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]

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.

Also take in account that base clock is different but game clock (boost) is almost equal between the 5600Xt and 5700. So it's not that the bandwidth is not important anymore, but only it is not the most limiting factor, even at 4K (the clear proof are the results of 5700XT vs 5500XT and 5700)

PS:in Assassin's Creed Odissey the 5600 XT is frame buffer limited at 4K

 

[Shivansps]

I thought the console APU's are low margin, very little profit, in it for the R&D dollars, offerings.....At least that's the way it went on the forum when Nvidia was arsed out of it. Ignoring their week SOC part of it of course.

We'll have to wait and see what AMD drops and then decide. I'm pretty sure you can't predict the price point of their dGPU's based on the cost of a console.

No idea what a 7nm wafer costs in the end. Most end users don't factor it in when making a buying decision.

I dont know, what i do know is right now AMD current gen CPU/GPU prices are crazy high compared to a whole console and that is not right as even desktop APU prices are going up.

So want to see AMD prices in desktop to drop due to this.

BTW: I wonder about the $299 Xbox S Hardware, diferent die? defect dies?

 

[Gideon]

I dont know, what i do know is right now AMD current gen CPU/GPU prices are crazy high compared to a whole console and that is not right as even desktop APU prices are going up.

So want to see AMD prices in desktop to drop due to this.

The cards between 200-300€ (that used to be the sweetspot long ago) are especially bad value right now. Hopefully this changes with the upcoming generation

 

[Timorous]

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

The 5500XT vs 5700XT is interesting because both have the same perf/watt @ 1080p (where the 128 bit bus is not a bottleneck). With AMD saying RDNA 2 has a 50% perf/watt increase over RDNA that would put the 5700XT performance at around 140W. With similar perf/watt scaling as RDNA that would put a 300W 'Big Navi' card at just over double that level of performance.

 

[reb0rn]

So far Big Navi info is that it will have 16 GB, 16 chip... guru3d say 256bit
I don`t see logic to be under 384bit and if its 16GB then it must be 512bit memory?

 

[TESKATLIPOKA]

80Cu won't have only 256bit memory controller, that's simply not enough to feed It, at least 384bit is a must.

 

[reb0rn]

But for 384bit you can't have 16GB, you can have 12GB

16chip is 256 or 512bit only

 

[JujuFish]

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.

I don't have a dog in this race, but 2x scaling and 10% IPC would be 2.2, not 2.1. Multiplicative, not additive.

 

[eek2121]

The 5500XT vs 5700XT is interesting because both have the same perf/watt @ 1080p (where the 128 bit bus is not a bottleneck). With AMD saying RDNA 2 has a 50% perf/watt increase over RDNA that would put the 5700XT performance at around 140W. With similar perf/watt scaling as RDNA that would put a 300W 'Big Navi' card at just over double that level of performance.

In theory, yes, however, Navi2X is a different beast. There is an IPC increase and the CUs are different from RDNA1. There are also quite a few unknowns that haven’t been revealed yet. A 40CU RDNA2 product is faster than a 40CU RDNA1 product at the same clock. Since things like VRS are supported, it will be significantly faster in some cases. Then on top of that, the clock ceiling is much higher.

I expect the RX 6600 XT will have around 40CUs and it will land somewhere between the 2080S and 2080ti performance wise. The rest of the products will scale up from there.

EDIT: a lot of people here and elsewhere have been claiming that there is ‘no way’ AMD can match NVIDIA’s RTX 3080. I don’t understand this logic at all. We know from the consoles that AMD hit their 50% perf/watt improvement (or at least came very close). I think AMD will be close, and they will manage to do it at a lower TDP. If you have owned a high end GPU you can appreciate why heat is a big issue and why NVIDIA went off the deep end this gen. The only component I hear in my PC when gaming is my GPU.

 

[kurosaki]

I don't have a dog in this race, but 2x scaling and 10% IPC would be 2.2, not 2.1. Multiplicative, not additive.

Very true, it came to me afterwards, but didnt care to adjust. A couple of more percent, now its for 5% IPC increase. But as said, ballpark figures..

 

[TESKATLIPOKA]

But for 384bit you can't have 16GB, you can have 12GB

16chip is 256 or 512bit only

Or 24GB for 384bit. It would be an overkill, but 12GB is not a lot.

 

[Saylick]

Or 24GB for 384bit. It would be an overkill, but 12GB is not a lot.

Better than 10GB.

 

[anthrax]

I would guess the 80CU navi will have something between 384bit or wider memory bus. I would guess 448bit for 14GB. Unless AMD is also getting access to GDDR6x

 

[maddie]

In theory, yes, however, Navi2X is a different beast. There is an IPC increase and the CUs are different from RDNA1. There are also quite a few unknowns that haven’t been revealed yet. A 40CU RDNA2 product is faster than a 40CU RDNA1 product at the same clock. Since things like VRS are supported, it will be significantly faster in some cases. Then on top of that, the clock ceiling is much higher.

I expect the RX 6600 XT will have around 40CUs and it will land somewhere between the 2080S and 2080ti performance wise. The rest of the products will scale up from there.

EDIT: a lot of people here and elsewhere have been claiming that there is ‘no way’ AMD can match NVIDIA’s RTX 3080. I don’t understand this logic at all. We know from the consoles that AMD hit their 50% perf/watt improvement (or at least came very close). I think AMD will be close, and they will manage to do it at a lower TDP. If you have owned a high end GPU you can appreciate why heat is a big issue and why NVIDIA went off the deep end this gen. The only component I hear in my PC when gaming is my GPU.

Also, does Navi 1x have a dual issue command processor?

 

[Stuka87]

Or 24GB for 384bit. It would be an overkill, but 12GB is not a lot.

If its HBM though, it should be 512bit?

 

[kurosaki]

If its HBM though, it should be 512bit?

Or 1024... 😁 Nah, it's gddr 6 this time around

 

[Saylick]

Also, does Navi 1x have a dual issue command processor?

My understanding is that RDNA 1's command processor was a single issue pipeline, with RDNA 2 being a dual-issue.

This is confirmed in the Xbox Series X GPU diagram as indicated by the dual input streams entering the command processor, as follows:
"Dual stream multi-core command processor"

Also, someone posted a video which mentioned this earlier but RDNA 2 apparently can issue 7 ops/clock per CU vs. the 4 ops/clock per SIMD (typically 2-4) that RDNA 1 could do. Refer to Page 9 of the RDNA 1 Whitepaper, which states "Each of the four SIMDs can request instructions every cycle and the instruction cache can deliver 32B (typically 2-4 instructions) every clock to each of the SIMDs – roughly 4X greater bandwidth than GCN."

In comparison, RDNA 2 issues 7 ops/clock:

So RDNA 1 = 2-4 ops/clock per SIMD, and there's (2) SIMDs per CU, so ~4 - 8 ops per CU/clock. RDNA 2 = 7 ops/clock per CU.

About comparable it seems.

 

[kurosaki]

My understanding is that RDNA 1's command processor was a single issue pipeline, with RDNA 2 being a dual-issue.

This is confirmed in the Xbox Series X GPU diagram as indicated by the dual input streams entering the command processor, as follows:
"Dual stream multi-core command processor"

Also, someone posted a video which mentioned this earlier but RDNA 2 apparently can issue 7 ops/clock vs. the 4 ops/clock max (typically 2-4) that RDNA 1 could do. Refer to Page 9 of the RDNA 1 Whitepaper, which states "Each of the four SIMDs can request instructions every cycle and the instruction cache can deliver 32B (typically 2-4 instructions) every clock to each of the SIMDs – roughly 4X greater bandwidth than GCN."

In comparison, RDNA 2 issues 7 ops/clock:

That's a heck of an improvement. Almost similar to the Amperes doubling of shaders.

 

[theeedude]

Only 2 of those 7 are vector ALU tho, the rest are data/control/scalar.

 

[Saylick]

That's a heck of an improvement. Almost similar to the Amperes doubling of shaders.

Only 2 of those 7 are vector ALU tho, the rest are data/control/scalar.

FYI, I made a last minute edit as I was parsing through the info from the whitepaper and the Xbox Series X presentation:

So RDNA 1 = 2-4 ops/clock per SIMD, and there's (2) SIMDs per CU, so ~4 - 8 ops per CU/clock. RDNA 2 = 7 ops/clock per CU.
About comparable it seems.

 

[maddie]

FYI, I made a last minute edit as I was parsing through the info from the whitepaper and the Xbox Series X presentation:

It depends.

2-4 ops/clk x 2 is 4-8 ops/CU (not 6-8). IF the average is 6 then 7 ops/clk is a utilization improvement.

 

[Saylick]

It depends.

2-4 ops/clk x 2 is 4-8 ops/CU (not 6-8). IF the average is 6 then 7 ops/clk is a utilization improvement.

Oops, brain fart. You're right!

 

[Kenmitch]

I'm not in the market to upgrade my RX 5700, but I do look forward to seeing the launch presentation and reviews when they drop.

 

[dr1337]

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.

kopite says its n21. If true that means it has to be a 512 bit bus with 16gb of vram. tbh 384 bit isnt enough for 80cu. This has to either be n21 at 512 bit or its n23 at 256 bit.

 

[Helis4life]

kopite says its n21. If true that means it has to be a 512 bit bus with 16gb of vram. tbh 384 bit isnt enough for 80cu. This has to either be n21 at 512 bit or its n23 at 256 bit.

He could be wrong, could be disinformation, could be very cut down n21, or n21 might not be the 80cu top die

Xbox has 52cu with 320bit, ps5 has 36 with 384. AMD is not releasing an 80cu with a 256 bus.


Also even though the sticker says 16gb, we can't verify that. What's to say this card should have 16 1gb chips, dual sided, but instead it has 8 single sided only for testing purposes