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

 

[Glo.]

Expecting the top end, call it the 6700 XT I suppose, to land about 10% behind the 3080 and be priced at either $549 or $599. I expect the step down 6700 to trade blows with the 3070, perhaps edge it out at $449.

Most interesting to me would be a potential 6600 XT at 2080S performance levels for $299-$349, or even a 3060 Ti at that same price point.

What I think may surprise people is that AMD will either tie or actually take the lead in power efficiency. It all depends on how aggressive AMD is with the voltage and frequencies.

What's funny is this will more or less be an exact repeat of last generation more or less. NVIDIA all by itself at the top with its largest chip and AMD competing nicely at the tier below it.

6700 XT is by far, not the top end SKU .

 

[Glo.]

Alleged AMD Radeon RX 6000 engineering sample spotted - VideoCardz.com

A photograph showing an engineering sample of the alleged RDNA2 graphics card has been published online. Is this Big Navi? A picture of a mysterious engineering sample has emerged on Bilibili. It is not possible to determine which graphics processor does the graphics card has, but one could...

videocardz.com

256 Bit GDDR6 is not for Big Navi.

N21 has 384 bit GDDR6 bus, N22 has 256 Bit GDDR6 bus.

So this GPU is the Navi 22.

 

[Kenmitch]

6700 XT is by far, not the top end SKU .

Chug-a-chug a choo-choo!....Glo says all aboard?

 

[blckgrffn]

Chug-a-chug a choo-choo!....Glo says all aboard?

Thomas_the_Train has entered the chat...

 

[Glo.]

Thomas_the_Train has entered the chat...

Im sorry...

 

[Konan]

Alleged AMD Radeon RX 6000 engineering sample spotted - VideoCardz.com

A photograph showing an engineering sample of the alleged RDNA2 graphics card has been published online. Is this Big Navi? A picture of a mysterious engineering sample has emerged on Bilibili. It is not possible to determine which graphics processor does the graphics card has, but one could...

videocardz.com

256 Bit GDDR6 is not for Big Navi.

N21 has 384 bit GDDR6 bus, N22 has 256 Bit GDDR6 bus.

So this GPU is the Navi 22.

I agree with you but KopiteKimi is saying N21

Well apparently there are 4 variants
Navi21 XTX Navi21 XT Navi21 XL Navi21 XE

https://twitter.com/x/status/1303890583123968000

 

[TESKATLIPOKA]

Short memory?

Nvidia teases an "Ultimate Countdown" a day before countdown was officially started (Aug 11, 2020)

https://twitter.com/x/status/1292808003897311232

AMD just did exactly the same thing, they just used a different form.

Thank you for your concern, but I don't have short memory, I just didn't see that, what I saw was only the official countdown.
BTW I don't see anything in that supernova clip or It's label which can imply something else than a countdown unlike the tweet from Frank Azor, but If you can good for you.

 

[TESKATLIPOKA]

Alleged AMD Radeon RX 6000 engineering sample spotted - VideoCardz.com

A photograph showing an engineering sample of the alleged RDNA2 graphics card has been published online. Is this Big Navi? A picture of a mysterious engineering sample has emerged on Bilibili. It is not possible to determine which graphics processor does the graphics card has, but one could...

videocardz.com

256 Bit GDDR6 is not for Big Navi.

N21 has 384 bit GDDR6 bus, N22 has 256 Bit GDDR6 bus.

So this GPU is the Navi 22.

That card is quite long. Great knowing that we can finally mount a CPU cooler on the GPU.

 

[leoneazzurro]

It is clearly an experimental card, not even a production sample. That is why there is the CPU cooler

 

[DisEnchantment]

It is a dev board used for bringup.
It is the AMD equivalent of this this Intel thing


In comparison a Vega dev board

 

[moinmoin]

I'd be disappointed if AMD doesn't have any card that clearly surpasses the 3080. The range between 3080 and 3090 is where the big margins are, so that's where AMD has to aim performance wise if it wants to make RTG more profitable.

 

[uzzi38]

Hehe

 

[leoneazzurro]

View attachment 29545

Hehe

It is quite obvious it is not BigNavi, but more probably N22

 

[kurosaki]

Just conservatively extrapolated values looks good enough for me. It's just a matter of price i guess.

 

[Gideon]

Just conservatively extrapolated values looks good enough for me. It's just a matter of price i guess.
View attachment 29546

You are never going to see 100% scaling. Even 2060 to 2080Ti was 87% scaling (at 4K) and only 75% scaling (at 1440p). That 10% IPC gain would have to be ~25% in practice (even more as your benches are 1440p).

EDIT:
Or if you don't like that example take 5700 XT vs 5500 XT. 40CUs vs 22CUs (with slightly slower clocks, so very near 2x), double the bandwidth, pretty much everything doubled up. Result? 83% faster in 1440p, 86% faster in 4k.

TL;DR you should either bump the IPC gain to 20% or subtract at least ~10% from the results to get a more accurate result at the same clocks

 

[leoneazzurro]

Just conservatively extrapolated values looks good enough for me. It's just a matter of price i guess.
View attachment 29546

Yep at 1850Mhz clock and same scaling as seen in RDNA cards, it should be in average 30% to 40% faster than 2080Ti, and at 2200MHz from 50% to 60-65% faster than 2080Ti, assuming the 384bit bus

 

[leoneazzurro]

You are never going to see 100% scaling. Even 2060 to 2080Ti was 87% scaling (at 4K) and only 75% scaling (at 1440p). That 10% IPC gain would have to be ~25% in practice (even more as your benches are 1440p).

These values were not assuming 100% scaling btw, they are extrapolated by looking at the lower cards' scaling with CU, bandwidth, and clock. Generally the scaling is slightly above 90% of the shading power, assuming other parts of the chip don't introduce severe bottlenecks.

 

[Gideon]

These values were not assuming 100% scaling btw, they are extrapolated by looking at the lower cards' scaling with CU, bandwidth, and clock. Generally the scaling is slightly above 90% of the shading power, assuming other parts of the chip don't introduce severe bottlenecks.

Slightly above 90% at 1440p?

Certainly not in the listed games (5500XT 8GB vs 5700XT)

Division 2 (1440p ultra): 65.3 / 38.4 = 70%
GTA V (1440p very high): 82.2 / 45.3 = 81%
Strange Brigade (1440p ultra): 118.4 / 65.1 = 82%
Metro Exodus(1440p ultra): 58.4 / 31.7 = 84%

IMO 5500XT and 5700XT are a very good comparison and as it's almost the same 2x CU count increase. On top of that 5700XT has 2x the memory bandwidth of the 5500XT, something many people seem to think won't be replicated for 80CU Big Navi (if it has 384 bit membus and no HBM). If it only has ~150% the memory bandwidth the scaling will surely be (at least a little bit) worse.

I want Big Navi to succeed as many others here, but most of the hyped performance gains ... seem too rosy. I can still see 2x scaling, but that would IMO need considerably higher clocks and HBM2E (or a really wide GDDR6 bus).

 

[kurosaki]

You are never going to see 100% scaling. Even 2060 to 2080Ti was 87% scaling (at 4K) and only 75% scaling (at 1440p). That 10% IPC gain would have to be ~25% in practice (even more as your benches are 1440p).

EDIT:
Or if you don't like that example take 5700 XT vs 5500 XT. 40CUs vs 22CUs (with slightly slower clocks, so very near 2x), double the bandwidth, pretty much everything doubled up. Result? 83% faster in 1440p, 86% faster in 4k.

TL;DR you should either bump the IPC gain to 20% or subtract at least ~10% from the results to get a more accurate result at the same clocks

No this isn't 100% spot on, but my guess is we should enter this ballpark, with margins in each direction of 10-20%(?)
It's good enough for me (1440p setup) as long as the price is ok. I'm getting tired of low end cards costing over 300usd.

 

[Gideon]

No this isn't 100% spot on, but my guess is we should enter this ballpark, with margins in each direction of 10-20%(?)
It's good enough for me (1440p setup) as long as the price is ok. I'm getting tired of low end cards costing over 300usd.

Agreed on all points.

These values were not assuming 100% scaling btw, they are extrapolated by looking at the lower cards' scaling with CU, bandwidth, and clock.

I presumed 100% scaling cause that's what basic math leads to:

58.4 * 2.1 = 122.64,
118.4 * 2.1 = 248.64,
etc ...

That's exactly 2x scaling + 10% added on top from IPC, as it was excplicitly stated that this estimates 0% improvement in clocks.

All I'm saying is that these result are not realistic for the given input data (10% IPC gain no clock-gain) especially at 1440p.

 

[leoneazzurro]

Slightly above 90% at 1440p?

Certainly not in the listed games (5500XT 8GB vs 5700XT)

Division 2 (1440p ultra): 65.3 / 38.4 = 70%
GTA V (1440p very high): 82.2 / 45.3 = 81%
Strange Brigade (1440p ultra): 118.4 / 65.1 = 82%
Metro Exodus(1440p ultra): 58.4 / 31.7 = 84%

IMO 5500XT and 5700XT are a very good comparison and as it's almost the same 2x CU count increase. On top of that 5700XT has 2x the memory bandwidth of the 5500XT, something many people seem to think won't be replicated for 80CU Big Navi (if it has 384 bit membus and no HBM). If it only has ~150% the memory bandwidth the scaling will surely be (at least a little bit) worse.

I want Big Navi to succeed as many others here, but most of the hyped performance gains ... seem too rosy. I can still see 2x scaling, but that would IMO need considerably higher clocks and HBM2E (or a really wide GDDR6 bus).

I used Techpowerup numbers, on a variety of titles, looking at numbers from their latest 5600XT and 5500Xt reviews (latest drivers and same test setup). Also "almost 2x" is not "2x" because 40/22=1.8 and not 2. I.e considering your numbers before, in Metro Exodus you have 0.84*1.11=93% scaling with shading resource and so on. 5600XT also is a demonstration that bandwidth is less an issue for RDNA1 cards (and scaling seems to be very similar going from 1080p to 4K) . Shading seems to be still the most important factor at these resolutions. Oh, I have to say, I used a more complex model, with 3 factors of scaling coming by Shading, bandwidth and ROP ratio

 

[kurosaki]

Agreed on all points.


I presumed 100% scaling cause that's what basic math leads to:

58.4 * 2.1 = 122.64,
118.4 * 2.1 = 248.64,
etc ...

That's exactly 2x scaling + 10% added on top from IPC, as it was excplicitly stated that this estimates 0% improvement in clocks.

All I'm saying is that these result are not realistic for the given input data (10% IPC gain no clock-gain) especially at 1440p.

Exactly, but I also believe that they have worked with the arch more than my figures show. Reducing the need for memory bandwidth, memory compression, Larger caches, a more effective arch, increase of clockspeed, more cores per CU as in Ampere? Who knows, the amount of stuff they might have done this time, makes the figures very uncertain. That's why I just picked a total of 10% in improvements, could easily have been 5%clocks, 5% IPC, and doubled it. It might perform in your worst case, like 1.8*5700XT, still, it would be kind of good enough for me.
I don't think it's going to perform any worse anyways. Then they will have to aim at competing in another price segment.

 

[Gideon]

Also "almost 2x" is not "2x" because 40/22=1.8 and not 2

That's right, it's a 10% difference in CUs and only a 2% difference in game clock, so yeah when taking that into account it's indeed usually over 90%

5600XT also is a demonstration that bandwidth is less an issue for RDNA1 cards

This is actually a very good data-point I've somehow missed (despite people hinting at it numerous times before)

5600XT has 288.0 GB/s of bandwidth vs 448.0 GB/s for 5700 (55% more) Yet 5700 is only 17% faster even at 4K which corresponds almost exactly to the 18% game clock difference.

Or from the same review. 5600 XT 12 Gbps vs 14 Gbps memory 17% memory bandwidth advantage leads to only 6% performance uplift at 4K.

In light of this data I'm feeling more hopeful about N21 even if it tops out at 384 bit GDDR6, gotta pump those clocks up!

 

[Glo.]

5600 Xt has 288 GB/s bandwdith only with 12 Gbps GDDR6(old BIOS). New BIOSes bumped that to 14 Gbps, and 336 GB/s bandwidth.

 

[Gideon]

5600 Xt has 288 GB/s bandwdith only with 12 Gbps GDDR6(old BIOS). New BIOSes bumped that to 14 Gbps, and 336 GB/s bandwidth.

The review I based my calculations on has both 14Gbps and 12Gbps results:

ASUS Radeon RX 5600 XT TUF EVO Review

The ASUS Radeon RX 5600 XT TUF EVO uses the improved version of the EVO cooler. Temperatures and noise levels are excellent because ASUS decided not to downgrade the heatsink for their RX 5600 XT despite the much lower heat output of the GPU. Idle fan stop and a backplate are included, too.

www.techpowerup.com

16.6% more bandwidth adds roughly 6% more performance at 4K. If memory bandwidth scaled linearly it would be 18% more performance from 50% more bandwidth (say 512 bit vs 384 bit), but in reality it should run into diminishing returns pretty quickly, so it would be even less.