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]

It makes the most sense in my mind to have each Shader Engine be its own chiplet when AMD moves RDNA to a chiplet architecture. The Shader Engine has all of the complete parts to run the graphics pipeline so its the most basic building block that gets duplicated. The memory controllers, last-level cache (Infinity Cache), display connectors, codec engines, etc can be on an IO die. This approach pretty much mirrors how it's done on Zen 2/3.

Agree that a 40 CU chiplet would probably be too large but disagree with cache being on IO die. In fact the cache would be needed on the chiplet to reduce data transfers from/to IO die for lower latency and power use.

 

[sandorski]

Hitler responds!

Best Hitler Responds video 2020!

 

[PhoBoChai]

I can't help but think that the biggest loser in this renjuvated competition between AMD and Nvidia is ... Intel. I wonder how Raja feels now, or how much of this he expected.

Speak of the man..

https://twitter.com/x/status/1321678187370196993

 

[PhoBoChai]

Agree that a 40 CU chiplet would probably be too large but disagree with cache being on IO die. In fact the cache would be needed on the chiplet to reduce data transfers from/to IO die for lower latency and power use.

Cache has to be on chiplet, to minimize traffic across the IF and reduce impact of bandwidth stress (POWER) and latency (performance).

Think 1 Shader Engine, 2 Arrays, with its own chunk of 32-64MB of L3 cache.

On 5N, the density for SRAM is insane and that cache is tiny.

 

[insertcarehere]

They will be once AIB cards drop. 300W power limit is the only reason clocks are this low. Basically marketing reasons to brag it’s lower than Ampere. This things will boost to 2.5GHz EZ PZ.

People in this very thread were pointing to earlier leaks + XSX/PS5 to argue that AMD can do 80CU/2.2ghz @ 250W TBP or 80CU/2.4+ghz @ 300W TBP...

Hm, up to 250W for 80CU at 2.2ghz is kind of nuts but that is exactly likely what we can expect.

https://twitter.com/x/status/1309895830044905472

 

[moonbogg]

There isn't a card spoken of this year that makes any sense compared to the 6800 XT. How does a person not get that card? You get 3080 performance with 16GB for $650. Game over. I swear to god Nvidia can take their ray tracing advantage and stuff it into a hole somewhere. No one cares about dumping half their frames for less than half-better reflections and lighting. Oh, but Tay Rraced Minecraft though! I'll alert my 6 year old right away!

 

[Saylick]

Agree that a 40 CU chiplet would probably be too large but disagree with cache being on IO die. In fact the cache would be needed on the chiplet to reduce data transfers from/to IO die for lower latency and power use.

Yeah, I was thinking about it some more about how the approach should follow Zen and you're right; the LLC needs to be local to the execution units in the same way how the fat L3 is on the CCD with Zen 2/3.

 

[dzoni2k2]

People in this very thread were pointing to earlier leaks + XSX/PS5 to argue that AMD can do 80CU/2.2ghz @ 250W TBP or 80CU/2.4+ghz @ 300W TBP...

I expect OCed AIB cards to pull around 350W. For additional 10% perf.

 

[KompuKare]

Wonder if Nvidia and AMD bid to try and get a Cyberpunk bundle for the holidays, and if so, who won?

Well, if AMD bid and won that would explain the last minute delays as they rush to implement and optimise RT for RDNA2 which is likely to be somewhat different than Nvidia's RTX implementation.
The reverse is also possible and more probable: Nvidia won and are currently helping CDPR to make sure ray tracing is implemented in a way which works well on Ampere and is slow on RDNA2 and Turing.

 

[IntelUser2000]

It looks like eDRAM cache. Intel and IBM used/are using it. Transistor count is "only" 28 billion so any transistor cache is smaller.
EDRAM on Wiki

Not sure serious or not, but its not using eDRAM.

6 transistor per bit x 8 bits per byte x 128MB = 6.44 billion transistors. RX 6800 series has 16 billion more transistors than RX 5700 and 13 billion more than Radeon VII.

If it was eDRAM they'd be able to put 500MB in the same space and transistors.

It seems really good if they have achieved it. There still might be scnearios where it hits out of the cache, but its a pretty large one.

 

[zinfamous]

Hitler responds!

OK, the final comment is worth it.

 

[uzzi38]

I expect OCed AIB cards to pull around 350W. For additional 10% perf.

I'm expecting more like 5% extra performance, but I especially don't want to comment on power draw. From increase in power limits alone it should be ~40W, but then different AIBs will have different components on board so power draw estimates get a little wonky.

 

[Veradun]

Was 300w expected for 6900xt and 6800xt? Seems higher than my recollection of rumors

Pending reviews I will most likely end up getting a 6900xt. $999 is firmly in the range of acceptably overpriced for me

Latest rumors put the card between 3070 and 3080 at 320W

 

[Timorous]

People in this very thread were pointing to earlier leaks + XSX/PS5 to argue that AMD can do 80CU/2.2ghz @ 250W TBP or 80CU/2.4+ghz @ 300W TBP...

We have not seen in game clocks yet.

5700XT has a boost clock of 1905 and in the TPU the median clockspeed was 1891. IF that applies to the 6800XT/6900XT then in game clocks should be around 2.23 ghz on average just like PS5.

There was talk of 2.4Ghz @ 300W TBP being possible and perhaps it is but the binning required to do that is not worth it just like it was not worth AMD binning the 5950X with a 5Ghz Boost clock.

 

[dzoni2k2]

I'm expecting more like 5% extra performance, but I especially don't want to comment on power draw. From increase in power limits alone it should be ~40W, but then different AIBs will have different components on board so power draw estimates get a little wonky.

You could be right, I guess it depends on how power limited 6800XT is at 300W. How long can it sustain that boost clock. More limited it is the bigger the perf jump AIB GPUs will have.

 

[leoneazzurro]

I think we could see even higher than 350W models from AIB (similarly to what happened to 3080/3090), scaling with frequency must be seem but it could be good due to the cache speed increasing.

 

[PhoBoChai]

There isn't a card spoken of this year that makes any sense compared to the 6800 XT. How does a person not get that card? You get 3080 performance with 16GB for $650. Game over. I swear to god Nvidia can take their ray tracing advantage and stuff it into a hole somewhere. No one cares about dumping half their frames for less than half-better reflections and lighting. Oh, but Tay Rraced Minecraft though! I'll alert my 6 year old right away!

This talk on better RT is also way too premature, as all the freaken synthetics are RTX optimized, and DXR 1.0 based. There isn't a single RTX game that uses DXR 1.1, taking advantage of RDNA2's hw RT capabilities or even optimized for RDNA2 on the market, yet.

The earliest we get it, is likely the console ports of Dirt 5, and some of the titles AMD mentioned.

And agreed, 6800 is nice, but a little bit more $, u get 6800XT with much more perf. 6900XT price is nuts for another 10%. No thanks. 6800XT is where it's at.

 

[uzzi38]

This talk on better RT is also way too premature, as all the freaken synthetics are RTX optimized, and DXR 1.0 based. There isn't a single RTX game that uses DXR 1.1, taking advantage of RDNA2's hw RT capabilities or even optimized for RDNA2 on the market, yet.

The earliest we get it, is likely the console ports of Dirt 5, and some of the titles AMD mentioned.

And agreed, 6800 is nice, but a little bit more $, u get 6800XT with much more perf. 6900XT price is nuts for another 10%. No thanks. 6800XT is where it's at.

There is actually one.

Minecraft RTX received a DXR1.1 patch recently.

EDIT: Ok, it was like nearly 2 months ago, but still.

New Minecraft RTX Beta brings DXR 1.1, updates DLSS, improves performance and more

Mojang and NVIDIA have released a new beta patch of Minecraft RTX that introduces DXR 1.1, updates the DLSS upscaling, and improves End Dimension in ray tracing.

www.dsogaming.com

 

[Kuiva maa]

Not surprised with the lack of RT benchmarks yeah, literally all games that have it are nvidia sponsored. AMD needs to work on their studio RT relations.

 

[zinfamous]

Not surprised with the lack of RT benchmarks yeah, literally all games that have it are nvidia sponsored. AMD needs to work on their studio RT relations.

They've got it as a feature sponsored by Sony and MSFT for their consoles, and whatever games are designed there. I assume the optimization will be different, and certainly some will likely still favor nVidia discreet cards on PC.

However, I wonder what the custom parts for the consoles will do in terms of RT-specific performance compared to the Zen3 + RDNA2 + 550 custom systems on PC.

 

[moinmoin]

However, I wonder what the custom parts for the consoles will do in terms of RT-specific performance compared to the Zen3 + RDNA2 + 550 custom systems on PC.

DXR1.1 should be the same on Xbox Series and under Windows. The wildcard are optimizations on PS5 which obviously won't and can't use DirectX.

 

[zinfamous]

DXR1.1 should be the same on Xbox Series and under Windows. The wildcard are optimizations on PS5 which obviously won't and can't use DirectX.

I mean hardware optimizations. The consoles use custom Zen 2, so whatever customization is going on in their boards is going to somehow emulate this Zen 3 + RDNA2 + 550 optimizations in a different way. I guess, if regular Zen2 can't work with DXR1.1 (I have no idea if that is the case), then that is already something...

AMD presented (SAM and RAGE) as a specific AMD ecosystem feature that requires Zen3. I'm not sure if that has any relation to RT performance or not, but I'm really curious about those console Zen 2 chips with RDNA2, and how it compares to Zen2 or Zen3 with RDNA2 on PC.

Even though it isn't a lot, we know quite a bit more about RDNA2 performance on PC than we do on consoles...just some leaked speculation and a few FPS charts. These things claim 4k/120. ...lol, seriously? I was always dubious about 4k/60 really being possible on these consoles and I still am...so I think it's a DLSS-like feature that they are utilizing to upscale 2k content. For me, if it looks great, I don't really care, but I don't really trust the CU counts in those RDNA chips, at that power draw, actually pushing true 4k/120.

 

[Krteq]

CPU architecture has nothing to do with DXR 1.1

 

[CastleBravo]

They've got it as a feature sponsored by Sony and MSFT for their consoles, and whatever games are designed there. I assume the optimization will be different, and certainly some will likely still favor nVidia discreet cards on PC.

However, I wonder what the custom parts for the consoles will do in terms of RT-specific performance compared to the Zen3 + RDNA2 + 550 custom systems on PC.

Microsoft? Skyrim DXR confirmed!

 

[Shivansps]

As I have said in this thread: 40 CU GPU should be between RTX 2080 Super(10% above) and RTX 2080 Ti, 60 CU should be 10-20% above RTX 2080 Ti, 80 CU die should be 40-50% above RTX 2080 Ti.

So far the two last one were pretty spot on, so I don't think my info was wrong on the 40 CU die .

You are assuming the 6700XT will be the 40CU part, what if its not?