

[Viking Warrior]

I expect the new AMD cards to very near these specs,with minor differences in clock speed.
Navi 21 6900xt 72 CU Clock 2020 Bus 384 Bandwidth 672 265 watts TFLOPS 18.6
Navi 21 6850 64 CU Clock 1850 Bus 384 Bandwidth 672 245 watts TFLOPS 15.1
Navi 23 6700 56 CU Clock 1900 Bus 256 Bandwidth 448 225 watts TFLOPS 13.6
Navi 23 6650 48 CU Clock 1800 Bus 256 Bandwidth 448 200 watts TFLOPS 11.1
Navi 10 6600 40 CU Clock 1850 Bus 256 Bandwidth 448 180 watts TFLOPS 9.4
Navi 10 6500 36 CU Clock 1700 Bus 256 Bandwidth 448 165 watts TFLOPS 7.8

 

[USER8000]

Its really shocking the night and day difference between the absolute waterfall of NV rumors and leaks vs the almost total blackout on AMD's side of the equation.

I can only hope we'll get a reveal shortly after the NV Ultimate event.

Nvidia has rained on AMD's parade for years, always having a product or price adjustment or reveal to steal some of their competitor's thunder. Would be interesting to see AMD hit back with the same tactics for once.

The hype is unreal,which reminds me of Fermi. Nvidia does not hype before launches.

 

[Glo.]

I expect the new AMD cards to very near these specs,with minor differences in clock speed.
Navi 21 6900xt 72 CU Clock 2020 Bus 384 Bandwidth 672 265 watts TFLOPS 18.6
Navi 21 6850 64 CU Clock 1850 Bus 384 Bandwidth 672 245 watts TFLOPS 15.1
Navi 23 6700 56 CU Clock 1900 Bus 256 Bandwidth 448 225 watts TFLOPS 13.6
Navi 23 6650 48 CU Clock 1800 Bus 256 Bandwidth 448 200 watts TFLOPS 11.1
Navi 10 6600 40 CU Clock 1850 Bus 256 Bandwidth 448 180 watts TFLOPS 9.4
Navi 10 6500 36 CU Clock 1700 Bus 256 Bandwidth 448 165 watts TFLOPS 7.8

Unfortunately incorrect CU counts for particular GPUs.

GPUs - GPU dies.

 

[soresu]

Nvidia does not hype before launches.

You must be occupying a different space time to the one where Jensen was 'cooking' a GPU in an oven.

If that isn't a CEO hyping something I really don't know what is.

 

[USER8000]

You must be occupying a different space time to the one where Jensen was 'cooking' a GPU in an oven.

If that isn't a CEO hyping something I really don't know what is.

You misunderstood my words. Ampere hype is huge.ie,21 days of twitter,the ceo and his oven,newsmedia saying unreal performance. Pascal did not have the hype,etc.

 

[Shmee]

Well Bulldozer and VEGA were hyped till release

Hmm, Bulldozer sure wasn't great, but Vega was pretty good, as long as you got a quality card. The launch wasn't super smooth, but the later cards with quality cooling were actually pretty impressive and competitive with GP 104(1070Ti, 1080)

 

[linkgoron]

Hmm, Bulldozer sure wasn't great, but Vega was pretty good, as long as you got a quality card. The launch wasn't super smooth, but the later cards with quality cooling were actually pretty impressive and competitive with GP 104(1070Ti, 1080)

Vega was probably supposed to compete with the 1080ti (given the die size was even bigger than GP102 and also used HBM2), and in that sense it was a complete failure.

I can only assume that Polaris lived for so many years just because Vega was so bad that it just didn't make sense for the lower-end. At ~500mm^2 and using HBM2 it barely competed with 312mm^2 GP104.

 

[Qwertilot]

Vega was probably supposed to compete with the 1080ti (given the die size was even bigger than GP102 and also used HBM2), and in that sense it was a complete failure.

I can only assume that Polaris lived for so many years just because Vega was so bad that it just didn't make sense for the lower-end. At ~500mm^2 and using HBM2 it barely competed with 312mm^2 GP104.

No need to speculate in that case - a very obvious reason why Polaris ran for so long. AMD were near bankrupcy for quite a while and quite rightly focused on Zen & console stuff. The GPU division got very short changed and so you never saw logical top to bottom refreshes etc.

That is now obviously somewhat behind them and their competitive situation in GPU's should start to improve over time. It will very likely take a while though, it isn't really entirely fair to expect instant miracles.

 

[Stuka87]

Vega was probably supposed to compete with the 1080ti (given the die size was even bigger than GP102 and also used HBM2), and in that sense it was a complete failure.

I can only assume that Polaris lived for so many years just because Vega was so bad that it just didn't make sense for the lower-end. At ~500mm^2 and using HBM2 it barely competed with 312mm^2 GP104.

No. Vega was designed to be a compute card for the enterprise, which AMD made a consumer version of. They squeezed whatever performance they could out of it, and shipped it. Wherever it landed is where it landed.

Work on Navi was most likely started before Polaris was released. As there is typically a 2-3 year run up to release a new GPU, especially if it also includes a new ISA. AMD sold tons of Polaris based cards, which held them over while they worked on creating a whole new product stack, which is basically the same thing that they did for the CPU side of things.

 

[Mopetar]

You must be occupying a different space time to the one where Jensen was 'cooking' a GPU in an oven.

If that isn't a CEO hyping something I really don't know what is.

You mean to tell me he wasn't giving a tour of NVidia's newest and state of the art GPU repair facility?

Kidding aside it does seem like NVidia is letting a lot more information out ahead of launch. Whether you want to call it "hype" or not is another matter, but it is a bit unusual for NVidia.

 

[DDH]

I was digging in to the hot chips presentation from Microsoft regarding Xbox series X. The quoted total transistors are 15.3 billion, but that seems a bit low for 56cu GPU, 8 core CPU and else. 5700xt with 8 gddr6 bus's was 10.3 billion, and a Zen2 chiplet is 3.9 billion. Already were at 14.2 without any additional CU's for the GPU, the soc or additional memory bus's. Seems weird?

Also, i was trying to measure up the CUs on the MS slide, i came out to 3.6mm2 per DCU. This was close to 20% smaller than the Navi DCU, but the same is also try of the gddr bus's and the zen2 chips. My total estimation for the GPU are of the chip was 47.8%, and for comparison Tom's hardware got 47.5

MS says 7nm Enhanced, but i thought they were referring to N7P, seeing as DUV and EUV are not ip compatible, and I didn't see AMD redesigning Zen2 for the smaller node, but now I'm not so sure

Sent from my SM-N975F using Tapatalk

 

[moinmoin]

MS says 7nm Enhanced, but i thought they were referring to N7P, seeing as DUV and EUV are not ip compatible, and I didn't see AMD redesigning Zen2 for the smaller node, but now I'm not so sure

Back when it became known that AMD would again do the chips for both Sony and Microsoft and TSMC announced both N7 and N7+, the latter was said to be less expensive than the former for mass production so it made sense for consoles to use that. As of now I have no ******* idea anymore what is what with all those subnodes by TSMC, so... *shrugs*

 

[Stuka87]

I was digging in to the hot chips presentation from Microsoft regarding Xbox series X. The quoted total transistors are 15.3 billion, but that seems a bit low for 56cu GPU, 8 core CPU and else. 5700xt with 8 gddr6 bus's was 10.3 billion, and a Zen2 chiplet is 3.9 billion. Already were at 14.2 without any additional CU's for the GPU, the soc or additional memory bus's. Seems weird?

Also, i was trying to measure up the CUs on the MS slide, i came out to 3.6mm2 per DCU. This was close to 20% smaller than the Navi DCU, but the same is also try of the gddr bus's and the zen2 chips. My total estimation for the GPU are of the chip was 47.8%, and for comparison Tom's hardware got 47.5

MS says 7nm Enhanced, but i thought they were referring to N7P, seeing as DUV and EUV are not ip compatible, and I didn't see AMD redesigning Zen2 for the smaller node, but now I'm not so sure

Sent from my SM-N975F using Tapatalk

Zen2 would not require an entire redesign to fit onto an updated N7 process. But even if so, the processors are already tweaked from a standard version because their memory interface is different, and they are being put on the same package as the GPU. So I don't think it should be surprising if its on a different process from a a standard Zen2 processor.

 

[A///]

I've followed GPU releases for a long time. This is one of the rare times that NVidia went above and beyond their typical marketing and hype, and that justification video this morning was interesting to say the least.

Edit: I'd expect something up to 2 weeks after NVidia's announcement. Remember, AMD is working on two fronts, not just one. They have a clear idea of what one competitor has in stock up to the next few years.

Without namedropping, do you recall a certain accented idiot who went on some weird deranged campaign in regard to Zen2 cores, speed and pricing? For some reason many in the community believed in that moron and were mad at AMD because AMD barely acknowledged something they'd said, and thus they became the word of God overnight.


Let your product sell itself. Not baseless promises, because the latter sure as hell didn't work for a decade.

 

[Veradun]

AMD hasn't told us they are seeing the 50% gains at TBP of 250W or on their 150W parts. It seems unlikely that it will scale linearly across the entire top to bottom card range.

I dare say linear power figures are impossible for very obvious reasons. But we are in the dark as to what they use for comparison (57XT or 57)

 

[Veradun]

Vega was probably supposed to compete with the 1080ti (given the die size was even bigger than GP102 and also used HBM2), and in that sense it was a complete failure.

I can only assume that Polaris lived for so many years just because Vega was so bad that it just didn't make sense for the lower-end. At ~500mm^2 and using HBM2 it barely competed with 312mm^2 GP104.

One might think a tsmc's specialized 16FF derivative and hbm2s delivered on promises would have told a different story, but who knows. The situation looks exactly reversed this time around and it's kinda fun and sad at the same time.

 

[A///]

One might think a tsmc's specialized 16FF derivative and hbm2s delivered on promises would have told a different story, but who knows. The situation looks exactly reversed this time around and it's kinda fun and sad at the same time.

?

 

[Veradun]

?

The one player with underdelivering process and memory subsystem overclocked to hell and beyond might not be AMD this time.

 

[A///]

The one player with underdelivering process and memory subsystem overclocked to hell and beyond might not be AMD this time.

The Green Giant? I figured that's what you were inferring with your vague comparison but I wasn't 100% sure. It's remarkable when you think about how the scripts have flipped here too if anything leaked is to be believed. We've been witnessing a slowed down train wreck the last 3 years, and even 3 years from now with the blue squad. It's enough to question the very fiber that's entrenched itself in the tech world that we've grown accustomed to.

 

[kurosaki]

And in five years all of our gaming rigs will be the ARM X4 anyways!