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

 

[Mopetar]

It's almost half of April and still no news about N23's release date. Does anyone have any new info?

I think everyone's too depressed that most of the cards will just be snatched up by scalpers or miners. It's a bit harder to be enthusiastically engaged with something when it seems impossible to obtain.

The last I heard about it was possibly April, but that seems rather unlikely right now. Supposedly Tesla is using Navi 23 in some of their vehicles, so they probably get their secured first, along with anyone else that would have ordered a bunch to slap in some pre-builts, etc. I don't know if Apple is going entirely with their own SoCs from now on, but Navi 23 sounds like the kind of part they'd put in an iMac.

 

[Glo.]

The worst part is that we should be seeing at this point any news about RX 6700 non-XT releases.

Nothing. Silence. Nada.

However, Elon was thanking lately the suppliers of Tesla for deliveries of equipment for latest Tesla cars.

And that means: Navi 23 deliveries.

 

[GodisanAtheist]

The worst part is that we should be seeing at this point any news about RX 6700 non-XT releases.

Nothing. Silence. Nada.

However, Elon was thanking lately the suppliers of Tesla for deliveries of equipment for latest Tesla cars.

And that means: Navi 23 deliveries.

-N22 yields are likely very good, supply is relatively low, demand is obscene.

AMD would be fools to sell the same piece of silicon for any less by way of artificial segmentation, and if there are defects the stockpile of defective parts is likely very very low.

 

[Mopetar]

We haven't seen Nvidia make any announcements regarding their cut down GA-106 dies (supposedly 3050 Ti and 3050) yet either, so AMD may also be waiting to see what they do before making any official moves.

Then again, it's not like it really matters now. They could say they're releasing both Navi 23 and the 6700 starting at $450 and they'd still all sell out immediately.

 

[uzzi38]

Did some V/f testing on my 6700XT. Clocks recorded and core-only power are in TimeSpy GT2 (its the tougher of the two GPU tests, required higher power on average than GT1). Full GPU power as recorded by Radeon Software is about 40W above these, and at stock Radeon Software record 12W less than the rated 230W TDP in the worst case. Although that being said, most reviews I've seen have had the reference 6700XT come in below the rated TDP as well, same with AIB cards. Unfortunately I don't have a way of recording power at the PCIe connector so I can't give a reliable estimate there - this is the best I can do.

Clock speed set (min-max, MHz)	Average Clock recorded (MHz)	Core-only power (W)	TimeSpy graphics Score
2401-2665	2565	175	12161
2401-2501	2404	140	11580
2300-2400	2295	116	11035
2200-2300	2200	101	10690
2100-2200	2098	88	10288
2000-2100	2004	77	9769
1900-2000	1912	69 (nice)	9358

Here, have a graph too.

 

[uzzi38]

Here's the complete chart scaling from 600MHz all the way up to the max I could do of 2565MHz.

Interestingly it's a very linear 3W per 100MHz up to 1700MHz. Then it starts to rise per 100MHz.

 

[Mopetar]

Seems like a really efficient design up to about 2 GHz. It should make for some good APUs once AMD gets around to incorporating RDNA2 into their other products.

Even though I'm not expecting to see those for quite a while they still may be the best shot at getting RDNA2 graphics.

 

[TESKATLIPOKA]

Great work.
So 1.9GHz N22 would have only 115W TBP? Very nice.
For mobile this wouldn't be so fantastic considering N22 should go against mobile RTX 3070.
RDNA2 needs ~25% more frequency to be on par with Ampere at similar config.
RTX 3070 at 100W has 1390Mhz boost clock -> N22 at 1738Mhz is Its equal.
RTX 3070 at 125W has 1620Mhz boost clock -> N22 at 1944mhz is Its equal.
If they could shave from those 40W 10-15W off for the mobile versions then It would look much better, hopefully they can.

 

[blckgrffn]

Seems like a really efficient design up to about 2 GHz. It should make for some good APUs once AMD gets around to incorporating RDNA2 into their other products.

Even though I'm not expecting to see those for quite a while they still may be the best shot at getting RDNA2 graphics.

Considering the wattage a 5700xt needs on the same node at essentially that clock speed it’s pretty amazing and makes you hopeful that AMD is firing on all cylinders in the GPU department.

Whatever good it does us 😂

 

[Saylick]

Wow, the perf/W if core power doubles with just dropping the clocks down to 2 GHz... RDNA3 having twice the CUs within the same TDP seems within reason then if all it takes is a small lowering of clocks.

 

[uzzi38]

Here's the complete chart scaling from 600MHz all the way up to the max I could do of 2565MHz.

Interestingly it's a very linear 3W per 100MHz up to 1700MHz. Then it starts to rise per 100MHz.

View attachment 43302

So I've realised that the linear scaling up to 1700MHz is due to there being a minimum voltage. I can work around it using MorePowerTool, so I guess I will at some point when I get the time.

 

[TESKATLIPOKA]

Here is a link to latency comparison between RDNA2 and Ampere
Latency comparison TPU

 

[TESKATLIPOKA]

It would be great If somehow you could partially disable Infinity cache and see how big impact It has on performance.
For N22 I would rather see only 64MB IC and +8CU with lower clocks, so It wouldn't have such a high TBP, yet the performance would stay the same. 230W TBP for a 40CU is very bad when a 80CU GPU has only 70W more.

P.S. I know 96MB IC is because of 192bit memory bus(16MB IC per 32bit memory bus) and 64MB wouldn't be possible, but 72MB would, If we set It to 12MB per 32bit.

 

[uzzi38]

Minimum voltage set to 725mv now. Still linear scaling up to 1300MHz, but everything past is exponential as expected. To my knowledge running a chip below 700mv isn't usually a good idea, so I figured it wasn't worth the risk.

 

[uzzi38]

It would be great If somehow you could partially disable Infinity cache and see how big impact It has on performance.
For N22 I would rather see only 64MB IC and +8CU with lower clocks, so It wouln't have such a high TBP. 230W TBP for a 40CU is nonsense when a 80CU GPU has only 70W more.

If I could, you can bet I would. But unfortunately, that sort of granularity is totally impossible.

Also, I don't think less IC and more CUs (even with lower clocks) would be a good idea. I'd still want to stick to 96MB IC personally.

 

[TESKATLIPOKA]

If I could, you can bet I would. But unfortunately, that sort of granularity is totally impossible.

Of course you would test It, that's a given.

Also, I don't think less IC and more CUs (even with lower clocks) would be a good idea. I'd still want to stick to 96MB IC personally.

What's the advantage?
I don't see much advantage in having 96MB IC Instead of 64MB, but clearly see an advantage in having 20% more CU at 2.3GHz for example. It would consume a lot less 116W * 1.2 = ~140W vs 175W and performance should be better.

 

[uzzi38]

Of course you would test It, that's a given.


What's the advantage?
I don't see much advantage in having 96MB IC Instead of 64MB, but clearly see an advantage in having 20% more CU at 2.3GHz for example. It would consume a lot less 116W * 1.2 = ~140W vs 175W and performance should be better.

Oh I'm just disagreeing with cutting the IC, not the increase in CU count. That would obviously be a significant improvement.

As for IC, it may seem a bit overkill - 96MB for 1440p, but thinking about it differently, every time you access the IC, you're not accessing VRAM, which is also extra power saved. Besides, 20% more CUs at 2.3GHz is still a 5-10% increase in compute, so I'd rather stick to the higher end of available access to memory.

 

[TESKATLIPOKA]

Not accessing Vram but IC is saved power, correct. The question is how much power we actually save by It and If It's worth the extra 32MB. I am quite skeptical about that, If N23 with only 20% fewer CUs has only 1/3 of IC and no problem.

P.S. I find It funny how we are discussing about a bigger IC saving a few Watts on a card where the GPU is clocked so high that It's power consumption is close to the much faster RX 6800.

 

[Paul98]

It's going to be real nice seeing IC in the next generations, as size increases won't be to crucial from where it's starting, also AMD doesn't have to use top end performance vram. Leaving more room and resources for other things

 

[uzzi38]

So yeah, I figured out how to do read-only links I think, so here's the spreadsheet with actual numerical values if anyone wants: https://docs.google.com/spreadsheets/d/1R57Xfp0pES5AFwbT-zAf8FXfE_iIRmkY7iKN0-Azbdc/edit?usp=sharing

 

[TESKATLIPOKA]

Great, thanks.
1/2 performance at 1146Mhz with only 16% power draw. That's "only" 3.125x better efficiency.

The question is what would happen to average clocks in a demanding game like Control and not just 3dmark.

P.S. Don't know how N23 could have only 135W TBP with supposedly ~2.5GHz clocks. Based on your findings It looks like N23 will be clocked at ~2.3GHz.

 

[uzzi38]

AmbitiousButRubbish on Twitter: "kinda wanted do a longer write-up, but since @uzzi38 beat me to the punch i will rush out some of my data as long as it's hot. @nerdtechgasm @JirayD @_rogame https://t.co/Xyy6rgPJgJ" / Twitter

Comparison between RDNA1 and RDNA2, similar testing methodology to me, but he's testing Firestrike GT1 instead.

 

[DiogoDX]

3.2GHZ on LN2!

 

[GodisanAtheist]

3.2GHZ on LN2!

-Impressive that it was an actual boot and benchmark run, not just one of those silly "Boot to desktop" OCs.

Mildly frustrating that we didn't get any performance numbers there... would have been nice to see the FPS curve.

Edit: Also somewhat anxiety inducing to see someone using LN2 without any apparent PPE and just freehanding it.

 

[eek2121]

3.2GHZ on LN2!

Find my comments before the 6000 series was launched. Questions?

Wait until you see next gen...