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

 

[dzoni2k2]

Well that makes
- Igors Lab
- Coreteks
- VideoCardz

All sharing the same AIB information. Don’t think it’s attention craving.
Only Coreteks has shared his perf. opinion on this info so far.
And VideoCardz following up with

https://twitter.com/x/status/1313613983027466240

AIB partners know absolutely nothing about N21.

 

[Karnak]

AIB partners know absolutely nothing about N21.

And you know that easily because there are like literally zero leaks out there (and the biggest one was/is MacOS). Three weeks before the official announcement and roughly a month before the launch.

The more AiB know, the more leaks you get. It's just that simple. Because of that there were so many information about GA102 besides performance numbers (because drivers were missing) even weeks before the official launch.

 

[Konan]

And you know that easily because there are like literally zero leaks out there (and the biggest one was/is MacOS). Three weeks before the official announcement and roughly a month before the launch.

The more AiB know, the more leaks you get. It's just that simple. Because of that there were so many information about GA102 besides performance numbers (because drivers were missing) even weeks before the official launch.

Very true. KatCorgi on Twitter is just a Chiphell poster with AIB info that came months before proven.
It will be interesting to me if what Coreteks said in his second to last video is true in that his info comes from a Tier 1 AIB. I like how he explains there are varying levels of AIB partners and that Those levels are exposed to certain information and that he got his from a partner who is at the very top.

 

[Hitman928]

Seems like AMD is maybe pulling their own 'FE' version where they will be offering N21 'FE' version and no custom designs will be available initially and N22/N23 was given to AIBs to design cards around. Then AIB only receiving up to N22 thought this was "Big Navi" and the leaks have all been based around this GPU. After official release, or maybe unveiling, AIBs will then get N21 to start designing cards around to be released at some point in the future. This way AMD cuts down on leaks but still gives AIBs the opportunity to do custom designs for the higher volume cards.

No idea if this is true, but it at least sounds plausible and if true makes me think AMD feels like they have good reason to not want any leaks about their top end card before they are ready to announce it themselves.

 

[PhoBoChai]

And you know that easily because there are like literally zero leaks out there (and the biggest one was/is MacOS). Three weeks before the official announcement and roughly a month before the launch.

The more AiB know, the more leaks you get. It's just that simple. Because of that there were so many information about GA102 besides performance numbers (because drivers were missing) even weeks before the official launch.

No custom N21 card = no AIB info until close to launch.

Custom N22 card = AIBs have the chips, making the cards, testing the cards, they know most things about it way ahead of launch date.

Two of the common tests AIBs run are 3dMark and Ashes of the Benchmark. What leaked? Ashes. It's N22.

 

[eek2121]

So AMD targeted big Navi (500+mm^2 7nm card) to fight a 3070 (390mm^2 8nm) which itself is slower than the 2080ti card? They also only gained 50% performance with a new architecture and double the size card? I find that hard to believe. This would put RDNA2 at Vega levels of fail.

I don't see how AMD could've expected anything less than 2080ti + 30-40% at least, for a die shrink.

The 40CU @ 2.5 ghz will have performance similar to a 2080ti. It is looking like it may beat the 3070 because the 3070 is actually behind the 2080ti for some stuff.

AIB partners know absolutely nothing about N21.

Correct. The reason AMD is waiting to push out AIB designs is to control leaks/misinformation and provide a reference card. It also helps them meet demand at launch. Rather than sending out kits to the AIBs and wait for the turnaround, they can do it themselves.

 

[Martimus]

Correct. The reason AMD is waiting to push out AIB designs is to control leaks/misinformation and provide a reference card. It also helps them meet demand at launch. Rather than sending out kits to the AIBs and wait for the turnaround, they can do it themselves.

AMD doesn't make their own cards. Sapphire is the OEM that makes AMD branded cards. (At least they were last I checked, but I highly doubt AMD had suddenly purchased manufacturing facilities to make their own cards in the last few years).

 

[PhoBoChai]

AMD doesn't make their own cards. Sapphire is the OEM that makes AMD branded cards. (At least they were last I checked, but I highly doubt AMD had suddenly purchased manufacturing facilities to make their own cards in the last few years).

That's true, Sapphire is a very close partner with AMD, they make AMD's reference GPUs and Coolermaster makes AMD's reference coolers.

 

[Konan]

That's true, Sapphire is a very close partner with AMD, they make AMD's reference GPUs and Coolermaster makes AMD's reference coolers.

What about XFX? Don't they too?

 

[mohit9206]

I think whatever performance AMD targets whether its 3070 or 3080, it will be the same price as Nvidia but with more Vram. So 3080 competitor having 16GB and 3070 competitor having 12GB but same price. Although i am not sure if it is the right strategy as that is not the strategy that Zen followed.

 

[PhoBoChai]

What about XFX? Don't they too?

They don't produce reference cards for AMD, at least as far as public info goes.

 

[PhoBoChai]

I think whatever performance AMD targets whether its 3070 or 3080, it will be the same price as Nvidia but with more Vram. So 3080 competitor having 16GB and 3070 competitor having 12GB but same price. Although i am not sure if it is the right strategy as that is not the strategy that Zen followed.

Zen could price aggressively due to chiplet design saving them on cost and yield, along with high volumes.

These GPUs are large chips, much fewer per wafer, wafers that AMD is limited in meeting all their markets. I don't expect N21 GPUs to be cheap.

 

[Timorous]

These GPUs are large chips, much fewer per wafer, wafers that AMD is limited in meeting all their markets. I don't expect N21 GPUs to be cheap.

And this begs the question why go with 500mm^2 by using cache when you can get 80CUs + HBM in a 350mm^2 die if you go as dense as Renoir.

While that might cost more per GPU it would reduce the number of wafers required to manufacture their desired number of parts allowing for more wafers to be used for Zen 3 which I think is more important at the moment. With Intel flailing around this is a golden opportunity for AMD to build up long lasting OEM partnerships so making sure they can meet demand on that side of things should be of upmost importance.

 

[BlitzWulf]

What about XFX? Don't they too?

XFX is super leaky AMD would be crazy to give them any more info than they absolutely have to.

well since Videocards is trying to hit the brakes on the hype train let me try to shovel some more coal into the engine. Not saying I believe this chart at all but it's fun to speculate, and 3070 perf from N21 is kind of hard to wrap my head around.

https://www.chiphell.com/forum.php?mod=viewthread&tid=2264233&mobile=1

 

[PhoBoChai]

And this begs the question why go with 500mm^2 by using cache when you can get 80CUs + HBM in a 350mm^2 die if you go as dense as Renoir.

While that might cost more per GPU it would reduce the number of wafers required to manufacture their desired number of parts allowing for more wafers to be used for Zen 3 which I think is more important at the moment. With Intel flailing around this is a golden opportunity for AMD to build up long lasting OEM partnerships so making sure they can meet demand on that side of things should be of upmost importance.

4 stacks of HBM and interposer step might be overkill $ when a 256 bit bus and cheap GDDR6 (not even X) will suffice, if you just bloat the die a little for cache.

Going denser tends to reduce clocks too. In the end, it could be a 400mm2 vs 500mm2 chip, not that much in the grand scheme of things.

 

[moinmoin]

Pretty embarrassing positioning by VideoCardz.

What I took from it was that Frank is not guaranteeing Or publicly reassuring that there will be sufficient volume for everybody. At the same time he was denying that would be infinite supply which I don’t think WCCF tech was actually saying to begin with.

Either you are confident that there will be supply and you will say so oh you’re not confident they will be supply and you’ll be quiet.

I guess he is confident in the demand and confident in the demand outstripping the supply. Doesn't need to mean that there is little supply (which is always relative to the demand anyway, thus the mention of "infinite supply").

 

[Saylick]

https://www.chiphell.com/forum.php?mod=viewthread&tid=2264233&mobile=1

I know it's pure speculation, but that chart probably isn't too far off from the truth. They have the 3080 with a high 1700s clockspeed and I know that 3080s already hit 1900s, so giving it another 7-8% matches up with the 6900 XT at ~2 GHz nicely, exactly where I'd expect it to land.

 

[Timorous]

4 stacks of HBM and interposer step might be overkill $ when a 256 bit bus and cheap GDDR6 (not even X) will suffice, if you just bloat the die a little for cache.

Going denser tends to reduce clocks too. In the end, it could be a 400mm2 vs 500mm2 chip, not that much in the grand scheme of things.

Sure but it means a bigger GPU that will either have fewer numbers or will impact Zen 3 production.

Renoir shows you can have density and clocks without hotspot or power issues so I am not convinced that improving density is out of the question.

 

[Panino Manino]

The long RDNA2 die was the PS5's SOC.
Surprisingly this thing was not a joke and really exists. If you want to see more:

 

[eek2121]

XFX is super leaky AMD would be crazy to give them any more info than they absolutely have to.

well since Videocards is trying to hit the brakes on the hype train let me try to shovel some more coal into the engine. Not saying I believe this chart at all but it's fun to speculate, and 3070 perf from N21 is kind of hard to wrap my head around.

https://www.chiphell.com/forum.php?mod=viewthread&tid=2264233&mobile=1

It may be a bit more accurate than you think. 😉

EDIT: RTG welcomes all the naysayers to the new AMD, including Jensen. I wonder how NVIDIA will respond. 😆

 

[Heartbreaker]

View attachment 31246

The long RDNA2 die was the PS5's SOC.
Surprisingly this thing was not a joke and really exists. If you want to see more:

That isn't the same die as the one that was claiming to be Big Navi. Proportions are different.

 

[eek2121]

Chiphell speculation, things were a bit confusing via Google Translate, but I think I got it - feel free to suggest corrections:

Radeon 6900XT - 80CU, 16 GB 256 bit GDDR6, 340w TDP - 2240 MHz
Radeon 6900 - 72CU, 16 GB 256 bit GDDR6, 290W TDP - 2050 MHz
Radeon 6800XT - ~64CU, 12 GB 192 bit GDDR6, 270W TDP - 2150 MHz
Radeon 6800 - 56CU, 192 bit GDDR6, 255W TDP - 2300 MHz
Radeon 6700XT - 50CU, 192 bit GDDR6, 240W TDP - 2300 MHz
Radeon 6700 - 40CU, 192 bit GDDR6, 195W TDP - 2300 MHz

Navi 21 - 6900XT, 6900, 6800 XT
Navi 22 - 6800, 6700XT
Navi 23 - 6700, etc.

Looks spot on to me. I did not believe AMD would differentiate XT and non XT chips via CU count, but I guess it's a possibility.

EDIT, it would not surprise me if all of the cards below the 6900 had 12 GB VRAM with the exception off the 6700 or 6600. The 6600 XT will probably be 36CU, and the 6500 will be Navi24.

EDIT #2: These chips have quite a bit of clock headroom, so I expect AMD may have their own "super" variants if NVIDIA manages to push something out that is faster. Radeon 6950XT @ 2.33-2.5 GHz! Given the fact I had the "old" Radeon 6950, I hope they do this

 

[Gideon]

The long RDNA2 die was the PS5's SOC.
Surprisingly this thing was not a joke and really exists. If you want to see more:

Looks really well built and even uses liquid metal TIM. Looking at the disassembly process and comparing it to Xbox Series S. This looks both considerably more expensive and much more laborious to assemble. It should be really quiet though.

 

[eek2121]

Chiphell speculation, things were a bit confusing via Google Translate, but I think I got it - feel free to suggest corrections:

Radeon 6900XT - 80CU, 16 GB 256 bit GDDR6, 340w TDP - 2240 MHz
Radeon 6900 - 72CU, 16 GB 256 bit GDDR6, 290W TDP - 2050 MHz
Radeon 6800XT - ~64CU, 12 GB 192 bit GDDR6, 270W TDP - 2150 MHz
Radeon 6800 - 56CU, 192 bit GDDR6, 255W TDP - 2300 MHz
Radeon 6700XT - 50CU, 192 bit GDDR6, 240W TDP - 2300 MHz
Radeon 6700 - 40CU, 192 bit GDDR6, 195W TDP - 2300 MHz

Navi 21 - 6900XT, 6900, 6800 XT
Navi 22 - 6800, 6700XT
Navi 23 - 6700, etc.

Looks spot on to me. I did not believe AMD would differentiate XT and non XT chips via CU count, but I guess it's a possibility.

EDIT, it would not surprise me if all of the cards below the 6900 had 12 GB VRAM with the exception off the 6700 or 6600. The 6600 XT will probably be 36CU, and the 6500 will be Navi24.

EDIT #2: These chips have quite a bit of clock headroom, so I expect AMD may have their own "super" variants if NVIDIA manages to push something out that is faster. Radeon 6950XT @ 2.33-2.5 GHz! Given the fact I had the "old" Radeon 6950, I hope they do this

One more thought here, if this speculation is close to accurate, I'm willing to bet the 6900XT will be priced under $1,000, and I dare say they'll probably price it at $699 or $799.

EDIT: A guess at pricing based on this speculation.

6900 XT - $699 - (3090 perf)
6900 - $649 - (3080 perf)
6800 XT - $549 (3070 perf)
6800 - $499
6700XT - $449
6700 - $399

 

[CastleBravo]

Anyone else think AMD will show us some RDNA2 stuff tomorrow? If Zen3 really beats Intel at gaming performance, they will want to pair it with a high end GPU to show it off, and I don't see them using an Nvidia product for that demo.