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

 

[ModEl4]

https://tpucdn.com/review/nvidia-geforce-rtx-3070-founders-edition/images/relative-performance_3840-2160.png

Taking this chart as base i would say 6000 lineup will perform more or less like this (just for fun prediction, we will see):

RX 6500XT 40-45%
RX 6600XT 65-70%
RX 6700 70-75%
RX 6700XT 80-85% (40CU)
RX 6800 110-115%
RX 6800XT 130-135%
RX 6900XT 140-145%

6700XT at 2.5GHz will be around 86% of 3070 imo so your assumption was spot on.
The 32CU part at 2.3GHz is going to be only + 6-7% from a 5700 imo, so 2070S will be 13-14% faster, no 2080 lofty goals...

 

[Glo.]

6700XT at 2.5GHz will be around 86% of 3070 imo so your assumption was spot on.
The 32CU part at 2.3GHz is going to be only +6% from a 5700 imo, so 2070S will be 14% faster, no 2080 lofty goals...

So suddenly RDNA2 on lower CU parts will lose IPC compared to big Navi die?

 

[exquisitechar]

6700XT at 2.5GHz will be around 86% of 3070 imo so your assumption was spot on.
The 32CU part at 2.3GHz is going to be only +6% from a 5700 imo, so 2070S will be 14% faster, no 2080 lofty goals...

Eh, 5700XT does really bad at 4k. Shouldn't be used to predict 6700XT's performance (it will be a 1440p card anyway).

 

[Veradun]

So suddenly RDNA2 on lower CU parts will lose IPC compared to big Navi die?

This is the new "can't scale out"

 

[ModEl4]

So suddenly RDNA2 on lower CU parts will lose IPC compared to big Navi die?

How are they losing IPC?
In your opinion what is 40CU Navi 22 vs 80CU Navi 21 clock for clock (let's suppose they had the same clock) difference at 4K and what at QHD?
Reply with your estimated clock for clock difference and we will take it from there...

 

[Glo.]

How are they losing IPC?
In your opinion what is 40CU Navi 22 vs 80CU Navi 21 clock for clock (let's suppose they had the same clock) difference at 4K and what at QHD?
Reply with your estimated clock for clock difference and we will take it from there...

I already have posted accurate performance targets, for 80, and 60 CU dies . Months ago.

Performance target for 40 CU die is Between RTX 2080 Super and RTX 2080 Ti(10% above RTX 2080 Super), which makes it around 90-95% of RTX 3070 performance.

Performance target for 32 CU die, with 128 bit bus, altho this is my speculation, should be between RTX 2070 Super and standard, non-Super 2080. Closer to which one? We will see.

 

[Midwayman]

Why would u want a 8GB GPU at 2080Ti perf levels for the next few years for $499, when you pay a tiny bit extra, you get +15% more performance AND 16GB VRAM?

The 3070 is the biggest loser here. No wonder NV rushing out another cut GA102 for 3070Ti.

It's not a 'tiny bit' its 15.8% more. Now you might say "But it has 15% better performance!" yes, but it also has way worse RT, their equivalent of DLSS isn't supposed to be nearly as good and AMD has a 'not great' track record with drivers. Good trade-off for 8gb vs 16gb? Maybe. Its really a 1440p card and 8gb shouldn't be an issue for awhile at that resolution. Its hardly a straight up win at that price anyways. AMD will have to cut its price when the 3070ti launches.

 

[Veradun]

it also has way worse RT

Bench suite used to assess this?

 

[Midwayman]

Bench suite used to assess this?

According to leaks "Microsoft DXR SDK tool called ‘Procedural Geometry '"

Not a lot of info of course, but the difference does appear pretty substantial.

 

[Veradun]

According to leaks "Microsoft DXR SDK tool called ‘Procedural Geometry '"

Not a lot of info of course, but the difference does appear pretty substantial.

Another not a game then.

I wonder how much time will we have to wait for RT to be a real thing.

Hopefully for Hopper/RDNA3?

 

[dzoni2k2]

According to leaks "Microsoft DXR SDK tool called ‘Procedural Geometry '"

Not a lot of info of course, but the difference does appear pretty substantial.

Are we talking 6800XT or 6800? 6800XT RT is on par with 3070 or as fast as 2080Ti in that bench.

 

[ModEl4]

Navi 1 requires more bandwidth to scale with clocks. IPC, clock for clock, SM for SM, with sufficient bandwidth will give you 1-2% higher IPC on RDNA1 than Turing.

In general I believe that base Navi 22 die with 40 CUs/192 bit bus and 2.5 GHz should be on par with RTX 2080 Ti in performance. Pessimistic calculations suggest that this die will be around 15% faster than RTX 2080 Super. Optimistic - even up to 15% performance higher than RTX 2080 Ti, if the IPC really increased by 20%, over RDNA1.

It is pretty mental architecture, after all, regardless of the outcome.

I already have posted accurate performance targets, for 80, and 60 CU dies . Months ago.

Performance target for 40 CU die is Between RTX 2080 Super and RTX 2080 Ti(10% above RTX 2080 Super), which makes it around 90-95% of RTX 3070 performance.

Performance target for 32 CU die, with 128 bit bus, altho this is my speculation, should be between RTX 2070 Super and standard, non-Super 2080. Closer to which one? We will see.

Regarding your speculations, all I know is what you replied to a post I made before 3 weeks (I tried to explain that +15% 2080Ti performance for 2.5GHz Navi22 would be highly unlikely) and according to your reply, you thought that Navi22 will be at 2080Ti level, with the most pessimistic projections +15% 2080S, up to +15% 2080Ti, if IPC is +20% vs RDNA1!!!
I wonder what is the IPC difference between RDNA2 and RDNA1...

Navi 1 requires more bandwidth to scale with clocks. IPC, clock for clock, SM for SM, with sufficient bandwidth will give you 1-2% higher IPC on RDNA1 than Turing.



In general I believe that base Navi 22 die with 40 CUs/192 bit bus and 2.5 GHz should be on par with RTX 2080 Ti in performance. Pessimistic calculations suggest that this die will be around 15% faster than RTX 2080 Super. Optimistic - even up to 15% performance higher than RTX 2080 Ti, if the IPC really increased by 20%, over RDNA1.



It is pretty mental architecture, after all, regardless of the outcome.

You didn't reply at all regarding the clock for clock Navi21/22 differences.
Anyway we will see indeed in a few months.
My predictions is at 4K for 3070 comparison vs Navi22 and at QHD for 5700 vs 32CU part.

 

[leoneazzurro]

With a clock of 2450Mhz, a Navi22 should be around 90-95% of a 2080Ti, in average, in rasterization.

 

[DiogoDX]

Some entertainment for those on the hype-train.

Not much to look at as far as RT goes. Game does look ok if your into the series.

They chose to do shadows that is the worst RT effect to showcase IQ. They should do GI, diffuse lighting or reflections.

 

[alexruiz]

It's not a 'tiny bit' its 15.8% more. Now you might say "But it has 15% better performance!" yes, but it also has way worse RT, their equivalent of DLSS isn't supposed to be nearly as good and AMD has a 'not great' track record with drivers. Good trade-off for 8gb vs 16gb? Maybe. Its really a 1440p card and 8gb shouldn't be an issue for awhile at that resolution. Its hardly a straight up win at that price anyways. AMD will have to cut its price when the 3070ti launches.

So RT and DLSS are now the most critical features?
If so, you can go and by the RTX 3070. Some other of us will get a Navi and "put up" with the "not so great drivers"

AMD won't cut the price, they really don't want to sell this one.
The price was set that high to push people to the RX 6800 XT.

 

[Glo.]

Regarding your speculations, all I know is what you replied to a post I made before 3 weeks (I tried to explain that +15% 2080Ti performance for 2.5GHz Navi22 would be highly unlikely) and according to your reply, you thought that Navi22 will be at 2080Ti level, with the most pessimistic projections +15% 2080S, up to +15% 2080Ti, if IPC is +20% vs RDNA1!!!
I wonder what is the IPC difference between RDNA2 and RDNA1...

You didn't reply at all regarding the clock for clock Navi21/22 differences.
Anyway we will see indeed in a few months.
My predictions is at 4K for 3070 comparison vs Navi22 and at QHD for 5700 vs 32CU part.

The point is, if your calculation is correct, from this post:

6700XT at 2.5GHz will be around 86% of 3070 imo so your assumption was spot on.
The 32CU part at 2.3GHz is going to be only + 6-7% from a 5700 imo, so 2070S will be 13-14% faster, no 2080 lofty goals...

... then 40CU die has lost IPC compared to not even RDNA1, but also Navi 21.

Performance targets have been already discussed.

 

[moinmoin]

They chose to do shadows that is the worst RT effect to showcase IQ. They should do GI, diffuse lighting or reflections.

Expect more such picks in the future as soon as consoles are involved, you won't see RT effects being used by default that tank the FPS.

 

[kurosaki]

Another not a game then.

I wonder how much time will we have to wait for RT to be a real thing.

Hopefully for Hopper/RDNA3?

I believe Cyberpunk 2077. Whichever card has the best performance in that game will get sold fast...

 

[Shivansps]

If it cant be 52CU N21 then ill have to agree that the 6700XT will be 40CU, 48CU is a lot to cut from the N21 die.

Im trying to see what they could do here:

6900XT: 80CU
6800XT: 72CU
6800: 60CU
6700XT: 40CU/12GB (N22)
6700: 32CU/12GB (N22)
6600XT: 32CU/8GB (N23)
6500XT: 24CU/8GB (N23)
6500: 20CU/4GB (N23)

This seems about right?

 

[Leadbox]

What chance that there's some built in redundancy here, that this is actually an 88 or 96 CU chip? Something to meet the Ti card that nvidia will surely drop in the coming months.

 

[Hitman928]

If it cant be 52CU N21 then ill have to agree that the 6700XT will be 40CU, 48CU is a lot to cut from the N21 die.

Im trying to see what they could do here:

6900XT: 80CU
6800XT: 72CU
6800: 60CU
6700XT: 40CU/12GB (N22)
6700: 32CU/12GB (N22)
6600XT: 32CU/8GB (N23)
6500XT: 24CU/8GB (N23)
6500: 20CU/4GB (N23)

This seems about right?

I would expect 36 CU for the 6700 just like the 5700.

 

[kurosaki]

What chance that there's some built in redundancy here, that this is actually an 88 or 96 CU chip? Something to meet the Ti card that nvidia will surely drop in the coming months.

Haha! Nvidia maxed it out this round!

They cant go higher on core count or watts, they pushed it with the 3090, with the 6900XT trades blows with. Nvidia wont release a 3090 Ti with 10 more shaders or what not. not until a new arch and process node.

 

[HurleyBird]

What chance that there's some built in redundancy here, that this is actually an 88 or 96 CU chip? Something to meet the Ti card that nvidia will surely drop in the coming months.

I think the chances are pretty low that there are hidden CUs, but pretty high that AMD can match a Titan if they want to. Theoretically, AMD could throw on a triple slot cooler, upgrade the memory system to 32 GB GDDR6X (assuming there isn't some hidden HBM controller), bin the heck of it so it clocks around ~2.5GHz power usage be damned. Add/unlock various professional features, and maybe add some extra display outputs.

 

[kurosaki]

I would expect 36 CU for the 6700 just like the 5700.

I'm a bit saddened that the price point surely will stay the same. still no clear upgrade path for the guy who bought a card 2 years ago.

 

[blckgrffn]

I believe Cyberpunk 2077. Whichever card has the best performance in that game will get sold fast...

False. Whichever card you can actually buy that runs Cyberpunk 2077 acceptably will sell lots 😂

Seriously though, that game better be great as it seems to be on a very high pedestal. I am not even considering it until it is cheap af because there are so many great games in my backlog. The way the launch is going I am expecting major patches after release as well.

And I just keep playing BL3 and SC2. Wake me up when BL3 has RT effects. I adore it’s HDR.

(I don’t think they are coming as it’s advertised as 60 FPS/4K on the X and that can’t happen with RT anything turned on without some other big compromises)