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

 

[Saylick]

Same density or, size?

Density. XSX size is closer to 360mm2. For transistor count, XSX is 15.4B.

 

[raghu78]

13B in 306mm² is 42.5M/mm² so practically the same as Series X.

Where did you get 13B transistors for Sony PS5 SoC. Did Sony give any numbers ?

 

[BlitzWulf]

https://twitter.com/x/status/1313944579826159623

https://patents.justia.com/patent/10795825

 

[Timorous]

Where did you get 13B transistors for Sony PS5 SoC. Did Sony give any numbers ?

It is an estimate.
Navi10 is 10.3B
zen2 with 8MB cache is around 2B.
Extra 0.7-1B for CPU-GPU interconnects, ray tracing and other 'stuff'.

Total ballpark is 13B.

 

[Tup3x]

So apparently Navi 21 XTX has more CUs than XT.

https://twitter.com/x/status/1314099938255605761

 

[exquisitechar]

So apparently Navi 21 XTX has more CUs than XT.

https://twitter.com/x/status/1314099938255605761

I'm pretty sure XTX is just a bit faster than XT, it doesn't explain their claim that AMD is only targeting the 3070 at all. This is probably only the beginning of VCZ backpedaling from their tweets yesterday...

 

[uzzi38]

If you checked his follow up Tweet he also stated that the XT was over 3070 perf. Not matching a 3070 or targeting a 3070 in terms of where performance should actually lie, just the 3070 is likely what would be have used as a comparison point.

Now you have an XTX version on top with more CUs and probably higher clocks (usually how AMD segments these dies). You can probably guess where performance will be closer to landing at this point.

 

[exquisitechar]

Now you have an XTX version on top with more CUs and probably higher clocks (usually how AMD segments these dies). You can probably guess where performance will be closer to landing at this point.

Barely matching a 3080 at best? Assuming "faster than a 3070" is like 15% faster, considering VCZ's words, and that XTX is another 15% faster like the 5700XT compared to the 5700.

In the first place, XTX having more CUs would be surprising, honestly.

 

[Timorous]

Barely matching a 3080 at best? Assuming "faster than a 3070" is like 15% faster, considering VCZ's words, and that XTX is another 15% faster like the 5700XT compared to the 5700.

In the first place, XTX having more CUs would be surprising, honestly.

If AMD have a 500mm^2 die (not something I am convinced of but anyway) with cache and a 256 bit bus it may be entirely possible they have decided to have 3 CU bins for it to make up for the naturally lower yield a larger die has. I could imagine 80CU, 72CU and 64CU parts all being built off of a single die.

This would also work with N23 at 40CUs, N22 at 56CUs (+40%) and N21 at 80CUs (+43%). That would seem like a stack AMD could use to fill the bulk of the market.

 

[moinmoin]

https://twitter.com/x/status/1313944579826159623

https://patents.justia.com/patent/10795825

Very interesting patent. Essentially it boils down to (to use the example used in the patent itself): Compressing data in cache is a good idea since cache space is limited. Since L2$ has low latency of few cycles, the additional latency of compressing the data makes it unfeasible to use there. But since L3$'s latency is significantly higher and L2$ needs to talk to it anyway, the latency from compressing the data is masked there. Thus compression of data in L2$ is achieved by letting L3$ compress the data while L2$ waits for the data to arrive during the usual latency after it requested it. (It also discusses having several levels of compression that are picked as best fit at that point.)

 

[uzzi38]

@stebler posted an update to the firmware stuff he found before:

https://www.reddit.com/r/Amd/comments/j7bpzs

 

[Olikan]

https://twitter.com/x/status/1313944579826159623

https://patents.justia.com/patent/10795825

Or i missed something, or It's used already

 

[Zstream]

Very interesting patent. Essentially it boils down to (to use the example used in the patent itself): Compressing data in cache is a good idea since cache space is limited. Since L2$ has low latency of few cycles, the additional latency of compressing the data makes it unfeasible to use there. But since L3$'s latency is significantly higher and L2$ needs to talk to it anyway, the latency from compressing the data is masked there. Thus compression of data in L2$ is achieved by letting L3$ compress the data while L2$ waits for the data to arrive during the usual latency after it requested it. (It also discusses having several levels of compression that are picked as best fit at that point.)

I still think this will enable some sort of multi-GPU down the line. Having the L3 shared, or through some sort of crossbar will fix many issues.

 

[eek2121]

The backtrack from VideoCardz was rather curious to me since it most definitely appears to be false. The numbers don't add up, and then there was that OpenVR leak...

For the record:

Navi 21: 2050-2200 MHz clocks
Navi 22: 2300-2500 MHz clocks

Navi 21: 80CU
Navi 22: 40 CU

If Navi had scaling issues, they would not be releasing Navi 21. Navi 22 is more than capable of competing with the 3070 and 3070ti. I suspect that one of the leakers got caught and is now attempting to backpedal...

EDIT: The numbers I posted came directly from BIOS and source code dumps provided by AMD.

 

[Hitman928]

Zen3 and Big Navi preview at Zen3 event.

 

[dr1337]

Zen3 and Big Navi preview at Zen3 event.

View attachment 31344

GeForce RTX 3090 Founder review (Page 16)

It is time to review and benchmark the new premium flagship graphics card, all hail the might GeForce RTX 3090. Armed with a shader core count that will make your eyebrows frown and a nauseating 24 GB...

www.guru3d.com

3fps slower in gears 5 according to guru3d

 

[CastleBravo]

Zen3 and Big Navi preview at Zen3 event.

View attachment 31344

TechPowerUp's RTX 3080 FE review numbers for comparison:

2080ti - 62.8fps in Gears5, and 49fps in BL3
3080 - 84.4fps in Gears5, and 70.3fps in BL3

That would put this unannounced 6000 series card at about half way between the 2080ti and 3080.

 

[Sable]

GeForce RTX 3090 Founder review (Page 16)

It is time to review and benchmark the new premium flagship graphics card, all hail the might GeForce RTX 3090. Armed with a shader core count that will make your eyebrows frown and a nauseating 24 GB...

www.guru3d.com

3fps slower in gears 5 according to guru3d

10fps according to

NVIDIA GeForce RTX 3080 Founders Edition Review - Must-Have for 4K Gamers

NVIDIA's new GeForce RTX 3080 "Ampere" Founders Edition is a truly impressive graphics card. It not only looks fantastic, performance is also better than even the RTX 2080 Ti. In our RTX 3080 Founders Edition review, we're also taking a close look at the new cooler, which runs quietly without...

www.techpowerup.com

 

[eek2121]

GeForce RTX 3090 Founder review (Page 16)

It is time to review and benchmark the new premium flagship graphics card, all hail the might GeForce RTX 3090. Armed with a shader core count that will make your eyebrows frown and a nauseating 24 GB...

www.guru3d.com

3fps slower in gears 5 according to guru3d

Depends on the benchmark.

Also note that AMD is using a new, unreleased CPU for the comparison.

That being said, those who claimed Big Navi was only competing with the 3070...RIP.

 

[Hitman928]

Yeah, you shouldn't try and directly compare to reviewer scores from this. There's all kinds of variables at play. This does show though that Big Navi is at least close to 3080 at worst in 3 these games.

 

[uzzi38]

Numbers change quite a bit between reviews. For the time being, wait and see.

 

[GoodRevrnd]

So if what they showed was the best seems like that was a crush 3070 target, or target 3080 and near miss.

 

[Qwertilot]

Well you always take a little bit off because, well, marketing slide!

Still should do fine overall.

 

[beginner99]

GeForce RTX 3090 Founder review (Page 16)

It is time to review and benchmark the new premium flagship graphics card, all hail the might GeForce RTX 3090. Armed with a shader core count that will make your eyebrows frown and a nauseating 24 GB...

www.guru3d.com

3fps slower in gears 5 according to guru3d

than the 3080. Even more in Borderlands 3. Hm...not good if top end model.

 

[Heartbreaker]

If they are very close to 3080, have 16GB of VRAM and undercut on pricing, it will still be a killer card that flies off the shelves.