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

 

[beginner99]

Why do ppl disrespect Raja? Read the Anandtech article on his return to AMD. Mark Papermaster asked him to develop a post-GCN, brand new architecture. Very similar goals to Jim Keller being hired back to AMD, post-FX CPU new architecture, Zen.

Yet Tiger Lake with Xe iGPU is trading blows with Zen 2 APUs in performance. In a single generation Intel manage to come back from what, 3 gens behind on graphics performance?

As other have said due to the overhyping (poor volta) and the obvious powergames he was playing (RTG) made him very unlikable.

As for tigerlake iGPU, performance per die size and performance/watt matter too and it's clear AMD has the clear lead in that area. In fact I think Intel is going the wrong way with too big iGPUs. AMD with renoir for once made a smart choice focusing on CPU vs GPU. AMD could easily add CUs and stay within sane power use but they preferred a smaller die.

 

[pandemonium]

... I think the "Rage mode" thing would make sense in that AMD could trumpet their perf/w lead using one mode, and their performance metrics using the other. Best of both worlds.

So what you're saying is we're going to have "vtec just kicked in, yo" for our GPUs? I'm down with that. (Joking, of course. Turbo modes/Boost clocks could be just as easily inferred as a more aggressive cam, er, uh, voltage/frequency profile.)

 

[Gideon]

As other have said due to the overhyping (poor volta) and the obvious powergames he was playing (RTG) made him very unlikable.

As for tigerlake iGPU, performance per die size and performance/watt matter too and it's clear AMD has the clear lead in that area. In fact I think Intel is going the wrong way with too big iGPUs. AMD with renoir for once made a smart choice focusing on CPU vs GPU. AMD could easily add CUs and stay within sane power use but they preferred a smaller die.

Yeah pretty much it. Plus his performances on events were always full cringe.

Raja went to Intel and ppl laughed it off, that Intel has no chance to be competitive in graphics since they were so far behind. Yet Tiger Lake with Xe iGPU is trading blows with Zen 2 APUs in performance. In a single generation Intel manage to come back from what, 3 gens behind on graphics performance?

It's so weird to see people criticize GCN iterations as if it was Raja's fault. Those were already on the roadmap, well under way without his input.

There are some double standards here. Xe was also well underway, before Raja joined. Can't have it both ways.

 

[Veradun]

80, 72 and 64 CUs. As was speculated by many for a long time already.

The multiplication table for the number 8 is quite simple

 

[Veradun]

I'm pretty sure it's another BS guess. Pretty reliable leakers have confirmed its 80, 72, 72. It also makes more sense since AMD never cuts a full die by that much. We'll see.

Oh god why did I even bother reading nonsense from Coreteks. The guy is saying a 64CU RX6800 with 1820MHz game clock (funny how it's the same as XSX) pulls 290W. Meanwhile a 52CU XSX peaks at 145W lol.

But does it have a traversal co-processor?

 

[PhoBoChai]

There are some double standards here. Xe was also well underway, before Raja joined. Can't have it both ways.

2017. Intel has a ton of funding for their graphics division, unlike AMD. That's enough time for a late 2020 release of early Xe, with more to come in 2021. In fact, when he joined Intel in 2017, people discussed 2021 time frames as when we'll get their new architecture.

 

[BlitzWulf]

https://www.igorslab.de/en/the-can-...-works-oc-may-maybe-a-glance-in-bios-details/


I couldnt fully grasp everything through the machine translation but it looks like big navi could be quite an overclocker.

 

[PhoBoChai]

https://www.igorslab.de/en/the-can-...-works-oc-may-maybe-a-glance-in-bios-details/


I couldnt fully grasp everything through the machine translation but it looks like big navi could be quite an overclocker.

Looks like some AIB custom 6800XT (N22) has boost clocks set to 2.57ghz!

 

[dzoni2k2]

This leaks are all over the place.

 

[dacostafilipe]

Looks like some AIB custom 6800XT (N22) has boost clocks set to 2.57ghz!

6800 should be N21

 

[HurleyBird]

As other have said due to the overhyping (poor volta) and the obvious powergames he was playing (RTG) made him very unlikable.

What makes him unlikeable to me is his pathological self-promotion. Besides being an annoying quality in and of itself, combined with his extremely poor track record heading RTG it makes me seriously question his ability. Did he climb the corporate ladder because he's competent, or is he just good at playing the corporate game? The same traits that helped Raja gain his limited celebrity are also the traits that would help a rung-climbing career-minded corporate sociopath attain a powerful position.

Under Raja's leadership it's not just that RTG took a swing and a miss ala. R600 or FX. They both failed to properly innovate on GCN while simultaneously falling on their faces whenever they did try anything novel (eg. primitive shaders). Sure, AMD doesn't have the resources that Nvidia does, and AMD was more focused on CPUs but, honest to God question, if you could go back in time and replace Raja with Keller (or the GPU equivalent), and then cut the already limited GPU budget in half do you think you would end up with a better performing RTG or a worse one?

Even without pointing out how R&D giants like Intel can flounder while plucky startups sometimes make magic, I think we all instinctively know the answer to that one. You do good engineering by having good engineers and good leadership.

Occam's razor here says that the only thing Raja is especially skilled at is promoting Raja.

 

[Dribble]

Why do ppl disrespect Raja? ....

It's all forum psychology, he is a scapegoat. AMD didn't deliver a card that could beat Nvidia while he was there and he has now left. Putting all the blame on him means that people can hope RDNA2 can now beat Nvidia because he's not in charge. Where as in-fact he was just a cog in a very big wheel and him being there, or not being there will have in the end made a very small difference to how RDNA2 turns out.

 

[mohit9206]

A source close to me (u can't really see it) has said 3 variants of Navi 21. 80CU Big Navi, 160CU Big Big Navi and then a Big Big Big Navi with 320CU, 5ghz boost and 4.5ghz game clock to destroy Nvidia and bring in the age of complete AMDominance. Look into it.

 

[Graphenewhen]

A source close to me (u can't really see it) has said 3 variants of Navi 21. 80CU Big Navi, 160CU Big Big Navi and then a Big Big Big Navi with 320CU, 5ghz boost and 4.5ghz game clock to destroy Nvidia and bring in the age of complete AMDominance. Look into it.

Raja, is that you?

 

[soresu]

It's all forum psychology, he is a scapegoat.

It's not all forum psychology.

He was a bit of a chatterbox at PR events and floated GPU chiplets in passing conversation before we even saw chiplets properly in Zen 2 platforms.

If nothing else he put ideas in peoples heads long, long before AMD was ready to make good on them - that does tend to put a bit of undue expectation on AMD to produce something equally astounding as Zen 2 earlier than the RTG budget allowed for.

 

[zinfamous]

80, 72 and 64 CUs. As was speculated by many for a long time already.

yeah, but the full die 80 CU unit being AMD exclusive, limited release?

Not gonna make people happy, if that ends up being true. So, I don't believe it, at least forever. I do think it's an industry-wide force on supply that is making these decisions necessary. So, I can definitely see full 80 CU dies being mostly unobtainable throughout the next year, until a refreshed mid-arch stack pops up and they release an "anniversary" version or something.

 

[Head1985]

Oh god why did I even bother reading nonsense from Coreteks. The guy is saying a 64CU RX6800 with 1820MHz game clock (funny how it's the same as XSX) pulls 290W. Meanwhile a 52CU XSX peaks at 145W lol.

Not to mention 2.5Ghz AIB Navi 21.If stock navi 21 is 300+ how many W will pull 2.5Ghz oc?500w?Its BS.
No way reference navi will be that power hungry.I still think all reference navi 21 will be bellowe 300w.

 

[Tup3x]

https://www.igorslab.de/en/the-can-...-works-oc-may-maybe-a-glance-in-bios-details/


I couldnt fully grasp everything through the machine translation but it looks like big navi could be quite an overclocker.

To me it looks like it's already at the limit, like RTX 3080 and ~300W card (probably over).

 

[Glo.]

Yeah, pretty much AMD is going all out, with balls to the wall, unlocking all of what those GPUs are capable of.

 

[Kuiva maa]

The rumors generally make sense the way I see, besides Navi 21 being one big GPU with 80CU and 22/23/24 being 40CU or smaller.

 

[leoneazzurro]

To me, the strange feeling in the RDNA2 lineup is the uncommon rumored gap between the top of the line (80CU) and the next smaller die (40 CU) and the very small gap between Navi 22, 23 and 23 to 24. Granted that we will see cut down Navi21 dies, but seeing that a certain performance target could have been covered with a slightly larger die (60-64CU) compared to the rumored 40CU of Navi22, while 40 is also quite near the 32 CU figure for NAVI23 (and both will be also cut down). One possibility is that AMD is trying to target these GPUs not only for the desktop marked but especially for the mobile market, where a smaller bus for the memory and possibly better clock/power/voltage curve compared to the competition could be a winning factor for OEMs.

 

[Timorous]

I think 505mm^2 N21 is nonsense.

That would give you 83 good dies per wafer. Even if they were all 6900XT and the 6900XT is 3080+5% it means AMD can look to charge around $700 per card. That is revenue of $58,100.

OTOH a wafer of Zen3 dies is 782 good CCDs. Even if all are used in 5600X's that would be a revenue of $230k.

A 500+ mm^2 GPU is far far far too costly for AMD when they can generate 4x the revenue by making Zen3 parts instead.

Edit To Add: The other option is that availability is like the 3080/3090, non existent.

 

[Panino Manino]

It's all forum psychology, he is a scapegoat. AMD didn't deliver a card that could beat Nvidia while he was there and he has now left. Putting all the blame on him means that people can hope RDNA2 can now beat Nvidia because he's not in charge. Where as in-fact he was just a cog in a very big wheel and him being there, or not being there will have in the end made a very small difference to how RDNA2 turns out.

And yet, the same team just without that special "clog" is now performing and behaving so differently, it's like, just like, his departure made a difference.

 

[Qwertilot]

And yet, the same team just without that special "clog" is now performing and behaving so differently, it's like, just like, his departure made a difference.

Or just maybe it has something to do with them moving from massively cash starved to actually having the resources to do a decent job? That's much more a priori probable than a single management person changing.

 

[marcUK2]

I think 505mm^2 N21 is nonsense.

That would give you 83 good dies per wafer. Even if they were all 6900XT and the 6900XT is 3080+5% it means AMD can look to charge around $700 per card. That is revenue of $58,100.

OTOH a wafer of Zen3 dies is 782 good CCDs. Even if all are used in 5600X's that would be a revenue of $230k.

A 500+ mm^2 GPU is far far far too costly for AMD when they can generate 4x the revenue by making Zen3 parts instead.

Edit To Add: The other option is that availability is like the 3080/3090, non existent.

500mm ZEN 4 FTW!

But seriously, why aren't CPUs a little bigger and far more performant if the *individual* core size was doubled, so that IPC was doubled (in theory) - Is the know-how just not there to scale IPC with individual core size?