Question Speculation: RDNA2 + CDNA Architectures thread

Page 208 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

uzzi38

Platinum Member
Oct 16, 2019
2,565
5,575
146
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
 

moinmoin

Diamond Member
Jun 1, 2017
4,934
7,619
136
Not of the Model S by itself no, but it's very likely that this infotainment system trickles down to the other Tesla models. ~130k cars (GPUs) per quarter isn't nothing with how tight supply is projected to be.
That's still a silly low quantity to complain about. Blame consoles instead, those actually sell in millions.

Oh, and it has pyhscial buttons and not effing capactive ones or worse touchscreens which in some countries are actually forbidden to operate during driving. The rate car usability is goign down in terms of lack of buttons, I will probably keep mine for quiet some time till this trend is fixed.
Agree so much. The worst part are all these big bright screens that fail to decently dim down or ideally turn off automatically during night drives.
 

Mopetar

Diamond Member
Jan 31, 2011
7,797
5,899
136
The 6800XT has a much higher effective memory bandwidth than even the 3090 so it's hard for me to believe that it would benefit all that much from a memory overclock.

The problem is that it's only effective up until the 256-bit bus is already saturated and there's a cache miss or a series of them leading to a lot of excess waiting.

Im not trying to argue that what AMD is doing isn't effective. Clearly there are results that speak for themselves, but there will be scenarios where a wider bus would provide a performance uplift.

Of course it's an overall matter of balancing the design. There would also be scenarios where the infinity cache will perform exceptionally well and the wider memory bus goes underutilized.

Im also excited to see what kind of generational improvements can be made to infinity cache, because I expect it to get a lot better over time. Any kind of option for greater developer control could also be massive.
 

MrTeal

Diamond Member
Dec 7, 2003
3,554
1,658
136
The GPU could be 90W or 290W TBP and it wouldn't matter for Tesla. 100kwh battery packs and the power delivery to deliver up to 500+kw of power to the motors without blowing up means the infotainment power/heat is a rounding error either way.
It might be a rounding error in the total heat and power budget of the vehicle, but a couple hundred watts is a lot to dissipate if it's going to be situated behind the dash somewhere.
 

Mopetar

Diamond Member
Jan 31, 2011
7,797
5,899
136
It might be a rounding error in the total heat and power budget of the vehicle, but a couple hundred watts is a lot to dissipate if it's going to be situated behind the dash somewhere.

Fortunately there's plenty of car that you could dissipate that heat through even if you don't have a great system otherwise. In the winter or colder areas it's basically a free source of heat and even in the summer I'm sure there's some air that could be vented past it to help with the cooling.

The biggest issues with these components in a PC is that they have to be kept cool inside of a small stationary case which limits the options for cooling or what can be done with the waste heat. Housing it in a car changes things considerably. I'm fairly certain Tesla of all companies could come up with something novel.
 

Mopetar

Diamond Member
Jan 31, 2011
7,797
5,899
136
Rumors are about TBP are both possibly correct, or at least were at the time given. Just like clock speeds, the TDP/TBP isn't set in stone and is somewhat flexible. AMD could have easily increased the clock speeds slightly based on capability of the silicon or just to better compete with Nvidia. Naturally a change like that increases the TBP. As long as they don't do this so close to launch that the AIBs can't adjust or it isn't too far out of initial guidance it shouldn't be a problem.

There are also other possibilities that the new information isn't fully explained or apparent from the tweet. "up to 150W TBP" could refer to custom AIB models that have higher clock speeds or faster VRAM that exceed a 130W TBP of a reference design from AMD. Just the fact that one is listed as 200W TBP but the other is "up to" 150W TBP suggests there's something different about the two numbers, but perhaps I'm just reading too much into it.

In either case, for something like this, I'd lean towards more recent rumors (assuming they come from a reliable source) as more likely to be correct.
 

TESKATLIPOKA

Platinum Member
May 1, 2020
2,329
2,811
106
Looking at how my perf. targets ended up, I believe he is right.

Why?

AMD may simply have traded power for more performance on this GPU.

P.S. IMO, N22 can have even up to 225W TBP.
High clocks will certainly do that. I was skeptical about 130W TBP combined with higher clocks than N21.
RTX 3060 12GB has 170W with performance a bit higher than 5700XT, so N23 won't be that much better in efficiency.
We will have to wait and see how much clockpeed mobile versions will loose by restricting TBP.
 
Last edited:

Mopetar

Diamond Member
Jan 31, 2011
7,797
5,899
136
I think it only matters if those TBP figures are for the reference spec. The AIB manufacturers are usually free to sell their models with a factory overclock to whatever they can achieve or want to build a cooling solution for.
 

TESKATLIPOKA

Platinum Member
May 1, 2020
2,329
2,811
106
Asus showed what RTX 3060 models they use in their lineup
TPU link
RTX 3060 versions:
60-65W RTX 3060 - 1382Mhz(Boost)
80-85W RTX 3060 - 1525Mhz(Boost)
80-95W RTX 3060 - 1525Mhz(Boost)
90-95W RTX 3060 - 1630Mhz(Boost)
115-130W RTX 3060 - 1802Mhz(Boost)
Not sure If this boost is achievable even with basic TGP(the first number) or only with TGP + dynamic boost = Max GPU power. After looking at Nvidia's specs where 60-115W TGP allows 1283-1703Mhz boost then these specs from Asus include dynamic boost.
So comparing the weakest against the strongest model, you gain 30% higher boost for 2x higher power, not sure what is the actual average clockspeed.

Now why did I post It?
If we compare RTX 3080 FE with 68SM vs RX 6800 XT with 72CU then Nvidia is 6% faster in 4K(Techspot review), in 1440p It performed the same.
RTX 3080 FE average clockspeed is 1931Mhz(TPU).
RX 6800 XT average clockspeed is 2257Mhz(TPU).
The average clockspeed is 17% higher for AMD. If we include the difference in performance then you need 6% higher clocks or 2393Mhz, which is 24% higher than Nvidia. So amD needs ~25% higher clockspeed to be on par with Nvidia, maybe there is a bottleneck in N21 so this can be different in lower segments, but let's say It's the same.
So If I apply this to RTX 3060 mobile with 30SM vs Navi 23 with 32CU:
60W -> 1283Mhz * 1.25 = 1604Mhz
65W -> 1382Mhz * 1.25 = 1728Mhz
85W -> 1525Mhz * 1.25 = 1906Mhz
95W -> 1630Mhz * 1.25 = 2038Mhz
115W -> 1703Mhz *1.25 = 2129Mhz
130W -> 1802Mhz * 1.25 = 2253Mhz
Navi 23 will need to have 1604-2253Mhz boost to perform as mobile RTX 3060 at different TGP settings.
 
Last edited:
  • Like
Reactions: Tlh97 and Mopetar

TESKATLIPOKA

Platinum Member
May 1, 2020
2,329
2,811
106
Now the question is how high can mobile Navi 23 be clocked at supposedly 65-80-90W TGP(GPU+VRAM), now let's say TDP(only GPU) is 50-65-75W and let's say Navi 23 with 150W TBP(TDP: 120W) has 2.45Ghz boost.
If you remember this slide(footnote) from AMD, where they claim 6900XT CU is 2x more effective than 5700xt CU at the same frequency measured in 3dmark11 and only TDP.
Now here is the same slide, where I illustrate possible clocks at 50, 65 and 75W TDP.
45-1080.d1ddab9b.png

The possible clocks for mobile N23 are:
50W TDP (65W TGP) - ~1820Mhz
65W TDP (80W TGP) - ~2025Mhz
75W TDP (90W TGP) - ~2140Mhz
It looks like Navi 23 should be faster within the same power consumption, but take this with a lot of salt, I wasn't so precise with It, and we don't even know If this frequency/power curve is accurate or If It can be applied to N23, and so on.
 
Last edited:

Glo.

Diamond Member
Apr 25, 2015
5,661
4,419
136
Now the question is how high can mobile Navi 23 be clocked at supposedly 65-80-90W TGP(GPU+VRAM), now let's say TDP(only GPU) is 50-65-75W and let's say Navi 23 with 150W TBP(TDP: 120W) has 2.45Ghz boost.
If you remember this slide(footnote) from AMD, where they claim 6900XT CU is 2x more effective than 5700xt CU at the same frequency measured in 3dmark11 and only TDP.
Now here is the same slide, where I illustrate possible clocks at 50, 65 and 75W TDP.
View attachment 39286
The possible clocks for mobile N23 are:
50W TDP (65W TGP) - ~1820Mhz
65W TDP (80W TGP) - ~2025Mhz
75W TDP (90W TGP) - ~2140Mhz
It looks like Navi 23 should be faster within the same power consumption, but take this with a lot of salt, I wasn't so precise with It, and we don't even know If this frequency/power curve is accurate or If It can be applied to N23, and so on.
N23 with 2.5 GHz has 135W TBP.

Only higher clocked versions are over 2.5 GHz in clock speed, and up to 150W TBP.

Have a nice day.
 

Glo.

Diamond Member
Apr 25, 2015
5,661
4,419
136
The average clockspeed is 17% higher for AMD. If we include the difference in performance then you need 6% higher clocks or 2393Mhz, which is 24% higher than Nvidia. So amD needs ~25% higher clockspeed to be on par with Nvidia, maybe there is a bottleneck in N21 so this can be different in lower segments, but let's say It's the same.
So If I apply this to RTX 3060 mobile with 30SM vs Navi 23 with 32CU:
60W -> 1283Mhz * 1.25 = 1604Mhz
65W -> 1382Mhz * 1.25 = 1728Mhz
85W -> 1525Mhz * 1.25 = 1906Mhz
95W -> 1630Mhz * 1.25 = 2038Mhz
115W -> 1703Mhz *1.25 = 2129Mhz
130W -> 1802Mhz * 1.25 = 2253Mhz
Navi 23 will need to have 1604-2253Mhz boost to perform as mobile RTX 3060 at different TGP settings.
I think people will be surprised at how well Navi GPUs perform with lower clock speeds, because of the IPC/Clock speeds curve.
 
  • Like
Reactions: Tlh97 and Mopetar

Mopetar

Diamond Member
Jan 31, 2011
7,797
5,899
136
I think people will be surprised at how well Navi GPUs perform with lower clock speeds, because of the IPC/Clock speeds curve.

Has anyone done some serious testing with Navi 21 underclocks to give us some kind of idea about what to expect? Just the fact that they can clock so high suggests the
potential for massive power savings.
 

uzzi38

Platinum Member
Oct 16, 2019
2,565
5,575
146
Has anyone done some serious testing with Navi 21 underclocks to give us some kind of idea about what to expect? Just the fact that they can clock so high suggests the
potential for massive power savings.
6800 non-XT tested here, MorePowerTool used to reduce power limits only. No undervolts etc applied - only the stock boost algorithm is being used here.
c51ca30b7f821b41a4885783df294653.jpg


Unfortunately this isn't mine so I can't do any testing, I'm hoping to grab a 6800XT at some point. When I do grab one, I'll make sure I do this kind of testing with it.
 

TESKATLIPOKA

Platinum Member
May 1, 2020
2,329
2,811
106
6800 non-XT tested here, MorePowerTool used to reduce power limits only. No undervolts etc applied - only the stock boost algorithm is being used here.
c51ca30b7f821b41a4885783df294653.jpg


Unfortunately this isn't mine so I can't do any testing, I'm hoping to grab a 6800XT at some point. When I do grab one, I'll make sure I do this kind of testing with it.
Is that power for whole board or just GPU?
 

Timorous

Golden Member
Oct 27, 2008
1,532
2,535
136
The problem is that it's only effective up until the 256-bit bus is already saturated and there's a cache miss or a series of them leading to a lot of excess waiting.

Im not trying to argue that what AMD is doing isn't effective. Clearly there are results that speak for themselves, but there will be scenarios where a wider bus would provide a performance uplift.

Of course it's an overall matter of balancing the design. There would also be scenarios where the infinity cache will perform exceptionally well and the wider memory bus goes underutilized.

Im also excited to see what kind of generational improvements can be made to infinity cache, because I expect it to get a lot better over time. Any kind of option for greater developer control could also be massive.

Memory overclocks on the 6900XT show no performance uplift @ 4k so I am not sure that a wider bus would benefit at the current GPU performance level.
 
  • Like
Reactions: Tlh97

Ajay

Lifer
Jan 8, 2001
15,332
7,792
136
Memory overclocks on the 6900XT show no performance uplift @ 4k so I am not sure that a wider bus would benefit at the current GPU performance level.
Shockingly, an architecture designed to reduce the dependence on DRAM bandwidth isn’t very sensitive to changes in DRAM bandwidth. It’s almost like some miracle occurred ;)
 
  • Like
Reactions: Tlh97 and Krteq

Mopetar

Diamond Member
Jan 31, 2011
7,797
5,899
136
Memory overclocks on the 6900XT show no performance uplift @ 4k so I am not sure that a wider bus would benefit at the current GPU performance level.

I could have sworn that some test results posted here did show scaling from increased memory bandwidth, but I could be wrong.

One thing to consider with an memory OC test is to look at if the OC is causing a lot of errors that actually drop performance due to error correction being necessary. Pushing it past the point of stability doesn't make it perform better.

The other possibility is that it's shader (or some other resource) bound at 4K for a particular title and memory bandwidth increases don't help. Perhaps the results that I (think that I) saw were for some non-4K resolution.
 
  • Like
Reactions: lightmanek

GodisanAtheist

Diamond Member
Nov 16, 2006
6,719
7,016
136

Looks like a couple more weeks until we get to wait a couple more months until we get to hear about people buying 6700XTs!
 

uzzi38

Platinum Member
Oct 16, 2019
2,565
5,575
146
So seeing as 3060 reviews are out and the card is mediocre at best, it'll be fun to hear what people's expectations are for Navi22 and Navi23 now.

Where does everyone think the 6700XT will land?
 
  • Like
Reactions: lightmanek

Stuka87

Diamond Member
Dec 10, 2010
6,240
2,559
136
So seeing as 3060 reviews are out and the card is mediocre at best, it'll be fun to hear what people's expectations are for Navi22 and Navi23 now.

Where does everyone think the 6700XT will land?

Mid way between the 5700XT and the 6800. Not expecting anything crazy from it. The question will be its price. Both suggested retail price, and actual retail price.