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

 

[senseamp]

I don't believe AMD is just sitting on a product that's good and ready to go or almost there. Time is money when your value add is depreciating at the rate of Moore's law. Something isn't ready for prime time and ETA is unknown, or the hype train would be leaving the station by now. Might be a "good from far but far from good" situation, like a silicon bug or driver maturity with all the new features.

 

[Kenmitch]

I don't believe AMD is just sitting on a product that's good and ready to go or almost there. Time is money when your value add is depreciating at the rate of Moore's law. Something isn't ready for prime time and ETA is unknown, or the hype train would be leaving the station by now. Might be a "good from far but far from good" situation, like a silicon bug or driver maturity with all the new features.

They're currently still in party mode. They know they got this one in the
bag! /s

No news/leaks = no leaks/news and nothing more.

 

[Helis4life]

AMD will announce and release when they are ready to. They have their own schedule and it isn't going to be dictated by NVIDIA or anyone crying they'll just buy the 3080.

I wouldn't expect any concrete announcement from AMD until mid October with a release time early November.

If you can't wait until November, buy the 3080. I promise you amd isn't going to care one way or another. They are launching potentially the most exciting CPUs in years, a brand new GPU architecture that will also power both major consoles which will also release immediately after RDNA2. This isn't a matter of coincidence, it's most likely by design and has been planned this way for quite a while.

 

[Glo.]

There is no reviews of Ampere GPUs and people expect that ALREADY there will be leaks for RDNA2?

As I have said. AMD does not want to spill all of the beans already cause Nvidia still has their own cards close to them. And there are reasons why they do not want to spoil the fun.

 

[Dribble]

I remember about 6 months ago when Glo was trying to convince the world that Nvidia wouldn't release anything this year saying don't count Nvidia out - historically they pretty well always execute well.
Now Glo is attempting to convince the world that AMD has a massive trump card hiding away I gotta say don't hype AMD up - historically they pretty well always fail to live up to it.

 

[itsmydamnation]

I remember about 6 months ago when Glo was trying to convince the world that Nvidia wouldn't release anything this year saying don't count Nvidia out - historically they pretty well always execute well.
Now Glo is attempting to convince the world that AMD has a massive trump card hiding away I gotta say don't hype AMD up - historically they pretty well always fail to live up to it.

Only when you pick and choose your history, also R&D dollar, align things to R&D dollar and then look at history.
But i guess that doesn't fit your narrative?

if you cant tell, im optimistic, but i always am.

 

[Glo.]

I remember about 6 months ago when Glo was trying to convince the world that Nvidia wouldn't release anything this year saying don't count Nvidia out - historically they pretty well always execute well.
Now Glo is attempting to convince the world that AMD has a massive trump card hiding away I gotta say don't hype AMD up - historically they pretty well always fail to live up to it.

Go back to what I was actually saying in Ampere thread.

That if anything Nvidia will release it will be at the end of the year.

I will ease it for you. 5th page of Ampere thread.

I will repeat this once again. Because Samsung failed to deliver 7 nm process, Nvidia hasn't even taped out their consumer, gaming architecture.

Gaming GPUs from Nvidia will not happen before late 2020, or early 2021. Period. Only 7 nm GPU we will see this year is Ampere, GA100, for HPC, that replaces Volta.

Period.

Its a post from JANUARY. Does this fit your narrative?

 

[pj-]

Go back to what I was actually saying in Ampere thread.

That if anything Nvidia will release it will be at the end of the year.

I will ease it for you. 5th page of Ampere thread.

Its a post from JANUARY. Does this fit your narrative?

Do you forget that other people can also check your posts?

Edit: And this is from 3 minutes of looking through January posts. I'm sure someone could go find dozens more

I will repeat this. 7 nm Products from Nvidia we will see in 2020 is only HPC GA100 Chip. Gaming cards are 7 nm, and Late 2020/Early 2021 products. And Gaming cards are not Ampere architecture. it will be completely different name.

wrong

This year is only HPC for Nvidia. Which good thing. Volta needed replacement, considering that just around the corner are MI100 GPUs from AMD. Competition at its best.

wrong

I know that Gaming cards have not been taped out. Also gaming cards are not based on Ampere, and are not named Ampere, which I have said before.

I was asking whether Ampere has been taped out on TSMC's process, or not... The HPC one.

wrong

Tape out should happen around April, for gaming, next gen GPUs, which makes them Q1/H1 2021 available.

Perfect for GTC 2021 release

wrong

 

[Glo.]

Do you forget that other people can also check your posts?

Edit: And this is from 3 minutes of looking through January posts. I'm sure someone could go find dozens more



wrong



wrong



wrong



wrong

And what were the contexts of the discussion? Hmmm?

Yes, I was wrong about it not being Ampere named.

I was not wrong on process, on timeline release for 2020 year, the reasons for the 8 nm process choices.

IN JANUARY of 2020.

 

[sandorski]

Even if RDNA2 on the PC is a dud, AMD is sitting pretty having the Consoles. If they compete with the 3800 and less and Nvidia has low supply, they'll do fine on the PC side. No matter what happens this gen, both companies are going to do well.

 

[Konan]

Even if RDNA2 on the PC is a dud, AMD is sitting pretty having the Consoles. If they compete with the 3800 and less and Nvidia has low supply, they'll do fine on the PC side. No matter what happens this gen, both companies are going to do well.

It won't be a dud mate. It'll be pretty competitive for sure. Their main focus will be on fulfilling console demand. We might not even see a super top end competitor until next year.

 

[pj-]

And what were the contexts of the discussion? Hmmm?

Yes, I was wrong about it not being Ampere named.

I was not wrong on process, on timeline release for 2020 year, the reasons for the 8 nm process choices.

IN JANUARY of 2020.

So this statement wasn't wrong?

"Tape out should happen around April, for gaming, next gen GPUs, which makes them Q1/H1 2021 available."

How about I use the Glo. school of speculation and post this:

"There will not be a mid-gen refresh of the new consoles"

and then a week later:

"There will be a mid-gen refresh of the new consoles"

Amazing! I can't be wrong!

Sorry, you don't get credit for guessing that nvidia would release gpus somewhere between "late 2020" (not even sure I'd grant you that mid september is "late") and h1 2021.

 

[Glo.]

Do you know when tape out happened for those GPUs?

If it happened in April does it make my stament wrong, considering that usually release is 9 months after a tape-out, and it is only 5 months since April?

 

[Konan]

I think its worth quoting this, with a little summary that can be gathered from rumors, to get a little context.

N23 - 40 CUs/192 bit bus.
N22 - 60 CUs/256 bit bus.
N21 - 80 CUs/384 bit bus.

Incorrect.

I forgot to add to that post, that N21 has 24 GB VRAM, N22 has 16 GB VRAM, N23 will have 12 GB VRAM and N24 will have 8 GB VRAM.

What is the source of your rumor? Moore's Law is Dead said exactly that yesterday in his last video....

He also said that each is "just behind the 3070 and 3080 counterparts with a significantly lower TDP, more memory and better prices. The HBM variant is to be used exclusively in data centers."
IMO seems then like the Halo SKU would probably have failed because they underestimated NV's will to raise the TDP so massively in this generation to get the performance.

 

[Glo.]

He also said that each is "just behind the 3070 and 3080 counterparts with a significantly lower TDP, more memory and better prices. The HBM variant is to be used exclusively in data centers."
IMO seems then like the Halo SKU would probably have failed because they underestimated NV's will to raise the TDP so massively in this generation to get the performance.

Yeah.

Navi 23 will be just behind RTX 3070, and Navi 22 will be just behind RTX 3080. That is correct.

Guys, use your logic. Look at RX 5700 XT and its performance per ALU/CU compared to Turing, add redesigned caches, which increases internal bandwidth massively, add 25% core clocks(2.3 GHz) and 10% IPC.

Then use those equations for 50% bigger GPU, with 60 CUs, and then for 2x bigger GPU in the mix.

Im sometimes baffled that even if people have those things so plain, in their faces, they fail to see them.

And Im not going to post where does the info about bus width, and memory sizes comes from.

If 40 CU RDNA1 GPU was able to trade blows with 40 SM Turing GPU, why would 40 CU GPU(RDNA2) with increased IPC lose to a 44 SM GPU that LOST IPC(Ampere)? Especially with masively increased clock speeds, on RDNA2 GPU?

P.S. I think we should stop calling ALUs in GPUs cores, and call SM's, CU's, EU's cores. That way will be factually correct, and easier for laymans to understand the performance of GPUs, instead of marketing gibberish that all of companies are trying to sell us.

 

[insertcarehere]

If 40 CU RDNA1 GPU was able to trade blows with 40 SM Turing GPU, why would 40 CU GPU(RDNA2) with increased IPC lose to a 44 SM GPU that LOST IPC(Ampere)? Especially with masively increased clock speeds, on RDNA2 GPU?

Yes, I clearly see from the benchmarks/claims that 48SM Turing (RTX 2080) is still faster than 46SM Ampere (RTX 3070), and 64SM Turing (RTX 2080ti) is faster than 64SM Ampere (RTX 3080), oh wait...

 

[Konan]

N22 - 60 CUs/256 bit bus.

Here is what Adored thinks about the 60 CU level -

But, what about the more affordable RTX 3070?, well, we don’t know as much about how that card performs, but taking NVIDIA at their word, it should be slightly faster than the RTX 2080 Ti, meaning its 20.4 TFLOPs is more like ~15 TFLOPs. At that rate, AMD could offer a cut-down variant of its Navi 2X GPU, a 60 CU variant would offer 3,840 SPs, and between 14 and 17 TFLOPs, again, using the console’s clock speeds as a template.

N21 - 80 CUs/384 bit bus.

and here is what they think about your 80CU level

That would mean the largest Navi 2X GPU would come in at 80 CUs (2x the 40 CUs of the 5700 XT’s GPU). If so, that would bring with it 5,120 SPs.
So, what kind of performance could we expect from something that big from AMD? Well, if the next-gen consoles’ clock speeds are anything to go by (1.8-2.23GHz), then the theoretical shader-TFLOPs of Navi 2X is likely somewhere between 18 and 23 TFLOPs, which is right where what AMD would need to compete with, and more importantly, potentially beat out the RTX 3080. As for the RTX 3090, well, it would likely get close at these numbers, but I believe AMD will likely target prices below $1,000, so it wouldn’t really be relevant to compare against directly. If we take AMD at their word, then theoretically, a graphics card based on the Navi 2X GPU should consume ~300W at most.

Navi 23 - RTX 2080 Super performance levels +10%, @150W TBP.
Navi 22 - 10-20% above RTX 2080 Ti performance levels @225 TBP.
Navi 21 - 40-50% above RTX 2080 Ti performance @ 275W TBP.
Those were performance targets for RDNA2, that I got hint few weeks ago.

Observations/Speculation to put something together...
The 3060 (and the Super/Ti variant) is going to be targeting 2080 to 2080S + levels. That will be your in and around your N23 level w/40CUs
The 3070 is suspected to be between 5-15% + of a 2080Ti. Nvidia quote "Faster than a 2080Ti" That will be around your N22 level estimate. Then it looks like the 3070S comes with 16GB GDDR6....
The 3080 looks like it will in and around your N21 level estimate. And who knows if there is an Super / Ti variant at this time.... Speculation - N21 = 384@16Gbps = 768GB/s and the 3080 is....

Edit: Then there is DLSS and RT and I am very much looking forward to seeing how AMD competes here.

 

[Bigos]

I would like to ask for any person that is certain about Big Navi ("N2x will") to tell their sources (so that we can verify them). It gets hard to discern wild delusions from facts here (or anything in between). Speculations are perfectly fine, just tell that they are...

I don't say that anything without sources is implicitly wrong. I just want to have a clear separation between facts and rumors.

 

[Helis4life]

I would like to ask for any person that is certain about Big Navi ("N2x will") to tell their sources (so that we can verify them). It gets hard to discern wild delusions from facts here (or anything in between). Speculations are perfectly fine, just tell that they are...

I don't say that anything without sources is implicitly wrong. I just want to have a clear separation between facts and rumors.

No one has any sources, at least not ones that have factual correct iformation. It is all just speculation and guesswork

 

[Glo.]

Here is what Adored thinks about the 60 CU level -





and here is what they think about your 80CU level




Observations to put something together...
The 3060 (and the Super/Ti variant) is going to be targeting 2080 to 2080S + levels. That will be your in and around your N23 level w/40CUs
The 3070 is suspected to be between 5-15% + of a 2080Ti. Nvidia quote "Faster than a 2080Ti" That will be around your N22 level estimate. Then it looks like the 3070S comes with 16GB GDDR6....
The 3080 looks like it will in and around your N21 level estimate. And who knows if there is an Super / Ti variant at this time.... Speculation - N21 = 384@16Gbps = 768GB/s and the 3080 is....

I would like to point to you out, that Adored did analysis on RTX 3070 in his video, and his analysis points to the reality that RTX 3070 is NOT GOING to achieve RTX 2080 Ti performance, and will be quite a few percent slower .

I don't want to sound rude towards Adored, and you, but using this logic, Compute performance should translate into 90% performance increase in games with RTX 3080 over RTX 2080 Ti.

And based on leaked benchmarks, on which even you commented, and liked posts about them, that is not going to happen.

Nvidia lost Performance per FLOPs with Ampere. AMD, according to them, is going to gain performance per FLOPs in games.

If anything previous generation of GPUs taught us is that we should never estimate the gaming performance based, SOLELY, on TFLOPs numbers.

 

[xpea]

Do you know when tape out happened for those GPUs?

If it happened in April does it make my stament wrong, considering that usually release is 9 months after a tape-out, and it is only 5 months since April?

G102 and 104 were tape out in December for a launch around Computex. But, and it's very unusual for Nvidia, they has to do a metal respin, which caused 3 months delay (gaming Ampere production silicon is revision A1, instead of A0 when everything goes well)

 

[exquisitechar]

If 40 CU RDNA1 GPU was able to trade blows with 40 SM Turing GPU, why would 40 CU GPU(RDNA2) with increased IPC lose to a 44 SM GPU that LOST IPC(Ampere)? Especially with masively increased clock speeds, on RDNA2 GPU?

Lost "IPC" (incorrect, let's call it "gaming performance per TFLOPS"), but doubled FP32 throughput. 40 SM Ampere is still >>> 40 SM Turing.

 

[Midwayman]

Rumor is PS5 announcement on the 9th.

Agreements?

Just that maybe they are waiting until the final launch announcement for ps5 and xbox. Probably unlikely, but who knows. I feel like the cat has been out of the bag awhile on the console specs. Its just price and timing.

 

[Shivansps]

This is not about "beliving" anything, it is just pure logic, Nvidia has already show his hand, there is nothing they can do if AMD comes out now with something better other than lower the prices and look very stupid by doing so. Not to mention is not a good idea to let rumors run free, they had to know this by now. So this "not wanting to share details" makes zero sence. They dont even have to share the prices, just a performance slide, thats all is needed to stop the Nvidia from taking all the spotlights.

So, something is going on, considering they already integrated RDNA2 on a console APU and we know that these integrations takes extra time, RDNA2 should be already on the street by now.

My question is... is AMD allowed to use RDNA2 for the general public? or consoles have to come out first and they have a exclusivity deal going on?

 

[A///]

Two different use cases.

I keep wanting to comment on RDNA2 but my guess is just as good as anyone else's. Someone on Crappit posted a header image on Computex back when it was canceled in May. End of September. I'm guessing we'll find out what we want to find out in the next 2-3 weeks?

I'm really happy I scored a platinum 850 watt PSU a few weeks ago. Sleep well, my precious.