AMD Navi 23 ‘NVIDIA Killer’ GPU Rumored to Support Hardware Ray Tracing, Coming Next Year

[humbug]

AMD have no intention to react to Nvidia, it was clear before Ampere AMD had no intention to get in before or around the same time as Nvidia and they are still in no rush, just saying "we are on our own release schedule"

I don't know what to make of that, since the Ampere kitchen reveal the 80 CU Navi seems to be no more.

The 2080TI is 35% faster than a 5700XT at 1440P, i'm not going to use 4K for this given that every 8GB card falls further behind the 2080TI, they are not 4K cards.
But at 4K compared to the 3080 it is 32% faster than the 2080TI.

2080TI vs 5700XT 135%
3080 vs 2080TI = 132% (@4K)

Not so Big Navi is rumoured to have 60 CU's, that's 50% more than the 5700XT, after scaling make that +45%

5700XT +45% = 145%

145 / 135 = 1.074. so normalized: 60 CU Not so Big Navi vs 2080TI = 107%

3080 vs 60 CU Not so Big Navi: 132 / 107 = 1.23 (+23%)

RDNA1 vs RDNA2 +10% IPC = 60 CU RDNA2 @1900Mhz. 3080 vs 60 CU RDNA2 @ 1900Mhz: 132 / 117 = 1.128 (+13%)

The PS5 clocks to 2.23Ghz, let's assume that's the limit of RDNA2: 2230 / 1900 = 1.173 (+17%)

Now the 6800XT is 34% faster than a 2080TI, 2% faster than the 3080.

Moores Law thinks AMD think they can compete with the 3080 without the 80 CU Big Navi. i think he's right. With the enhancements of the improved 7nm node and improvements to the RDNA architecture AMD could make a still small 350mm2 GPU at around <2.3Ghz, 250 Watts, 16GB GDDR6 and at least trade blows with the 3080 at £150 less and still get good margins.

I think its a shame if AMD don't slam an 80 CU RDNA2 GPU in Nvidia's face anyway but i think AMD were expecting more from Ampere, they are not impressed.

https://www.techspot.com/review/2099-geforce-rtx-3080/

 

[LePhuronn]

humbug, you never factor in the 50% performance per watt AMD are targeting over RDNA 1 (10% IPC increase alone won't do a 50% uplift, surely). Do you simply not believe it'll happen? Do a direct replacement of the 5700XT with that uplift: 40 CUs at 225W and the same clock speed goes about 10% over the 2080 Ti. 40 RDNA 2 CUs then would drop in pretty much mid way between the 3070 and 3080. Now do a 45% scaling for 60 CUs and see what happens.

...but i think AMD were expecting more from Ampere, they are not impressed.

Which makes it the perfect opportunity to wheel out 80 CUs as a limited run halo and smack Nvidia around the chops with it. Drop 32GB HBM2e on it, push it to 350W, sell it for £1,300. Hell, take money from RTG's marketing budget to support it, such a halo product would be a pure marketing exercise.

 

[gpuerrilla]

Positive thoughts. It cant really get any worse than the 3080 paper release, can it?

 

[humbug]

humbug, you never factor in the 50% performance per watt AMD are targeting over RDNA 1 (10% IPC increase alone won't do a 50% uplift, surely). Do you simply not believe it'll happen? Do a direct replacement of the 5700XT with that uplift: 40 CUs at 225W and the same clock speed goes about 10% over the 2080 Ti. 40 RDNA 2 CUs then would drop in pretty much mid way between the 3070 and 3080. Now do a 45% scaling for 60 CUs and see what happens.


Which makes it the perfect opportunity to wheel out 80 CUs as a limited run halo and smack Nvidia around the chops with it. Drop 32GB HBM2e on it, push it to 350W, sell it for £1,300. Hell, take money from RTG's marketing budget to support it, such a halo product would be a pure marketing exercise.

With +50% PPW a 60 CU RDNA2 would be <225 Watts, i think AMD will and will have to push the card to around 2.2 - 2.3Ghz to get it at the 3080 level, that's about a 15 to 20% clock difference vs RDNA1, the 225 Watts is board power and includes the GDDR6, the GPU is between 180 and 200 Watts, the clock difference will push the board power on a 6800XT to about 250 Watts.

 

[humbug]

Which makes it the perfect opportunity to wheel out 80 CUs as a limited run halo and smack Nvidia around the chops with it. Drop 32GB HBM2e on it, push it to 350W, sell it for £1,300. Hell, take money from RTG's marketing budget to support it, such a halo product would be a pure marketing exercise.

Maybe they will, at a later date when they are not so wafer constrained.

 

[pugheaven]

Maybe they will, at a later date when they are not so wafer constrained.

Yep and it would be fantastic to see something getting to "possible" 3090 levels, as I think nVidia think they're 100% safe.... with the 3090 so they can at least say we're still No.1... be nice to actually say, hold on a minute... check this out... we shall see soon!

 

[LePhuronn]

With +50% PPW a 60 CU RDNA2 would be <225 Watts,

I always just take the 5700XT as the reference: keep 40 CUs, keep 225W total board power, add 50% performance. Plenty of leeway in napkin maths then to say "well, we can slap another 50W onto this".

Either which way, even for educated guesswork the numbers paint a very tantilising picture.

Maybe they will, at a later date when they are not so wafer constrained.

Are they wafer constrained? Has anybody said such a thing? I know Lisa Su said a while back that having so much on TSMC 7nm was a challenge, but I've not seen anything that suggests there's issues. TSMC have different 7nm nodes, AMD aren't building everything on just one of them, surely?

 

[Chuk_Chuk]

no, just chuck in a one per customer rule and some sort of captcha. If stock is low, tell people and let people preorder for later deliveries earlier than when they go live.

On their own site, a week before going live, a randomised lottery system could work (like festival tickets sometimes do) but that would be a high expectation.

How can AMD force other websites to do that though? One per customer, maybe they can convince retailers to do that. But i can't see, for example, the rainforest implementing a captcha system because AMD told them to.

 

[Lokken86]

Lisa Su - "I would love it if we beat them. LOVE it!"

 

[humbug]

There is also the possibility that all these leaks are deliberately engineered by AMD, misdirection, keep people talking about it.

Its frustrating that AMD are taking such a blasé approach to this, i get that AMD are trying to learn lessons and not rush it, but 2 months after Nvidia's reveal for what might be AMD's reveal, we hope its the paper launch with availability a couple of weeks later and not just "RDNA2 is a thing, see you in another 6 weeks for a paper launch"
With your competition pushing hard you can't afford to be this relaxed, another mistake AMD need to learn is that they are always too late to the party, by the time AMD eventually get there everyone has spent their money and got what they wanted while Nvidia start the marketing cycle for the next batch of releases.

Nvidia are here now, why should i wait 3 months for AMD?

I always just take the 5700XT as the reference: keep 40 CUs, keep 225W total board power, add 50% performance. Plenty of leeway in napkin maths then to say "well, we can slap another 50W onto this".

Either which way, even for educated guesswork the numbers paint a very tantilising picture.

Are they wafer constrained? Has anybody said such a thing? I know Lisa Su said a while back that having so much on TSMC 7nm was a challenge, but I've not seen anything that suggests there's issues. TSMC have different 7nm nodes, AMD aren't building everything on just one of them, surely?

Slap another 50 Watts on to this... i think that's where you are looking at bigger GPU's tho, i think that's what you're saying, sure AMD have plenty of headroom in the power envelope.

I'm assuming they are Wafer constrained. Zen 3 probably comes from the same wafer supply, and they will sell an absolute cluster-crisis load of them, it might not be a wafer supply issues in the sense of enough wafers total, AMD own pretty much all of TSMC's 7nm but that they just can't make them fast enough to keep up with the immediate short term demand, perhaps later when the demand settles down a bit there might be room to make the large 80 CU GPUs dies and that not resulting in starving the market from Zen 3 CPU's and the smaller Navi cards. these Big Navi dies would take up a lot of wafer space compared with the smaller ones and especially Zen 3 CPU's, 4 to 6 times more wafer space.

 

[queamin]

Would you buy a car with 3 wheels or wait 6 weeks for the other wheel.

EDIT:- and then moan it doesn't drive well.

 

[Ayahuasca]

Nvidia are here now, why should i wait 3 months for AMD?

If you haven't already ordered then it could take 4-6 weeks for an nVidia card anyway. I would hazard a guess you would be more interested in the 3070 and that could be even longer.

 

[g67575]

I think its a good idea to lower expectations. We have no reason at all to believe that the RX 6000 series will have 80 CUs, AMD has not confirmed it. 64 (same as vega 64) is probably a given, as well as similar clocks to a PS5 GPU. We dont know if AMD can / will produce a large 500-600mm2 die.

The only thing we know is that 7nm EUV and 5nm fab. processes are planned for CPUs and probably GPUs next year.

question - Can I ride it out with a R9 390 at 1080p for a bit longer? Are there any games that require a better GPU as a minimum?

Thoughts for an eventual large die Navi 2x name - maybe it will just be called Navi 80 or Navi 96 (depending on CU count)

 

[humbug]

If you haven't already ordered then it could take 4-6 weeks for an nVidia card anyway. I would hazard a guess you would be more interested in the 3070 and that could be even longer.

Yeah, maybe.....

PS: AMD had this supply problem when they released Zen 2 Laptop APU's, they were very well received and what supply there was sold out instantly, OEM's cried out for more and AMD had to concede they couldn't make them fast enough to keep up.

I think its a good idea to lower expectations. We have no reason at all to believe the RX 6000 series will have 80 CUs, AMD has not confirmed it. 64 is probably a given, as well as similar clocks to a PS5 GPU. We dont know if AMD can / will produce a large 500 600mm2 die.

The only thing we know is that EUV and 5nm are planned for CPUs and probably GPUs next year.

question - Can I ride it out with a R9 390 at 1080p for a bit longer? Are there any games that require a better GPU as a minimum?

Agreed.... this is why i'm open to the idea that there actually is no 80 CU Big Navi.
It depends on what games you play, plenty games these days where the 390 will struggle but if you don't play them and your 390 is still serving you well? Sure stick with it until something doesn't run well on it. My GTX 1070 couldn't cut it for me anymore, so it got replaced.

 

[timorous]

AMD have no intention to react to Nvidia, it was clear before Ampere AMD had no intention to get in before or around the same time as Nvidia and they are still in no rush, just saying "we are on our own release schedule"

I don't know what to make of that, since the Ampere kitchen reveal the 80 CU Navi seems to be no more.

The 2080TI is 35% faster than a 5700XT at 1440P, i'm not going to use 4K for this given that every 8GB card falls further behind the 2080TI, they are not 4K cards.
But at 4K compared to the 3080 it is 32% faster than the 2080TI.

2080TI vs 5700XT 135%
3080 vs 2080TI = 132% (@4K)

Not so Big Navi is rumoured to have 60 CU's, that's 50% more than the 5700XT, after scaling make that +45%

5700XT +45% = 145%

145 / 135 = 1.074. so normalized: 60 CU Not so Big Navi vs 2080TI = 107%

3080 vs 60 CU Not so Big Navi: 132 / 107 = 1.23 (+23%)

RDNA1 vs RDNA2 +10% IPC = 60 CU RDNA2 @1900Mhz. 3080 vs 60 CU RDNA2 @ 1900Mhz: 132 / 117 = 1.128 (+13%)

The PS5 clocks to 2.23Ghz, let's assume that's the limit of RDNA2: 2230 / 1900 = 1.173 (+17%)

Now the 6800XT is 34% faster than a 2080TI, 2% faster than the 3080.

Moores Law thinks AMD think they can compete with the 3080 without the 80 CU Big Navi. i think he's right. With the enhancements of the improved 7nm node and improvements to the RDNA architecture AMD could make a still small 350mm2 GPU at around <2.3Ghz, 250 Watts, 16GB GDDR6 and at least trade blows with the 3080 at £150 less and still get good margins.

I think its a shame if AMD don't slam an 80 CU RDNA2 GPU in Nvidia's face anyway but i think AMD were expecting more from Ampere, they are not impressed.

https://www.techspot.com/review/2099-geforce-rtx-3080/

If a 60 CU card were 350mm^2 then that density is lower then NV are getting on a reworked 10nm Samsung node with Ampere.

AMD could quite easily push density up to that used by Renoir and GA100 which would make a 60 CU card with a 256bit bus in the 250mm^2 region. (around 15B xtors)

With that density they could also have an 80CU card with 384 bit bus in a 350mm^2 die size region. (21B xtors)

I think 60CUs at moderate clockspeeds on a 250mm^2 die would be a great card. It could go toe to toe with the 3070 and would probably be really power efficient because AMD could back off from max clocks. Same with an 80CU part, that many CUs at reasonable clocks would be playing in the 3080 region and be using less power. This would make the cards easier to cool, they would probably overclock well and it gives AMD the chance to bring out clock boosted versions ala the 4890 or 7970 Ghz edition if Nvidia have something to respond with.

EDIT: I should probably add the caveat that this is on the basis that there is not a huge increase in cache which would use transistors and die area.

 

[LePhuronn]

I'm assuming they are Wafer constrained. Zen 3 probably comes from the same wafer supply, and they will sell an absolute cluster-crisis load of them, it might not be a wafer supply issues in the sense enough wafers total, AMD own pretty much all of TSMC's 7nm but that they just can't make them fast enough to keep up with the immediate short term demand, perhaps later when the demand settles down a bit there might be room to make the large 80 CU GPUs dies and that not resulting in starving the market from Zen 3 CPU's and the smaller Navi cards. these Big Navi dies would take up a lot of wafer space compared with the smaller ones and especially Zen 3 CPU's, 4 to 6 times more wafer space.

Ultimately it all comes down to which 7nm process (or processes) AMD are using. That is something we just don't know. Zen 3 isn't using the vanilla 7nm used for Zen 2. But is it performance one or the EUV one? I should imagine RX 6000 will use the performance one, but do the consoles because of their RDNA DNA? What about the refined 6nm node that's tool and design compatible?

N7, N7+, N7P, N6 and AMD have 30% of TSMC's 7nm capacity. Mix and match, smash out everything

 

[humbug]

If a 60 CU card were 350mm^2 then that density is lower then NV are getting on a reworked 10nm Samsung node with Ampere.

AMD could quite easily push density up to that used by Renoir and GA100 which would make a 60 CU card with a 256bit bus in the 250mm^2 region. (around 15B xtors)

With that density they could also have an 80CU card with 384 bit bus in a 350mm^2 die size region. (21B xtors)

I think 60CUs at moderate clockspeeds on a 250mm^2 die would be a great card. It could go toe to toe with the 3070 and would probably be really power efficient because AMD could back off from max clocks. Same with an 80CU part, that many CUs at reasonable clocks would be playing in the 3080 region and be using less power. This would make the cards easier to cool, they would probably overclock well and it gives AMD the chance to bring out clock boosted versions ala the 4890 or 7970 Ghz edition if Nvidia have something to respond with.

I'm assuming RDNA2 will have a lot more transistors, RT, extra Cache, Dual CU's.......... truth is its really difficult to know as @LePhuronn points out. So i'm just going with similar density, because its not likely to be worse....

Ultimately it all comes down to which 7nm process (or processes) AMD are using. That is something we just don't know. Zen 3 isn't using the vanilla 7nm used for Zen 2. But is it performance one or the EUV one? I should imagine RX 6000 will use the performance one, but do the consoles because of their RDNA DNA? What about the refined 6nm node that's tool and design compatible?

N7, N7+, N7P, N6 and AMD have 30% of TSMC's 7nm capacity. Mix and match, smash out everything

Crisis knows.... but i don't think its unreasonable to assume both Zen 3 and RDNA2 are on the best 7nm avalable.

I did some napkin maths on the XBox Series X and that looks firmly on the same 7nm that RDNA1 is on, the old one... they are near identical density.

 

[Dontrocktheboat]

AMD have no intention to react to Nvidia, it was clear before Ampere AMD had no intention to get in before or around the same time as Nvidia and they are still in no rush, just saying "we are on our own release schedule"

I don't know what to make of that, since the Ampere kitchen reveal the 80 CU Navi seems to be no more.

The 2080TI is 35% faster than a 5700XT at 1440P, i'm not going to use 4K for this given that every 8GB card falls further behind the 2080TI, they are not 4K cards.
But at 4K compared to the 3080 it is 32% faster than the 2080TI.

2080TI vs 5700XT 135%
3080 vs 2080TI = 132% (@4K)

Not so Big Navi is rumoured to have 60 CU's, that's 50% more than the 5700XT, after scaling make that +45%

5700XT +45% = 145%

145 / 135 = 1.074. so normalized: 60 CU Not so Big Navi vs 2080TI = 107%

3080 vs 60 CU Not so Big Navi: 132 / 107 = 1.23 (+23%)

RDNA1 vs RDNA2 +10% IPC = 60 CU RDNA2 @1900Mhz. 3080 vs 60 CU RDNA2 @ 1900Mhz: 132 / 117 = 1.128 (+13%)

The PS5 clocks to 2.23Ghz, let's assume that's the limit of RDNA2: 2230 / 1900 = 1.173 (+17%)

Now the 6800XT is 34% faster than a 2080TI, 2% faster than the 3080.

Moores Law thinks AMD think they can compete with the 3080 without the 80 CU Big Navi. i think he's right. With the enhancements of the improved 7nm node and improvements to the RDNA architecture AMD could make a still small 350mm2 GPU at around <2.3Ghz, 250 Watts, 16GB GDDR6 and at least trade blows with the 3080 at £150 less and still get good margins.

I think its a shame if AMD don't slam an 80 CU RDNA2 GPU in Nvidia's face anyway but i think AMD were expecting more from Ampere, they are not impressed.

https://www.techspot.com/review/2099-geforce-rtx-3080/

Your napkin maths is, how should i put this nicely, D+ at best hambug.


5700XT power per watt lead vs Radeon VII is
21% ( 1080P )
14% ( 1440P )
9% ( 4K)

15% overall.

If you get same uplift on the RDNA2 expect 15% across the board.

 

[timorous]

I'm assuming RDNA2 will have a lot more transistors, RT, extra Cache, Dual CU's.......... truth is its really difficult to know as @LePhuronn points out. So i'm just going with similar density, because its not likely to be worse....

Reasonable.

Your napkin maths is, how should i put this nicely, D+ at best hambug.


5700XT power per watt lead vs Radeon VII is
21% ( 1080P )
14% ( 1440P )
9% ( 4K)

15% overall.

If you get same uplift on the RDNA2 expect 15% across the board.

The +50% from GCN -> RDNA was from Vega 64 to 5700XT. The worst RDNA did over Vega was Vega 56 vs 5700XT where the perf/watt increase was 49%

EDIT: https://d1io3yog0oux5.cloudfront.ne...gy+Overview_AMD+FAD+2020_For+Distribution.pdf This is the full presentation. On a previous slide they break down the 50% perf/watt gain and point to end note RX-362 where it says Testing done by AMD performance labs on June 4 2019. Systems were tested with Intel i7-5930K CPU @3.5Ghz (6 core) with 16GB GDDR4 @ 2133 Mhz using an Asus X99-E Motherboard running Windows 10 Enterprise 64 bit (Ver: 1809, build 17763.053). Using the following graphics cards. Navi 10 (driver 19.30_1905161434 (CL# 1784070)) with 40 compute units, versus a Vega 64 (driver 19.4.1) with 40 compute units enabled.

As it so happens the actual gain over Vega 64 was 63% perf/watt in the computerbase.de test suite and 59% perf/watt in the Techpower up test suite.

 

[LePhuronn]

Radeon VII

What's Radeon VII got to do with it? That's an enhanced GCN and a repurposed compute card. RDNA 1 already had a 50% lift over GCN. We're talking about the 50% lift RDNA 2 has over RDNA 1. GCN is irrelevant here.