Question Speculation: RDNA2 + CDNA Architectures thread

Page 124 - 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
2,627
5,927
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
 

Vope45

Member
Oct 4, 2020
114
168
86
Imagine bringing up Ampere in the 'small and efficient' convo lol.

If Ampere is on TSMC 7nm from the very beginning your opinion will be vastly different.

That's not out of the box clocks. That's an overclock.

Out of the box a stock 5700XT caps out at 1905MHz, and even after an undervolt will sit around the 1850MHz mark.

Again you're right. I guess sky high is an overreach.

This is top - down BS.

First of all. Maxwell - Pascal jump was on a new node.

RDNA2 is a jump in clock speeds, on the same node.

Nvidia were not able to clock their GPUs higher, specifically BECAUSE they added all of that Compute capability into gaming architecture. Remember, 3080 consumes 320W of power, and 3090 - 350. USING BRAND NEW NODE.

AMD will still be within reasonable 250-280W power draw range, while clocking largest GPU to 2.2 GHz. ON THE SAME NODE as previous generation architecture!

Dude. Its all due to physical design. Stop moving the goalposts.

Those compute capabilities cause excessive power draw, which disallowes Nvidia to clock their GPUs higher.

Polaris didn't experience any clock uplift going to 14nm neither, Pascal did.

Like I said before AMD physical design team has a lot of work cut out for them. Over several generations of GCN the CU relatively remains the same so any changes to the design would yield an good result. You guys should really need to tamper expectation for RDNA 2, it could backfire. biggest Navi will be hot but it's not gonna be as hot as GA102 that's all I can say.
 

moinmoin

Diamond Member
Jun 1, 2017
4,948
7,656
136
I'm not sure how well you can simulate a pretty radical change like a including a big cache and a smaller bus. I suppose you could use FPGAs to model the behavior, but this isn't just tweaking the design in small ways on an established process. The proposed inclusion of what's being dubbed the infinity cache and the reduction in bus width is a major architectural shift.
Actually simulating the effect of changed cache size, changed bus size etc. are the easy part, those boil down to rather simple mathematical equation. What's hard but very important to get right is simulating the physical characteristics of the specific node used, like hotspots, heat transfer and the impact of more/less transistor density on those, as well as all the interaction of those with the chip's features. Without working simulation one couldn't design chips targeting specific clocks at all and it would be a total gamble burning through mask revisions.

(My impression is that TSMC at this point has a really good grasp on the physical characteristics of its nodes, which may very well be its most unknown competitive advantage it has over Intel and Samsung.)
 

Glo.

Diamond Member
Apr 25, 2015
5,707
4,551
136
Polaris didn't experience any clock uplift going to 14nm neither, Pascal did.

Like I said before AMD physical design team has a lot of work cut out for them. Over several generations of GCN the CU relatively remains the same so any changes to the design would yield an good result. You guys should really need to tamper expectation for RDNA 2, it could backfire. biggest Navi will be hot but it's not gonna be as hot as GA102 that's all I can say.
Polaris CUs have zero to do with not even with RDNA2 CUs but with RDNA1, even. RDNA1 already was better than than any previous physical design AMD brought in previous ... at least couple of years, and was way better architecture than GCN, at least for gaming. It was on par with Turing. So why would AMD not have a good ground to go further? Increase performance per flop, versus RDNA1, increase clocks, make it do more with each cycle, and make it do more cycles each second?

If we would judge the present by what was in the past, we would with confidence say, that AMD could've NEVER in history catch up to Intel, even Intel making mistake,after mistake, so they simply cannot do the same thing, when Nvidia makes mistake, right?

Nobody here has insane expectations, considering that we have seen the data from Power Tables. I don't know what fight you are fighting here, but its wrong, from the ground up.

RDNA2 is way more efficient than RDNA1, and from Ampere. Sure, AMD can clock it the hell out, put its balls to the wall, at 300W of power.

But it will result not in a tie with Nvidia, but with full blown straight up win, if - the specs we have seen in Power Tables are anything to go by.

And lastly. We have seen how efficient are consoles. Those consoles are using the same RDNA2 architecture, and sip power, compared to any previous generation of GPUs. Its because of that silicon design engineers that went to RTG after Zen 2, long after Raja left. And those people are responsible for this physical design, and those high clocks in relatively "humane" power envelopes.

Let me repeat.
Navi 21A - 2.2 GHz max boost clock at 238W per MacOS power tables.
Navi 21B - 2050 MHz at 200W of power.
Navi 22 - 2500 MHz at 170 of power.
Navi 23 - no specs visible, so far.
Xbox Series X 52 CU GPU with 1825 MHz core clock using the same amount of power as Xbox One/One S - 130-140W estimated.

I don't know what is to hamper here. We have all the data, and we can make pretty good guesses, about how powerful RDNA2 is.
 

Vope45

Member
Oct 4, 2020
114
168
86
What are you talking about?

Polaris 10 went up to 1.266 ghz compared to Fiji which was around 1 ghz.

Progressive changes to P20/580 and P30/590 clocked even higher, albeit at disastrous power efficiency, but they did clock higher.

Polaris was already pushed to the max in contrast to Hawaii and Tahiti.
 

ksosx86

Member
Sep 27, 2012
105
44
101
Polaris was already pushed to the max in contrast to Hawaii and Tahiti.
True.
  • Hardware/Software optimization; relevant up to 3+ years (market dominance)
  • Hawaii/Tahiti - Hardware not software optimized at launch, slowly over course of over a year +
RDNA1 - Within 1 month of release relative hardware performance optimized
RDNA2 - Speculated to be at least the same
 
  • Like
Reactions: lightmanek

uzzi38

Platinum Member
Oct 16, 2019
2,627
5,927
146

eek2121

Platinum Member
Aug 2, 2005
2,930
4,026
136
For anyone wondering this is the maintainer of the TPU database.

He also did clarify he was talking about base/boost later on too here. However, he also said he expected reference cards to have a base between 1300MHz - 1600MHz.

Just keep all the data points in mind.
Well yeah, it matches up with the Apple power tables leak. No surprises.
 

Konan

Senior member
Jul 28, 2017
360
291
106
Let me repeat.
Navi 21A - 2.2 GHz max boost clock at 238W per MacOS power tables.
Navi 21B - 2050 MHz at 200W of power.
Navi 22 - 2500 MHz at 170 of power.
Navi 23 - no specs visible, so far.
Xbox Series X 52 CU GPU with 1825 MHz core clock using the same amount of power as Xbox One/One S - 130-140W estimated.

Wait so you are picking the highest amount clocks from the power tables and posting them as actual? That is going to be the definition of disappointed.
Yes SKUs are likely to have high clocks with lots of indications leaning that way (and looking forward to it); but there were ranges in the tables, so don't get how you can randomly pick what you want from that and post as fact?? Help me understand :)
 

uzzi38

Platinum Member
Oct 16, 2019
2,627
5,927
146
Wait so you are picking the highest amount clocks from the power tables and posting them as actual? That is going to be the definition of disappointed.
Yes SKUs are likely to have high clocks with lots of indications leaning that way (and looking forward to it); but there were ranges in the tables, so don't get how you can randomly pick what you want from that and post as fact?? Help me understand :)
You sure you want to say that three posts down from one from the TPU database guy on Navi21 clocks? :p
 

Glo.

Diamond Member
Apr 25, 2015
5,707
4,551
136
Wait so you are picking the highest amount clocks from the power tables and posting them as actual? That is going to be the definition of disappointed.
Yes SKUs are likely to have high clocks with lots of indications leaning that way (and looking forward to it); but there were ranges in the tables, so don't get how you can randomly pick what you want from that and post as fact?? Help me understand :)
Looking at what Matthew Smith posted, it appears those clocks are not the highest.

Its safe to assume at this point, that those power tables from MacOS, are for actual Radeon Pro GPUs that will land in Macs, because they are so low, compared to what we actually see sent out to AIBs in the form of BIOS(this is where MS got the clocks).
 

Stuka87

Diamond Member
Dec 10, 2010
6,240
2,559
136
If Ampere is on TSMC 7nm from the very beginning your opinion will be vastly different.

Like I said before AMD physical design team has a lot of work cut out for them. Over several generations of GCN the CU relatively remains the same so any changes to the design would yield an good result. You guys should really need to tamper expectation for RDNA 2, it could backfire. biggest Navi will be hot but it's not gonna be as hot as GA102 that's all I can say.

First, even if Ampere was on TSMC 7nm, GA102 would still be the largest GPU ever. Going from SS8 to TSMC 7 would not result in any noticeable shrink.

Second, WHY are you comparing Navi to GCN? They are entirely different ISA's. And the ISA has nothing to do with physical layout of a chip.

But, I fear I am wasting my time as this is obviously a troll account that was just created last week.
 

Justinus

Diamond Member
Oct 10, 2005
3,174
1,516
136
First, even if Ampere was on TSMC 7nm, GA102 would still be the largest GPU ever. Going from SS8 to TSMC 7 would not result in any noticeable shrink.

Second, WHY are you comparing Navi to GCN? They are entirely different ISA's. And the ISA has nothing to do with physical layout of a chip.

But, I fear I am wasting my time as this is obviously a troll account that was just created last week.

GA100 is ~50% more dense on whatever TSMC 7nm node it's on compared to GA102 on SS8
 

uzzi38

Platinum Member
Oct 16, 2019
2,627
5,927
146

But it's also comparable to the total heat output of most modern consoles. Series X's power draw—a theoretical maximum of 315 watts, but closer to 190 watts as measured by a Kill-a-Watt device while playing "next-gen" games—is a good metric, compared to the 120W average gameplay draw of the original Xbox One and 185W average gameplay draw of Xbox One X. (You read that correctly: In my limited testing, Series X and XB1X are absolutely comparable in power draw.) Series X just happens to vent its air at a particularly noticeable rate, which is probably why early reports claimed it ran hot.

Nice.
 

eek2121

Platinum Member
Aug 2, 2005
2,930
4,026
136
Wait so you are picking the highest amount clocks from the power tables and posting them as actual? That is going to be the definition of disappointed.
Yes SKUs are likely to have high clocks with lots of indications leaning that way (and looking forward to it); but there were ranges in the tables, so don't get how you can randomly pick what you want from that and post as fact?? Help me understand :)

You are right, the power tables are for pro cards. Actual AIB clocks should be higher...
 

DisEnchantment

Golden Member
Mar 3, 2017
1,602
5,788
136
Nice.
Woooah. Assuming a conservative 40W for CPU cores, that's 150 W for the GPU + G6. That's already something comparing 52CU w/ 320bit G6 @1825 MHz vs 256bit G6 40CU N10 at 225W@1850MHz
Assuming 0 perf/clock increase vs N10 that is way more than 50% perf/watt improvement. Granted, it is in its sweetest spot.
 

eek2121

Platinum Member
Aug 2, 2005
2,930
4,026
136
I stumbled across some old news posts regarding the 5700XT. Apparently there is (or was?) a way to unlock higher power limits. This allows a 5700XT to boost to around 2.2 Ghz while consuming 250W of power. Looks like RDNA1 already had some headroom?
 

raghu78

Diamond Member
Aug 23, 2012
4,093
1,475
136
Woooah. Assuming a conservative 40W for CPU cores, that's 150 W for the GPU + G6. That's already something comparing 52CU w/ 320bit G6 @1825 MHz vs 256bit G6 40CU N10 at 225W@1850MHz
Assuming 0 perf/clock increase vs N10 that is way more than 50% perf/watt improvement. Granted, it is in its sweetest spot.

Zen 2 at 3.66 Ghz (SMT off) draws 55w approx. That 12 TF RDNA2 GPU in Series X is drawing less than 130w. Even against mobile RX 5700M that Series X GPU is atleast >1.5x perf/watt if you account for twice the VRAM capacity