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

 

[moinmoin]

I still believe that the strategy of having a big chunk of the die dedicated to just RT and DLSS is not sustainable in the long run. It locks nvidia to big die chips and is detrimental to general rasterization performance. Especially with consoles going the RDNA2 way.

It seems Nvidia wanted to avoid creating GPUs optimized for gaming, so reused the same dies as in its datacenter chips. Now Nvidia's datacenter business is actually surpassing its gaming business in revenue, while "suddenly" facing pressure on both price and performance by the competition in the latter. Nvidia has to react to it by serving both markets separately as AMD already is underway doing with RDNA and CDNA. It's only a question of how long it takes Nvidia (my guess at least 2 years unless effort is already underway) and how much worth it puts on the gaming market (which is as mentioned of decreasing importance revenue wise).

Well played by AMD making the RDNA/CDNA split even before it became clear how much it's able to catch up.

 

[Kepler_L2]

It only makes sense if N22 is HBM, designed with density in mind, and targeting lower frequencies to make a mobile beast

N22 is GDDR6, there are Linux patches for G6 memory training.

 

[Veradun]

N22 is GDDR6, there are Linux patches for G6 memory training.

The idea still stands even if less energy and size efficient :>

 

[zinfamous]

Didn't they just delay that game another month? Its like some dude if that nut on the ground came from his Harley.

yeah, that prompted me to ask that question, and also before I watched the Hitler video. It would be great timing for those X-Mas/New Year's bundles and with the constant delays, I assume some gamers have gotten salty over Cyberpunk, and it might be a good idea for CDPR to look into a bundle at or very soon after release.

The release date is now what, 1 day before or after 6800XT releases?

 

[zinfamous]

Any news yet on how N21 will mine?

BAN!

 

[raghu78]

So according to you a 40CU N22 has 50% more bandwidth and 3x more cache compared to a 32CU N23 while having only 1/3 more CUs?
So according to you a 32CU N23 has the same amount of Infinity cache as a 8-12CU APUS?
So according to you a 16-20CU N24 has only half of the infinity cache of an 8-12CU APU?
Are you so sure about all this?

I am only saying what cache sizes are possible based on what we know about Navi 10 die size of 251 sq mm with 40 CU and 256 bit GDDR6. The Zen 2 L3 cache was roughly 16-17 sq mm for 16 MB L3. The math suggests N23 cannot have 64 MB Infinity Cache with 128b it G6 and only 32 MB. Similarly N22 is large enough at 340 sq mm to fit 96 MB Infinity Cache with 192 bit G6. The RDNA2 APUs can fit 32 MB Infinity Cache given the Renoir die size and the size of Zen 2 core and the L3 cache.

https://www.reddit.com/r/Amd/comments/dw7msh

As always we can only be reasonably confident given what we know already about Renoir, Navi 10 die size. N24 die size is unknown but given that AMD would want a RDNA2 GPU die to sell graphics cards at $129-$169 my guess is the die size of that GPU is <= 150 sq mm which rules out even 32 MB Infinity cache.

 

[TESKATLIPOKA]

The demand for the cache is primarily driven by display resolution, as that determines the size of the render targets, not by how much computing power the card has. The cards are therefore targeted at specific resolutions. 128MB is for ~ 4k, 64MB ~ 1440p, 32MB ~ 1080p, 16MB ~ 720p. The wildcard is the N22, which has a 192-bit bus -- there might be 96MB on it if the cache needs to match up with the memory channels.

There is a 100m^2 difference in size between N22 and N23, with only a 8CU difference, while N23 has half the bus width of N10, yet is not that much smaller than it (so it's not cacheless). I also think the numbers lean very strongly to what Raghu posted.

N24 will probably just be a very weak bottom-tier card.

Between N22 and N23 is not a big difference, just 8CU or 25% more for N22. So I don't see any reason why N23 should have only 32MB cache while N22 has 96MB along with 50% wider memory bus. It's possible that 96MB is because of 192bit bus, but then N23 should have 64MB instead of 32MB because It has 128bit memory bus and N24 should have 32MB instead of 16MB because It has 64bit memory bus. Navi23 is suppossedly 235mm2, but even that is 77mm2 bigger than Navi14(158mm2), while having 8 more CU + infinity cache. You can say the difference in die size is not enough for that, but we actually don't know transistor density of RDNA2, It could be the same as RDNA1 or more.
N24 is a bottom tier card, but It should have more CU than Van Gogh(8CU) or Rebrant(12CU) IGP, so I don't see a good reason why It should have less infinity cache than any of them.
Do you have any official die size info of any RDNA2 product? I don't.

 

[MrTeal]

I would guess it would mine terribly for the price. The IC probably won't help in mining, and a 256-bit G6 bus on a $600 card is going to be junk value.

So, yay.

 

[Glo.]

According to you, a 2.5GHz 40CU RDNA2 part with +20% IPC in relation with RDNA1 would be at +15% of a 2080Ti, which means that if RDNA2 has no IPC benefit vs RDNA1 then we would have:
2080Ti=100
+20% IPC part= 100+15%= 115
0 IPC part= 115/120*100= 95.8
In the last 3 weeks you changed the tune saying that it will be at 2080S+10% level which is way slower than 95.8 (Navi22 with 0 IPC improvements versus RDNA1 according with your calculations)
Who is saying again that RDNA2 lost IPC?
I really didn't expect all this controversy, I redone my calculations carefully just to check exactly and I revised my predictions by just 1% so really no essential change:
2.5GHz 40CU part is 87% 3070 at 4K (86% in my first post)
and regarding the 2.3GHz 32CU part, 5700 will only be 7-8% slower than this part at QHD (6-7% in my first post)
Regarding IPC, 2.25GHz 72CU 6800XT = 2 X 1.9GHz 40CU 5700XT according to AMD, so RDNA2 lost IPC?
This was a rhetorical question...
I feel confident regarding the above predictions, but if I'm wrong I'll have no problem admitting it when the reviews come.
In the mean time let's agree to disagree.

You still don't get it?

If your calculations are correct, 40 CU RDNA2 GPU having 87% of performance of RTX 3070 will lose IPC compared to RX 6800XT and 6900XT.

And that is flawed logic, mainly because in your calculations you do not take into account for scaling.

40 CU GPU Clocked at 2.2 GHz will not be 50% of performance of 80 CU 2.2 GHz RDNA2 GPU.

Why? Two things. Scaling. 80 CU die may have 90% of CU scaling, compared to 40 CU die, so it alone effectively pushed performance target higher. Secondly, we don't know how 192 bit bus affects performance.

Using basic, armchair logic, for 40 CU die to be exactly 50% of 80 CU die, it would have to have 128 bit bus. But it doesn't. It has 192 bit bus.

So no. Just taking into accounts, those two factors, 40 CU GPU, clocked at 2.5 GHz(which it won't ) will not be "only" 87% of RTX 3070.

Even RedGamingTech info says that it will be about the same as RTX 3070 in performance.

So I would trust a guys who leaked Infinity Cache name, its size, etc way before everyone thought it is real .

 

[Glo.]

Between N22 and N23 is not a big difference, just 8CU or 25% more for N22. So I don't see any reason why N23 should have only 32MB cache while N22 has 96MB along with 50% wider memory bus. It's possible that 96MB is because of 192bit bus, but then N23 should have 64MB instead of 32MB because It has 128bit memory bus and N24 should have 32MB instead of 16MB because It has 64bit memory bus. Navi23 is suppossedly 235mm2, but even that is 77mm2 bigger than Navi14(158mm2), while having 8 more CU + infinity cache. You can say the difference in die size is not enough for that, but we actually don't know transistor density of RDNA2, It could be the same as RDNA1 or more.
N24 is a bottom tier card, but It should have more CU than Van Gogh(8CU) or Rebrant(12CU) IGP, so I don't see a good reason why It should have less infinity cache than any of them.
Do you have any official die size info of any RDNA2 product? I don't.

There is one reason for it.

Layouts of the dies.

N22 may be almost 50% of N21. But N23 is not. Ergo. We have to wait. N23 is the biggest mystery of them all.

 

[Glo.]

My question was how sure we were on the specs. I think someone else said it was taken from Apple software updates that have checked out, so I'll trust it unless there's a good reason to think maybe Apple isn't getting any of the full die cards so don't have an entry, but Illthat's unlikely.



Why assume it's below? I'll admit that there's certainly room since there aren't any N21 cards with disabled cache or memory controllers which is a little suspect.

However, there's also no 6900 and there's a big price hole between the 6800 XT and 6900 XT that AMD might want to hold off on filling until they see what NVidia will do. It's odd that there aren't any 80 CU parts that can't hit 6900XT clocks. Sure many can be made into serviceable 6800XT, but I think there's another bin in there. An 80 CU part with a 1915 MHz game clock is about 5% on either side of the two XT cards. Sounds like a possible $750 card.

Why assume its below 6800?

Well, we have seen that in the typical order: 6900XT is XTX die, 6800 XT is XT die, 6800 is XL die, so by that reason XE should be below 6800.

 

[TESKATLIPOKA]

I am only saying what cache sizes are possible based on what we know about Navi 10 die size of 251 sq mm with 40 CU and 256 bit GDDR6. The Zen 2 L3 cache was roughly 16-17 sq mm for 16 MB L3. The math suggests N23 cannot have 64 MB Infinity Cache with 128b it G6 and only 32 MB. Similarly N22 is large enough at 340 sq mm to fit 96 MB Infinity Cache with 192 bit G6. The RDNA2 APUs can fit 32 MB Infinity Cache given the Renoir die size and the size of Zen 2 core and the L3 cache.

https://www.reddit.com/r/Amd/comments/dw7msh

As always we can only be reasonably confident given what we know already about Renoir, Navi 10 die size. N24 die size is unknown but given that AMD would want a RDNA2 GPU die to sell graphics cards at $129-$169 my guess is the die size of that GPU is <= 150 sq mm which rules out even 32 MB Infinity cache.

Please also read my reply to Tuna-Fish.
We don't know the actual die sizes of RDNA2 or their transistor density so this is just speculation.
I don't think N24 would be slower than Navi14. I think N24 could have 20-24CU and coupled with much higher clocks It will end up faster than Navi14. With this in mind It's not possible that AMD would ask only $129-$169 for such card, much less If they put 8GB Vram on It. Performance would be also enough for Full HD and for that 16MB infinity cache paire with onlz 64bit memory bus is most likely not enough.

 

[Glo.]

https://www.amd.com/en/gaming/graphics-gaming-benchmarks

Uhhh, has anyone seen these benchmarks?

What the hell?

 

[Midwayman]

Custom 6900XT might happen

AMD in talks with AIBs to allow custom Radeon RX 6900 XT graphics cards - VideoCardz.com

AMD is allegedly considering allowing its board partners to launch custom Radeon RX 6900 XT graphics cards. Custom AMD Radeon RX 6900 XT might happen Prior to launch, we reported that AMD had not confirmed to AIBs the existence of Radeon RX 6900 XT SKU, although the AIBs we spoke with already...

videocardz.com

Some additional benchmarks are out.

RX 6800, 6800 XT und 6900 XT: AMD veröffentlicht weitere Benchmarks in 4K und WQHD

Nach der Präsentation der Radeon RX 6800, 6800 XT und 6900 XT hat AMD nun weitere Benchmarks der Grafikkarten in 4K und WQHD veröffentlicht.

www.computerbase.de

 

[CastleBravo]

Don't know if anyone posted this yet, but AMD's website has detailed benchmark data for both 1440p and 4k on all three GPUs. At 1440p, the 6800 XT looks like a match for the 3090, and the 6900 XT is even faster. If these cards overclock to 2.4Ghz+, wow...

https://www.amd.com/en/gaming/graphics-gaming-benchmarks

EDIT: damn, beat by one minute

 

[MrTeal]

They only ever list one API. Perhaps the discrepancy to the presentation slides are that this is the best API for AMD and not necessarily the best one for the RTX cards?

 

[Hitman928]

They only ever list one API. Perhaps the discrepancy to the presentation slides are that this is the best API for AMD and not necessarily the best one for the RTX cards?

That is a good point. Also, SAM is on in those benchmarks, most likely Rage mode as well. In their presentation they never showed the 6800XT with those on in direct comparison to the 3080 so all of that is probably what is making the difference. I expect the reviewed results to be much tighter between the respective card tiers. Who knows exactly where it's going to land, their 5700XT presentation results were actually slightly lower than most reviews showed, so maybe a bit of a positive surprise is waiting here as well. Either way, the main thing is, AMD is competing again with Nvidia across the full GPU stack. That hasn't happened in a long time and is really nice to see.

 

[Zepp]

It seems Nvidia wanted to avoid creating GPUs optimized for gaming, so reused the same dies as in its datacenter chips. Now Nvidia's datacenter business is actually surpassing its gaming business in revenue, while "suddenly" facing pressure on both price and performance by the competition in the latter. Nvidia has to react to it by serving both markets separately as AMD already is underway doing with RDNA and CDNA. It's only a question of how long it takes Nvidia (my guess at least 2 years unless effort is already underway) and how much worth it puts on the gaming market (which is as mentioned of decreasing importance revenue wise).

people say Nvidia hasn't been slacking like Intel regarding AMD's competition but Ampere seems like a clear example of exactly that. RTX and DLSS always seemed like heavily marketed by-products of Nvidia's data-center engineering.

 

[pj-]

https://www.amd.com/en/gaming/graphics-gaming-benchmarks

Uhhh, has anyone seen these benchmarks?

What the hell?

Very mixed results at 4k. It's also a bit questionable to have results with SAM enabled since most people will not be using these with 5000 series ryzen

 

[Mopetar]

Why assume its below 6800?

Well, we have seen that in the typical order: 6900XT is XTX die, 6800 XT is XT die, 6800 is XL die, so by that reason XE should be below 6800.

I don't see why this has to be the case. Compare N21 with Cypress (old 5000 series) which had XT, PRO, and LE that were the 5870, 5850, and 5830 respectively.

If you look at those cards compared to N21 the similarities are eerie. Cypress had 3 bins with 100/90/70 percent of hardware enabled. Navi21 is 100/90/75. The price spread for Cypress was 100/68/63 percent relative to full die cost. N21 is 100/65/58 which is similar.

AMD hasn't a die with 4 bins (I'm excluding refreshes where the same die on a much mature process was given a new designation like Tahiti Pro2) since RV770, which is a decade ago. It may be possible that no 4th bin exists at all. The TSMC 7nm is reported as being quite mature and having a low defect rate relative to other processes at this point in its lifecycle, so an even more cutdown die feels a little less likely.

Maybe we eventually do see something that's a lowest common denominator salvage die, but it would have to be something with reduced memory controllers or infinity cache or a disabled shader engine because it's hard to imagine too many of those existing and at some point the bin is bad enough that even a high clock N22 will come close to it. If the yields are good enough maybe that part is so rare that AMD just scraps them or saves them for some special OEM-only part.

 

[soresu]

RTX and DLSS always seemed like heavily marketed by-products of Nvidia's data-center engineering

DLSS is software - previously run on the regular compute shaders and now on the tensor cores.

RTX is a marketing term for their RT acceleration and associated middleware code (GameWorks RT Super Shiny edition).

 

[MrTeal]

Very mixed results at 4k. It's also a bit questionable to have results with SAM enabled since most people will not be using these with 5000 series ryzen

We'll have to see how Zen3 benchmarks when it hits shelves. If it does provide clear leadership over Intel in gaming, I would imagine a good number of people running 3090/6900XT or even 3080/6800XT would be on the platform that provides the best results.

 

[MrTeal]

I don't see why this has to be the case. Compare N21 with Cypress (old 5000 series) which had XT, PRO, and LE that were the 5870, 5850, and 5830 respectively.

If you look at those cards compared to N21 the similarities are eerie. Cypress had 3 bins with 100/90/70 percent of hardware enabled. Navi21 is 100/90/75. The price spread for Cypress was 100/68/63 percent relative to full die cost. N21 is 100/65/58 which is similar.

AMD hasn't a die with 4 bins (I'm excluding refreshes where the same die on a much mature process was given a new designation like Tahiti Pro2) since RV770, which is a decade ago. It may be possible that no 4th bin exists at all. The TSMC 7nm is reported as being quite mature and having a low defect rate relative to other processes at this point in its lifecycle, so an even more cutdown die feels a little less likely.

Maybe we eventually do see something that's a lowest common denominator salvage die, but it would have to be something with reduced memory controllers or infinity cache or a disabled shader engine because it's hard to imagine too many of those existing and at some point the bin is bad enough that even a high clock N22 will come close to it. If the yields are good enough maybe that part is so rare that AMD just scraps them or saves them for some special OEM-only part.

The 5830 also didn't launch until 5 months after the 5870 and 4 months after the 5770 to plug a hole in the lineup, so by that point they would have had a good supply of bad dies on the relatively new and notoriously picky 40nm node.
I think the fact that AMD is leading with a fully enabled 80CU die is pretty indicative that yields are at the least reasonable.

 

[pj-]

We'll have to see how Zen3 benchmarks when it hits shelves. If it does provide clear leadership over Intel in gaming, I would imagine a good number of people running 3090/6900XT or even 3080/6800XT would be on the platform that provides the best results.

I will be one of them but I don't think it will be even close to a majority, especially if the leadership over intel is minor or game dependent.

 

[Veradun]

Very mixed results at 4k. It's also a bit questionable to have results with SAM enabled since most people will not be using these with 5000 series ryzen

I would assume people on the market for $$$$$ cards are the same people that often upgrade to achieve plaid FPS zone, thus buying the best CPU and not sticking to the old one :>