• Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.

Digital Foundry: AMD Zen 2/Navi Reaction + What It Means For Next-Gen Consoles

Mista

Dragonslayer
Nov 21, 2014
7,378
8,309
1,085


A little later than planned (flights, jetlag, other things), here's Rich and John with their take on the recent AMD Zen 2 and Navi reveals - along with what it all means for the next generation consoles.
 
Last edited:
Jul 29, 2013
1,259
1,310
615
Oregon, US
Basically said based on Navi reveal 40CU with 4 disabled would be ideal for console APU size. I commented the other day that based on power consumption of Navi and $399 price point, 5700 Pro at 1.8GHz would be a good fit.
 

ethomaz

Member
Mar 19, 2013
24,294
3,468
800
37
Brazil
Basically said based on Navi reveal 40CU with 4 disabled would be ideal for console APU size. I commented the other day that based on power consumption of Navi and $399 price point, 5700 Pro at 1.8GHz would be a good fit.
Funny enough they needs to have a bit of knowledge before made these claims.

Actual AMD Navi is a chip with 2 Shader Engine, 10 DCU each, 2 CUs per DCU. If you look at the chip layout it is symmetrical.

So how to increase these CUs?

To continue simétricas they will need to add 2 DCU to both SEs... that means the CU count will go to 48 and with 4 disabled it will have 44CUs.

Other option is to add two new SEs but that way the limit of 80CUs so I won’t enter is speculation because there is dozens of symmetrical options with 4SEs (64, 72, 56, 80, etc).

3 SEs I don’t believe AMD will use due the the balance... there is no way to get symmetrical units that way.

But hey looking at GCN 44CUs with 4 disabled is pretty easy.

Of course an odd number of CUs with different numbers in each SE is possible but I never saw AMD doing that because you create bottleneck to the division of work.
 
Last edited:
  • Like
Reactions: DanielsM

Panajev2001a

GAF's Pleasant Genius
Jun 7, 2004
13,509
1,670
1,955
Surprised at how they missed the SIMD units doubling in width, handling 32 wide warps in one clock and the old 64 wide warps in two (and the latter being more of the “enhanced backward compatibility with GCN” part than the number of CU’s)...
 
Last edited:
  • Like
Reactions: llien
Jul 29, 2013
1,259
1,310
615
Oregon, US
Funny enough they needs to have a bit of knowledge before made these claims.
They were talking about the size of the APU.

I personally tend to focus more on power consumption. You can see power consumption of Navi with either 36CU at 1.6GHz(5700 Pro), or 40CU at 1.7GHz(5700XT), or 40CU at 1.8GHz(50th Anniversary).
 

ethomaz

Member
Mar 19, 2013
24,294
3,468
800
37
Brazil
Surprised at how they missed the SIMD units doubling in width, handling 32 wide warps in one click and the old 64 wide warps in two (and the latter being more of the “enhanced backward compatibility with GCN” part than the number of CU’s)...
They missed that too?

That is probably the biggest change.

In GCN you have to fill 64 wide warps with SIMD-16... that required 4 cycles to execute the 64.

Now RDNA has both SIMD-32 and just 32 wide warps... so in 1 cycle you processed the instructions and are ready to receive more.

That is what make RDNA more efficient.
 
Last edited:

ethomaz

Member
Mar 19, 2013
24,294
3,468
800
37
Brazil
They were talking about the size of the APU.

I personally tend to focus more on power consumption. You can see power consumption of Navi with either 36CU at 1.6GHz(5700 Pro), or 40CU at 1.7GHz(5700XT), or 40CU at 1.8GHz(50th Anniversary).
RDNA power draw didn’t allow too big chip or clocks in console form.

That is one of the biggest issues with GCN that carry on RDNA... they only had improvement due the shrink... no architecture improvement in power draw like NVidia did with Maxwell.
 
  • Like
Reactions: DanielsM

Armorian

Member
Jan 17, 2018
785
579
360
For Digital Foundry the PS5/Scarlett will be at best GTX 1080 level (y)
The best 2019 Navi discrete highly clocked GPU is 1080 level (few % better)... so consoles would have to be based on bigger chip or/and have higher clocks. Big chip would occupy more die space and high clock needs better cooling/consumes more power (and gets you more shit chips in production), I think performance around 9TF RDNA is quite possible and that would be ~12TF if it was Polaris chip.
 
Last edited:

Evilms

Member
Dec 16, 2018
673
1,462
380
The best 2019 Navi discrete highly clocked GPU is 1080 level (few % better)... so consoles would have to be based on bigger chip or/and have higher clocks. Big chip would occupy more die space and high clock needs better cooling/consumes more power (and gets you more shit chips in production), I think performance around 9TF RDNA is quite possible and that would be ~12TF if it was Polaris chip.
Hopefully, but in the meantime wait and see.
 
Jul 29, 2013
1,259
1,310
615
Oregon, US
RDNA power draw didn’t allow too big chip or clocks in console form.

That is one of the biggest issues with GCN that carry on RDNA... they only had improvement due the shrink... no architecture improvement in power draw like NVidia did with Maxwell.
Everything about Navi seems good except the power draw. Feel like there was a bigger perf/watt jump with Polaris.
 

LordOfChaos

Member
Mar 31, 2014
9,089
1,139
710
8 core Zen 2 has a 32MB L3 cache, holy. We're coming up from zero L3 cache and 4MB of shared L2 (2MB per 4 core CCX) in the PS4/XBO

I like how their solution to cross CCX communication limiting maximum game framerates was "well, we'll just double the caches everywhere" lol.

We're coming a loong way from Jaguar and even over Ryzen where max framerates suffered vs Intel, this should allow a lot more high framerate games, but as every generation it's always the devs choice where to apply the extra power.
 

thelastword

Member
Apr 7, 2006
8,169
2,012
1,380
Consoles will not even be based on current Navi (7/7), but the evolved NAVI in 2020...…..Perhaps on 7nm+/5nm......

Consoles will pack a punch...I think if in the prior gens or in this one in particular, top tier devs could give us such amazing feats on 1.84TF and a Jaguar CPU.....(Talking SWWS), then their output with a top shelf CPU and GPU will really evolve beyond our imaginations, especially Sony First Party........SWWS was able to give us the best looking games this gen on PS4 spec....So it shows, if you have the hardware and don't have the people to push it, it means nothing........

It's the same with RTX, the only way RTX can really take off is when all devs are on board, all GPU manufacturerers and most importantly, the real top brass dev heads...Till then, all we're getting with RTX games will be overdone like devs high on bloom and parallax mapping......
 

SonGoku

Member
Aug 16, 2018
3,572
3,359
550
This video is a waste of time, building a chip around 7nm weakness (clocks) instead of its strength (density increase) is shortsighted.
72CU APU = 399.1 mm2 (64 enabled)
64CU APU = 385.6 mm2 (56 enabled)

With the possibility of quick cost reductions come 2021:
Launch die size at 7nm"6nm" die size (15% reduction)
400 mm2​
340 mm2​
390 mm2​
331.5 mm2​
380 mm2​
323 mm2​
A 36CUs chip power consumption at 1.8Ghz would be just as high if not more than a 64CU chip at 1600Mhz.
Everything about Navi seems good except the power draw. Feel like there was a bigger perf/watt jump with Polaris.
Because AMD is pushing high clocks... Sacrificing perf/watt for max performance
 
Last edited:

LordOfChaos

Member
Mar 31, 2014
9,089
1,139
710
Fwiw


Node16 nm10 nm7 nm7nm/10nm Δ
Gate90 nm66 nm57 nm0.86x
Min Metal64 nm42 nm40 nm0.95x


Never realized that the 16nm-10nm jump was actually quite substantial, and the 10nm to 7 is much less so. Though the transistor profile improved. TSMC calling this 7nm was even more aggressively...Optimistic, than fab naming bs already was, either this was the feature size on a ring oscillator or some other un-chip component, or else an original target long abandoned but used the name of. The fin is also very tall.


Node10 nm7 nmΔ
Transistor Profile
Fin Pitch36 nm30 nm0.83x
Fin Width6 nm6 nm1.00x
Fin Height42 nm52 nm1.24x

7nm+ is where density would get somewhat close, and is in fact closer to Intels 10nm.

"2nd Generation 7nm
TSMC also developed a 2nd generation of their 7nm process. This is an optimized process which uses the same design rules and DUV and is unrelated to 7nm+ which is EUV-based. This process is entirely design-compatible with the first generation but enjoys additional power and performance enhancements. For their second generation process, TSMC made some additional optimizations.

  • Fin profile Optimizations
  • Epi Optimizations
  • MOL resistance Optimizations
  • FEOL capacitance
  • Metal gate Optimizations
All in all, the 2nd-generation 7nm process is said to deliver over 5% improvement in performance. Additionally, at the same leakage, at high frequencies, the second-generation 7nm process has improved the Vmin by 50 mV. Qualcomm has stated that the second-generation 7-nanometer node will be used for their next-generation cellular 5G modem which will deliver 2x peak data rate over their first-generation 5G modem."
 

Racer43

Member
Jun 14, 2018
306
234
265
"2nd Generation 7nm
TSMC also developed a 2nd generation of their 7nm process. This is an optimized process which uses the same design rules and DUV and is unrelated to 7nm+ which is EUV-based. This process is entirely design-compatible with the first generation but enjoys additional power and performance enhancements. For their second generation process, TSMC made some additional optimizations.

  • Fin profile Optimizations
  • Epi Optimizations
  • MOL resistance Optimizations
  • FEOL capacitance
  • Metal gate Optimizations
All in all, the 2nd-generation 7nm process is said to deliver over 5% improvement in performance. Additionally, at the same leakage, at high frequencies, the second-generation 7nm process has improved the Vmin by 50 mV. Qualcomm has stated that the second-generation 7-nanometer node will be used for their next-generation cellular 5G modem which will deliver 2x peak data rate over their first-generation 5G modem."
This is what next gen consoles is going to be manufactured with in my opinion. No way in hell next gen consoles will be manufactured with regular 7nm, when both 5nm (which Sony said they wouldnt use) and 7nm+ is available in a years time (or probably before that). That would be pure stupidity in a size/power sensitive product. (Which is why mobile phones usually are out first on a new node).

In AMD`s slides, it says next gen RDNA will be manufactured with 7nm+, which is what they say they will use with raytracing. If Microsoft will have hardware based raytracing, its almost a given Sony will to. Big Navi will be 7nm+ with next gen RDNA and raytracing. (Consoles are probably a derivative of Big Navi).
 
Last edited:

truth411

Member
May 19, 2014
1,514
56
390
Basically said based on Navi reveal 40CU with 4 disabled would be ideal for console APU size. I commented the other day that based on power consumption of Navi and $399 price point, 5700 Pro at 1.8GHz would be a good fit.
Wait a sec. If my math is correct wouldn't that only be 8.3tflops? Kick rocks with that. 12Tflops minimum.
 

ArabianPrynce

Member
Jun 4, 2019
154
125
195
This video is a waste of time, building a chip around 7nm weakness (clocks) instead of its strength (density increase) is shortsighted.
72CU APU = 399.1 mm2 (64 enabled)
64CU APU = 385.6 mm2 (56 enabled)

With the possibility of quick cost reductions come 2021:
Launch die size at 7nm"6nm" die size (15% reduction)
400 mm2​
340 mm2​
390 mm2​
331.5 mm2​
380 mm2​
323 mm2​
A 36CUs chip power consumption at 1.8Ghz would be just as high if not more than a 64CU chip at 1600Mhz.

Because AMD is pushing high clocks... Sacrificing perf/watt for max performance
How confident are you that sony will go with these specs? the thing that rich said in a different video is that always lower your expectation when it comes to console specs. I mean dont these speculative discussions happen every generation where people expect higher specs than the final product?. I have a slight feeling we all might get dissapointed with the final specs.

I asked to see what you have to say.
 

Mass Shift

Member
Jun 12, 2019
115
83
270
"Developers are quite happy with PS5" - Rich

Megaton :messenger_open_mouth:
Because they actually HAVE kits from Sony. I'm not sure what MS has provided but it's nothing near final specs. Or at least still alpha stage.
 
Last edited:

SonGoku

Member
Aug 16, 2018
3,572
3,359
550
How confident are you that sony will go with these specs? the thing that rich said in a different video is that always lower your expectation when it comes to console specs
My absolute minimum expectation is 11TF
56CUs @1500-1540Mhz = 10.7TF-11TF

My more optimistic "mid" expectation
56CUs @1680Mhz = 12TF
64CUs @1500Mhz = 12.2TF

My most optimistic of expectations (needs 7nm EUV):

64CUs @1710Mhz = 14TF
72CUs @1550Mhz = 14.28TF
 
  • Like
Reactions: ArabianPrynce

quest

Banned from OT
Jul 17, 2004
3,997
85
1,525
My absolute minimum expectation is 11TF
56CUs @1500-1540Mhz = 10.7TF-11TF

My more optimistic "mid" expectation
56CUs @1680Mhz = 12TF
64CUs @1500Mhz = 12.2TF

My most optimistic of expectations (needs 7nm EUV):

64CUs @1710Mhz = 14TF
72CUs @1550Mhz = 14.28TF
I think it will be close to option 1 for them. There is no point in a small navi cu count won't have the juice to be next gen imo. The cu count needs to be north of 50 to be really next generation. I am also interested how far behind the AMD RT hardware will be compared to the nvidia version. How far the effects will have to be scaled back vs the 2070 and higher nvidia cards.
 
Last edited:
  • Like
Reactions: SonGoku

SonGoku

Member
Aug 16, 2018
3,572
3,359
550
I think it will be close to option 1 for them. There is no point in a small navi cut count won't have the juice to be next gen imo. The current count needs to be north of 50 to be really next generation. I am also interested how far behind the AMD RT hardware will be compared to the nvidia version. How far the effects will have to be scaled back vs the 2070 and higher nvidia cards.
It would be a dream if consoles licensed imgtech RT
 
  • Like
Reactions: Panajev2001a

ArabianPrynce

Member
Jun 4, 2019
154
125
195
12TF Navi would match a RTX 2080
Well hot damn, Its hard to grasp of how amazing this will translate into real time graphics. Like holy fuck, if naugty dog got their hands on this or guerrilla damn. I still doubt it, richard in the df video mentioned that devs are "quite happy". For my interpreation it if the ps5 was good devs would be "overwhelmed" or stunned.
 

SonGoku

Member
Aug 16, 2018
3,572
3,359
550
I still doubt it, richard in the df video mentioned that devs are "quite happy". For my interpreation it if the ps5 was good devs would be "overwhelmed" or stunned.
I would take anything DF says with a grain of salt, they have no inside info and Richard is butthurt Sony chose Wired over DF
That being said my baseline expectation is 11TF which its still beast
 
  • Like
Reactions: demigod

ArabianPrynce

Member
Jun 4, 2019
154
125
195
I would take anything DF says with a grain of salt, they have no inside info and Richard is butthurt Sony chose Wired over DF
That being said my baseline expectation is 11TF which its still beast
Hmm i am surpised if true, still DF gives the best most researched opinion in regards gaming tech. Curious as to the source of richard being mad at sony not interviewing with them lol
 

LordOfChaos

Member
Mar 31, 2014
9,089
1,139
710
Well hot damn, Its hard to grasp of how amazing this will translate into real time graphics. Like holy fuck, if naugty dog got their hands on this or guerrilla damn. I still doubt it, richard in the df video mentioned that devs are "quite happy". For my interpreation it if the ps5 was good devs would be "overwhelmed" or stunned.
Could just be the old English penchant for understatement

pants literally on fire
"It is quite warm"