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

 

[Kepler_L2]

XT is 80 CUs, I don't know why people keep thinking its 72.

 

[Krteq]

72 CUs, 255W TDP, 2.4 GHz Game Clock is too good to be true... That's literally like 2x perf/W. I'm not going to keep my hopes up for that tbh.

Also add 128 (NAVI2) ROPs

But I'm quite skeptical about this... if that's true than Suzanne Plummer's team must be full of magicians

... but you know, Ryzen/next-gen semi-customs...

 

[Glo.]

XT is 80 CUs, I don't know why people keep thinking its 72.

XTX is 80 CU. XT is 72.

 

[Kepler_L2]

XTX is 80 CU. XT is 72.

XT codename literally means full die. All variants like XTX (higher TDP/clocks), XTL(lower clocks) are still full die.

 

[eek2121]

Well, they already said there is an "XTX," and I just love that this nonsense is being resurrected. Hopefully, it means a true XT > XTX > XTXTX > XTXXTTTXXX sort of thing and that there is real performance (like actually used to happen) attached to the nonsense (and so it is defensible), but I doubt it is more than a 3090 competitor, whatever that means (price, performance, efficiency, etc.)

and whenever it arrives, with whatever Nvidia has to throw it at that unannounced time.

It is very likely that 'XTX' refers to the die, AMD will likely have XT and non-XT variants.

AMD also will likely hold at least a few SKUs back to deal with any response NVIDIA has.

I personally am hoping for a Radeon 6950XT. I own a Radeon HD 6950, so it would be fitting that I purchase a 6950 XT.

 

[Glo.]

XT codename literally means full die. All variants like XTX (higher TDP/clocks), XTL(lower clocks) are still full die.

If its 80 CU die, and has 2.4 GHz GAME CLOCK, at 255 TGP then this is freaking the most efficient GPU architecture, that has ever been made by any company.

 

[eek2121]

XT is 80 CUs, I don't know why people keep thinking its 72.

The largest variant is 80CUs. The next largest is 72 CUs. GPUs are more prone to defects than CPUs (especially chiplet based CPUs) so AMD has to have multiple SKUs for every die. Remember that it isn't just about defective CUs, but CUs that don't hit clock speed targets.

Yields will determine whether the full fat 80CU part gets launched right away or not.

A Radeon GPU die is 4X (or more) more likely to have a defect than a Ryzen chiplet.

 

[Kepler_L2]

The largest variant is 80CUs. The next largest is 72 CUs. GPUs are more prone to defects than CPUs (especially chiplet based CPUs) so AMD has to have multiple SKUs for every die. Remember that it isn't just about defective CUs, but CUs that don't hit clock speed targets.

Yields will determine whether the full fat 80CU part gets launched right away or not.

A Radeon GPU die is 4X (or more) more likely to have a defect than a Ryzen chiplet.

Yes they have 4 SKUs, XT and XTX are full die (80 CU), XL is 72 CU and XE is probably 64 CU trash dies and likely OEM only.

 

[eek2121]

If its 80 CU die, and has 2.4 GHz GAME CLOCK, at 255 TGP then this is freaking the most efficient GPU architecture, that has ever been made by any company.

It's entirely possible, it's just nuts. I've been told that they've managed to get in-lab silicon to almost 3 GHz. Obviously shipping parts won't reach that. What is likely helping them a lot are the refinements they made to the 7nm process when the Ryzen XT chips launched.

I will be completely shocked if these chips don't perform amazingly well. Clearly AMD isn't concerned about memory bandwidth since they didn't increase the bus size (they decreased it for the 40CU part even). They reworked other areas of the GPU that could have caused scaling issues.

EDIT: Jensen is going to have a heart attack.

 

[Glo.]

The largest variant is 80CUs. The next largest is 72 CUs. GPUs are more prone to defects than CPUs (especially chiplet based CPUs) so AMD has to have multiple SKUs for every die. Remember that it isn't just about defective CUs, but CUs that don't hit clock speed targets.

Yields will determine whether the full fat 80CU part gets launched right away or not.

A Radeon GPU die is 4X (or more) more likely to have a defect than a Ryzen chiplet.

No. He is right. XT always has been the full die, and XTX was extreme version of that Full die.

RX 5700 XT Lisa Su edition was XTX version. RX 5700 XT was Navi 10 XT.

This is the actual FULL DIE.

This makes it actually even more mind blowing, guys. AMD actually have pulled the greatest physical design ever.

Yes they have 4 SKUs, XT and XTX are full die (80 CU), XL is 72 CU and XE is probably 64 CU trash dies and likely OEM only.

XE is supposedly coming next year.

 

[Kepler_L2]

XE is supposedly coming next year.

Correct, they need to get enough trash dies to release the lowest SKU first and 7nm yields are too good

 

[Glo.]

Yes, guys. Kepler_L2 is correct.

Historically XT and XTX are full dies. They just differ by clock speeds, and power targets.

So its not 72 CUs. Its full, 80 CUs, with 2.4 GHz game clock, at 255W of power drawn(hence the TGP).

Im lost. Its GG. Its beyond EVERYTHING we thought it could possibly be.

 

[jpiniero]

If that was really true, it should be faster than even the 3090. Bandwidth limited perhaps?

 

[TESKATLIPOKA]

Then the question is, which GPU's scores did they show us during Zen3 presentation and we don't know If that tweet is correct or not.

 

[Glo.]

Then the question is, which GPU's scores did they show us during Zen3 presentation.

If it was just for demo, they could've used ANY configuration in the die, with ANY clock speeds.

They could've used Navi 21 Die, with special, "demo" BIOS, that was specifically designed to blatantly sandbag its performance.

Im not saying this is what happened. But its possible, after all.

 

[maddie]

It's entirely possible, it's just nuts. I've been told that they've managed to get in-lab silicon to almost 3 GHz. Obviously shipping parts won't reach that. What is likely helping them a lot are the refinements they made to the 7nm process when the Ryzen XT chips launched.

I will be completely shocked if these chips don't perform amazingly well. Clearly AMD isn't concerned about memory bandwidth since they didn't increase the bus size (they decreased it for the 40CU part even). They reworked other areas of the GPU that could have caused scaling issues.

EDIT: Jensen is going to have a heart attack.

They did claim reduced complexity and switching power. Simpler circuits doing the same job as before, so less transistors, etc, should allow for higher speed limits.

 

[Glo.]

speed limits.

Ah yes. The hype train ignored the 50 MPH speed limit and travels now at 200 MPH!

 

[moinmoin]

In another tweet he's saying that's not for the reference card but rather some AiB.

https://twitter.com/x/status/1317538064034562050

If the source was an AiB surely all the rumor mongers who claimed having AiB sources before should be able to get up-to-date?

AMD can't keep getting away with them sandbags!

 

[Glo.]

https://twitter.com/x/status/1317576922621149184

Oh and last thing for today: the 2.4GHz game clock was from before Ampere launched. Goodnight

Ok. Can somebody stop this train? We're approaching Earth's Orbit right now!

 

[NTMBK]

It was already pretty crazy that the PS5 GPU was hitting 2.2GHz... if Big Navi is even higher clocked I will be stunned.

 

[uzzi38]

Redacted it these people aren't letting me sleep.

https://twitter.com/x/status/1317584994156617729

Profanity is not allowed in the tech sub-forums.

AT Mod Usandthem

 

[zinfamous]

I need XFX to come out with a RX6900XTX XTRA THICCC Pro III triple slot cooler version.

 

[Saylick]

This hype train is more like a hype missile. It just can't be stopped. These clock speeds just defy all logic (pun intended)!

https://twitter.com/x/status/1317572881233182720

https://twitter.com/x/status/1317578852969963520

 

[dzoni2k2]

Goddamn it these people aren't letting me sleep.

https://twitter.com/x/status/1317584994156617729

What will they price and position the XL against? 3070?

 

[uzzi38]

What will they price and position the XL against? 3070?

Wild guess says 3070+15% perf for $550.

Value for money usually degrades as you increase performance, so offering 15% more performance at 10% extra cost will sound like an amazing deal to some.