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

 

[Karnak]

So what's with that large gap between N21 (80CU) and N22 (40CU) if N23 is a 32CU chip? Any ideas?

 

[DiogoDX]

https://www.reddit.com/r/Amd/comments/j06xcd

Navi 23: 32 CUs/128 bit bus.

Bloody hell.

Looks like AMD will choose to cut down the N21 instead of making another die with 60CU. 64, 72 and 80 configurations would be my guess.

 

[Glo.]

So what's with that large gap between N21 (80CU) and N22 (40CU) if N23 is a 32CU chip? Any ideas?

There is a possiblity that Navi 24 might be 64 CU GPU, or that AMD will reuse Navi 21 die.

Guys. There is mind blowing info there.

Navi 21 has 240W TDP at 2200 MHz.
Navi 22 has 170W at ... 2500 MHz.

O_O

What the hell?

 

[itsmydamnation]

Looks like AMD will choose to cut down the N21 instead of making another die with 60CU. 64, 72 and 80 configurations would be my guess.

Given 40cu at 2.5 GHz that will probably work OK in terms of perf delta .

 

[Glo.]

2500 MHz is 33% above RX 5700 XT clock speeds. And that excludes all IPC gains of RDNA2 architecture.

My mind is blown. Even I did not expected stuff like this.

Edit. With IPC increases, this, 40 CU GPU will rival RTX 2080 Ti performance levels. At 170W of power drawn.

 

[uzzi38]

You guy's feeling that big Wang energy at work?

Look at dem clocks.

 

[Hitman928]

There is a possiblity that Navi 24 might be 64 CU GPU, or that AMD will reuse Navi 21 die.

Guys. There is mind blowing info there.

Navi 21 has 240W TDP at 2200 MHz.
Navi 22 has 170W at ... 2500 MHz.

O_O

What the hell?

I would expect that to just be a frequency bin cap, doesn't necessarily mean it will actually boost that high, especially under any kind of real load. AMD seems quietly confident with RDNA2 to me, we'll see if that's true or not soon enough.

 

[itsmydamnation]

I would expect that to just be a frequency bin cap, doesn't necessarily mean it will actually boost that high, especially under any kind of real load. AMD seems quietly confident with RDNA2 to me, we'll see if that's true or not soon enough.

If they are the power play tables then they are real world clocks, just look at navi 10 for reference.

 

[Glo.]

If they are the power play tables then they are real world clocks, just look at navi 10 for reference.

Yep. Its worth adding those are GPUs for Apple, and those usually have lower power and lower clock speeds, compared to desktop GPUs.

I think the PP values are upper limit(hard power gate) that those GPU should not exceed at any conditions. But here I might be wrong. It was long time ago that I was interested and rooted in Apple software.

 

[Konan]

There is a possiblity that Navi 24 might be 64 CU GPU, or that AMD will reuse Navi 21 die.

Guys. There is mind blowing info there.

Navi 21 has 240W TDP at 2200 MHz.
Navi 22 has 170W at ... 2500 MHz.

O_O

What the hell?

Add memory (say 50w) and power delivery (40-50W?) watts to the 238W of N21
Then you have something going for the 3080

 

[TESKATLIPOKA]

There is a possiblity that Navi 24 might be 64 CU GPU, or that AMD will reuse Navi 21 die.

Guys. There is mind blowing info there.

Navi 21 has 240W TDP at 2200 MHz.
Navi 22 has 170W at ... 2500 MHz.

O_O

What the hell?

Is that the clockspeed max turbo? Is socket_power_limit_ac TBP or just TDP?
Navi 21 will have a lot better perf/W ratio than Navi 22.
I kinda don't understand why between Navi 21 vs Navi 22 is 40CU difference, but between Navi 22 vs Navi 23 is only 8CU? Navi 23 won't be much smaller than Navi22. BTW It looks like Navi 21 will have 384bit GDDR6.
I want to know what is the actual die size for Navi22 and Navi23, because It doesn't look like It's 360mm^2 and 240mm^2 what was leaked before.
BTW I don't understand how is AMD planning to go against Tiger Lake with only 8CU Vega. At least Rebrand has 12CU RDNA2, but It will take some time until It's released.

 

[Glo.]

Add memory (say 50w) and power delivery (40-50W?) watts to the 238W of N21
Then you have something going for the 3080

No. Its the upper limit, for total GPU board power.

Look at the Power tables for Navi 1X GPUs. That 238W is the total board power for N21 if the history of how Apple treats their GPUs is anything to go by.

 

[DisEnchantment]

Yep. Its worth adding those are GPUs for Apple, and those usually have lower power and lower clock speeds, compared to desktop GPUs.

I think the PP values are upper limit(hard power gate) that those GPU should not exceed at any conditions. But here I might be wrong. It was long time ago that I was interested and rooted in Apple software.

PP tables are the same everywhere for the same HW. Concept is the same.
PP tables are soft limits. You can overwrite them, like how Igor is doing with his tool. (MorePowerTool)
Hard limits are in the atombios.

 

[Konan]

No. Its the upper limit, for total GPU board power.

Look at the Power tables for Navi 1X GPUs. That 238W is the total board power for N21 if the history of how Apple treats their GPUs is anything to go by.

I saw the power tables. Might not have anything to do with apple as the OP says.
So it’s an unknown right now it could be potentially another hundred watts added on.

 

[Glo.]

PP tables are the same everywhere for the same HW.
PP tables are soft limits. You can overwrite them, like how Igor is doing with his tool.
Hard limits are in the atombios.

Even in Apple hardware? Those specs are specifically for GPUs that will go to Apple computers. Somewhat like LInux GPU microcode.

 

[DisEnchantment]

Even in Apple hardware? Those specs are specifically for GPUs that will go to Apple computers. Somewhat like LInux GPU microcode.

Those are not specific to Apple, firmware blob is same everywhere for the same HW/SKU.
Apple gets it own SKU, which has its own firmware(and PP table).
Afterwards OEM can change whatever (PP table part) they want as long as it is with the limits defined inside atombios.
The firmware has microcode as well which instructs how the different ip blocks will function. Like CPU microcode.

Firmware is provided by AMD.
AMD provides an interface via driver to overwrite the PP table.

 

[Olikan]

Lmao, there was rumours early this year of rdna2 reaching 2.7ghz.. we thought it was insane...

now AIB cards will likely reach that

 

[Glo.]

I saw the power tables. Might not have anything to do with apple as the OP says.
So it’s an unknown right now it could be potentially another hundred watts added on.

Looking at Navi 10 and Navi 14 power tables, not really that is the case.

It appears that most likely picture is that 238W and 170 is the actual power target for those GPUs.

 

[Konan]

Looking at Navi 10 and Navi 14 power tables, not really that is the case.

It appears that most likely picture is that 238W and 170 is the actual power target for those GPUs.

possible, but as the OP says N21 is really still unknown. Exciting though so far.

 

[Karnak]

Even if it's a 300W GPU. I mean with like 2.1GHz+. Damn, I'll take it all day.

Imagine overclocking these beasts to 2.4GHz with water cooling (or maybe even air cooled, noise aside). Damn.

 

[Glo.]

possible, but as the OP says N21 is really still unknown. Exciting though so far.

The thing is.

In the clock tables we see that N21 has 2050 MHz clock target, while maximum is 2200 MHz. 2050 is the "gameclock" clock, and 2200 MHz is the boost clock?

If so, why N22 does not have "gameclock" and has exactly the same clock table as maximum Boost clock?

Also, Even 238W at 2050 MHz for whole board in 80 CU GPU is pretty darn impressive efficiency achievement.

It basically means that AMD achieved at least 2x the efficiency, on the same node, if those clock and power tables are going to have any relation to the real world performance.

2x performance/watt. On the same node.

O_O

 

[TESKATLIPOKA]

Another weird thing is that N22 has 40CU, 192bit 16GHz GDDR6 and N23 has 32CU but only 128bit 14-16Ghz gddr6. N22 should be 48CU, then It would be 50% difference in CU and bandwidth. That would make much more sense and there wouldn't be so much difference between N21 and N22.

 

[Glo.]

Another weird thing is that N22 has 40CU, 192bit 16GHz GDDR6 and N23 has 32CU but only 128bit 14-16Ghz gddr6. N22 should be 48CU, then It would be 50% difference in CU and bandwidth. That would make much more sense and there wouldn't be so much difference between N21 and N22.

What if:
N21 - 80 CUs
N22 - 40 CUs
N23 - 32 CUs
N24 - 64 CUs

Doesn't this make sense?

 

[Veradun]

So what's with that large gap between N21 (80CU) and N22 (40CU) if N23 is a 32CU chip? Any ideas?

In a vacuum I would have expected Navi 22 to cap at 48 and navi23 at 24.

Picture me confused

 

[stebler]

OP here. These pptables are test configurations, they are not used by production hardware (you can likely boot with some parameters to use them though). And yes they are the same in macOS and in Linux.