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

 

[eek2121]

How 40 CUs , same amount as RX5700XT will reach 2080 Super performance that is 20% faster today ??
52 CUs will not get close to 2080Ti/3070, perhaps close to 2080 Super/3060
72 CUs will not get 40-50% faster vs 2080Ti

Hell the entire stack is way off

7% higher IPC and much higher clock speeds. Call me out if you want, but set a reminder for launch day.

EDIT: The 6700XT will have clocks north of 2 ghz, for example.

 

[leoneazzurro]

Hmm... I made some calculations based on the Xbox series S, Xbox X (Scorpio) and the 5500XT. According to Techpowerup, at 1080p both Scorpio and 5500XT GPUs perform like the RX580. There are some difference in the bandwidth, but quite limited. Microsoft declared that Xbox Series S is faster than Scorpio, but let's assume their GPU is performing the same and the difference is due to the better CPU cores in the Series S. This also means that 20 RDNA2 CU @ 1565 MHz perform the same as 22 RDNA1 CU @1717 MHz.
This means that IPC increase from RDNA1 to RDNA2 is around 20%. Let's consider that the Xbox Series X has a 25% more bandwidth (but not quite fully useable due to being divided in two asymmetrical segments). Let's say that this amounts for another 5% of IPC gain. Thsi means that, CU vs CU, RDNA2 should have around 15% better IPC respect to RDNA1.

 

[eek2121]

Nope.

Prove it. I will go even further than this and possibly even estimate clock speeds and FP32 TFLOPs. (22 on the biggest chip)

EDIT: The last time I estimated this crap was with Ampere, it was many months ago. I said Q4 and 30% faster than a 2080ti. I don’t remember if it was here or elsewhere (or both), but I was ridiculed and made fun of. Those people are eating their hats now. Ampere appeared...exactly as expected. Q4 and 30% faster. The only unexpected surprise was the higher TDP.

 

[Glo.]

Prove it. I will go even further than this and possibly even estimate clock speeds and FP32 TFLOPs. (22 on the biggest chip)

Its simple.

You have no idea which Die will be used for which SKU. AMD will not harvest 40 CU dies, which will have performance of RTX 3070 to go to 6500 SKU range. They will have apropriate price point, if they will perform accordingly.

 

[Glo.]

This means that IPC increase from RDNA1 to RDNA2 is around 20%. Let's consider that the Xbox Series X has a 25% more bandwidth (but not quite fully useable due to being divided in two asymmetrical segments). Let's say that this amounts for another 5% of IPC gain. Thsi means that, CU vs CU, RDNA2 should have around 15% better IPC respect to RDNA1.

Considering this is speculation thread, I'll bite, even tho Im a complete layman in terms of highly technical stuff.

First of all, RDNA2 appears to schedule more work with each cycle, based on Microsoft's own papers on RDNA2. Secondly, we have all of that Cache rumors that will happen with RDNA2, and most importantly we have leaked recently patents that are discussing new L1 cache design and CU clustering designs.

Both: more work scheduled each cycle, and L1 cache redesign would bring massive IPC increases compared to previous generation.

Important note. In this discussion I use word: "Would". We have no idea if RDNA2 will have that new Cache design, even tho, rumors point into this possibility.

 

[Glo.]

AMD Dimgrey Cavefish and VanGogh Support Lands In Radeon Linux OpenGL Driver - Phoronix

www.phoronix.com

Dimgrey Cavefish is the new GPU codename found in Mesa Drivers.

Bloody hell.

 

[leoneazzurro]

Mine was a speculation based on the projected performance of the Xbox Series S - of course it could be entirely wrong but IIRC rumors of a 10-13% better IPC for RDNA2 were floating around recently.

 

[Glo.]

Mine was a speculation based on the projected performance of the Xbox Series S - of course it could be entirely wrong but IIRC rumors of a 10-13% better IPC for RDNA2 were floating around recently.

"I'll bite" as in "I'm gonna chime in" with my uneducated speculation .

P.S. I believe that you MIGHT be correct, IF the rumors of new cache, and patent materialize in the new architecture, and it really might have 20% IPC uplift over previous generation, after all.

 

[eek2121]

Its simple.

You have no idea which Die will be used for which SKU. AMD will not harvest 40 CU dies, which will have performance of RTX 3070 to go to 6500 SKU range. They will have apropriate price point, if they will perform accordingly.

AMD already does this for the 5600 XT and 5700 XT.

EDIT: Do you own a hat?

 

[moinmoin]

AMD Dimgrey Cavefish and VanGogh Support Lands In Radeon Linux OpenGL Driver - Phoronix

www.phoronix.com

Dimgrey Cavefish is the new GPU codename found in Mesa Drivers.

Bloody hell.

AMD: Countering rampant speculation with fishy code names. 👍

 

[DisEnchantment]

VGH use RDNA2.

+++ b/src/amd/addrlib/src/core/addrlib.cpp
@@ -226,6 +226,9 @@ ADDR_E_RETURNCODE Lib::Create(
                     case FAMILY_NV:
                         pLib = Gfx10HwlInit(&client);
                         break;
+                    case FAMILY_VGH:
+                        pLib = Gfx10HwlInit(&client);

@@ -177,6 +177,7 @@ const char *ac_get_llvm_processor_name(enum radeon_family family)
    case CHIP_SIENNA_CICHLID:
    case CHIP_NAVY_FLOUNDER:
    case CHIP_DIMGREY_CAVEFISH:
+   case CHIP_VANGOGH:
       return "gfx1030";

 

[Glo.]

AMD already does this for the 5600 XT and 5700 XT.

5700 XT was priced apropriately - on the same level as slower RTX 2060 Super, and 100$ less that roughly the same RTX 2070. 5600 XT was priced at 280$, 20$ less than RTX 2060, from which RX 5600 XT was faster.

So no, AMD won't price 40 CU die, with performance of RTX 3070 at the price of RTX 3060, or less.

AMD has smaller RDNA2 die, in the pipeline that will fit in 199$ price tag, and 6500 SKUs.

 

[AtenRa]

Higher IPC and higher clock speeds? .

7% higher IPC and much higher clock speeds. Call me out if you want, but set a reminder for launch day.

EDIT: The 6700XT will have clocks north of 2 ghz, for example.

20% higher performance per CU ??? not a chance in hell

 

[eek2121]

The 5600XT has the exact same die size as the 5700 XT, yet it was priced at $279, $120 cheaper than the 5700 XT. It also has a *gasp* 192 bit memory bus. Additional CUs would be utilized due to stuff like RT. You seem to forget that GPU performance has always historically stepped down a tier across generations. With RDNA2, AMD has a substantial perf/watt increase (they said 50%, but evidence suggests 40-45%), allowing the midrange cards to really shine.

The only outlier in the lineup is, IMO, the 6500XT. AMD may compensate for this with a small price increase, small bus size, and less VRAM.

Prices on GDDR6 have dropped since the 5000 series launch.

EDIT: In case I wasn’t clear, AMD needs a way to bin the 40CU die, and I think the 6500XT would be it.

If I am following what you are writing, you appear to think the 6700XT will have 40CUs, but at much faster clocks. Given the bus size, I don’t believe this will be the case. By limiting the bus size to 192 bit, AMD effectively makes the 6600XT a 5700XT class card (it will have a higher clockspeed, but will be bandwidth limited). The 4 extra CUs will help with RT.

EDIT: It is possible that the 6600 non-XT could be the 36CU part, but that would be odd.

 

[Heartbreaker]

case CHIP_DIMGREY_CAVEFISH:

Gotta love AMD code names this time.

 

[Glo.]

If I am following what you are writing, you appear to think the 6700XT will have 40CUs, but at much faster clocks. Given the bus size, I don’t believe this will be the case. By limiting the bus size to 192 bit, AMD effectively makes the 6600XT a 5700XT class card (it will have a higher clockspeed, but will be bandwidth limited). The 4 extra CUs will help with RT.

You are creating unneeded hype over the prices. If 40CU/192 bit GDDR6 die has RTX 3070 performance it will be priced accordingly, mainly - with undercutting Nvidia offering. Its really that simple. It will also be named as RX 6700 XT.

Its really that simple.

Product segmentation has little to do with how wide the memory bus is, but with how much performance this GPU have. If 40 CU die has 192 bit GDDR6 bus, but the performance of RTX 3070, because of the rumored architectural advancements - AMD can charge premium, and they do not need to slot it as RX 6500 XT.

Secondly, as Rogame on Twitter found out, Navi 22, which is that 40CU/192 bit die has 12 GB VRAM buffer. It won't be used as 6500XT.

For RX 6500 XT SKU there are really only two possibilities.

Either its based on Smallest Navi 2 die, with 128 Bit Bus, or it will be based on reused/refreshed Navi 10 die.

 

[Glo.]

20% higher performance per CU ??? not a chance in hell

Just like with RDNA1 it was impossible, hmmm?

 

[leoneazzurro]

https://twitter.com/x/status/1308480364353130498

Dimgrey Cavefish seems to be Navi 23

 

[GoodRevrnd]

If AMD exceed nvidia performance the prices will NOT be lower. They will be the same unless the performance is >5% over competing part. eek's $600 3080 competitor is super wishful thinking. It doesn't matter how good these parts are, AMD doesn't have nvidia's level of market mindshare. They'll need superior performance WITH competitive RTX/DLSS feature sets to price at the same perf/$, and they might still need a halo win to even do that.

EDIT: Although much lower TDP would shift things in their favor. Inferior features and a couple % lower performance than competing sku they could probably snag enough people preferring the lower TDP to price the same. That comes down to a volume vs margin analysis.

 

[Krteq]

...

Dimgrey Cavefish seems to be Navi 23

We already know that from ranges defined in AMDGPU for gfx1030 GPUs. These MESA drivers uses these ranges, so we have this already confirmed:

NAVI 21 - Sienna Cichlid
NAVI 22 - Navy Flounder
NAVI 23 - Dimgrey Cavefish

 

[AtenRa]

Just like with RDNA1 it was impossible, hmmm?

RDNA 2 vs RDNA 1 is not the same as RDNA 1 vs VEGA

 

[Glo.]

RDNA 2 vs RDNA 1 is not the same as RDNA 1 vs VEGA

What makes you think so?

 

[eek2121]

You are creating unneeded hype over the prices. If 40CU/192 bit GDDR6 die has RTX 3070 performance it will be priced accordingly, mainly - with undercutting Nvidia offering. Its really that simple. It will also be named as RX 6700 XT.

Its really that simple.

Product segmentation has little to do with how wide the memory bus is, but with how much performance this GPU have. If 40 CU die has 192 bit GDDR6 bus, but the performance of RTX 3070, because of the rumored architectural advancements - AMD can charge premium, and they do not need to slot it as RX 6500 XT.

Secondly, as Rogame on Twitter found out, Navi 22, which is that 40CU/192 bit die has 12 GB VRAM buffer. It won't be used as 6500XT.

For RX 6500 XT SKU there are really only two possibilities.

Either its based on Smallest Navi 2 die, with 128 Bit Bus, or it will be based on reused/refreshed Navi 10 die.

Every single piece of evidence I've seen so far indicates that what I says is mostly accurate. Even a new article VideoCardz just published hints that I'm on the right path, though he thinks that the large die will be for the RX 6900 XT only, and the 6800XT and 6800XT will share the medium die, and the 6600 XT/6500 XT will share the small die.

For Reference, VideoCardz says:

NAVI 21 - Sienna Cichlid - RX 6900 XT
NAVI 22 - Navy Flounder - RX 6800 XT / RX 6700 XT
NAVI 23 - Dimgrey Cavefish - RX 6600 XT / RX 6500 XT

Note that The "small die" here has been rumored numerous times to be 251mm2, the "large die" at 505mm2, and the "medium die" around 331mm2. I do have my doubts about those sizes being accurate. I expect they will be slightly larger/smaller due to some of the rework that has gone into the chip. However, the RX 5700 XT was 251mm2, and we know the leaked card with similar specs isn't the 6700 XT because the memory bus would actually make it perform slower than the 5700 XT. That means it's the 6600 XT. If the 6600XT is a 40 CU card, then something has to be the 36 CU card or those dies with defects go to waste. You can dispute that all day long, but at the end of the day I suspect you will be wrong.

As far as pricing is concerned, AMD will likely price the cards to be a bit cheaper than NVIDIA, without sacrificing margins. They won't raise the price of the 6700 XT. They have no reason to. They lack brand recognition in the market place (look at how many people bash AMD, including yourself). That's why that my pricing estimate will be pretty close (note the question marks, +- $50). Not cheap, but not NVIDIA expensive either. The 3080 is priced at $699, so I guarantee you AMD will come in a bit cheaper, and $599 for the equivalent is what is floating around in the rumor mill currently.

For reference:
RTX 2060: $349
RTX 2070: $499
RTX 2080: $699

RTX 3060: $399?
RTX 3070: $499
RTX 3080: $699

Radeon RX 5500 XT: $199
Radeon RX 5600 XT: $279
Radeon RX 5700 XT: $399

EDIT: Some time ago, there was a Open VR benchmark that leaked and showed an AMD GPU scoring 17% higher than the 2080TI. That was probably an early engineering sample of Navi 21. The author of the benchmark confirmed it was legit. Right on the heals of the 3080 even in January. Who knows how far they have progressed now?

 

[Glo.]

Every single piece of evidence I've seen so far indicates that what I says is mostly accurate. Even a new article VideoCardz just published hints that I'm on the right path, though he thinks that the large die will be for the RX 6900 XT only, and the 6800XT and 6800XT will share the medium die, and the 6600 XT/6500 XT will share the small die.

You do realize that there are FOUR Navi 2 dies, right?

80 CU die - X900 SKUs
60 CU die - X800 SKU
40 CU die - X700/600 SKUs
smallest die - X500 SKUs.

 

[eek2121]

You do realize that there are FOUR Navi 2 dies, right?

80 CU die - X900 SKUs
60 CU die - X800 SKU
40 CU die - X700/600 SKUs
smallest die - X500 SKUs.

So where is the 4th? I'm seeing nothing in driver source code and little mention of it anywhere. There may be a 4th eventually, but I suspect AMD is finally planning on retiring Polaris, so we'll see a backfill of even lower tier cards.

EDIT: Speaking of which, AMD just added (probably initial) Van Gogh support to Mesa 4 hours ago.

EDIT: For the record:

 switch (amdinfo->family_id) {
   case FAMILY_SI:
      identify_chip(TAHITI);
      identify_chip(PITCAIRN);
      identify_chip2(CAPEVERDE, VERDE);
      identify_chip(OLAND);
      identify_chip(HAINAN);
      break;
   case FAMILY_CI:
      identify_chip(BONAIRE);
      identify_chip(HAWAII);
      break;
   case FAMILY_KV:
      identify_chip2(SPECTRE, KAVERI);
      identify_chip2(SPOOKY, KAVERI);
      identify_chip2(KALINDI, KABINI);
      identify_chip2(GODAVARI, KABINI);
      break;
   case FAMILY_VI:
      identify_chip(ICELAND);
      identify_chip(TONGA);
      identify_chip(FIJI);
      identify_chip(POLARIS10);
      identify_chip(POLARIS11);
      identify_chip(POLARIS12);
      identify_chip(VEGAM);
      break;
   case FAMILY_CZ:
      identify_chip(CARRIZO);
      identify_chip(STONEY);
      break;
   case FAMILY_AI:
      identify_chip(VEGA10);
      identify_chip(VEGA12);
      identify_chip(VEGA20);
      identify_chip(ARCTURUS);
      break;
   case FAMILY_RV:
      identify_chip(RAVEN);
      identify_chip(RAVEN2);
      identify_chip(RENOIR);
      break;
   case FAMILY_NV:
      identify_chip(NAVI10);
      identify_chip(NAVI12);
      identify_chip(NAVI14);
      identify_chip(SIENNA_CICHLID);
      identify_chip(NAVY_FLOUNDER);
      identify_chip(DIMGREY_CAVEFISH);
      break;
   case FAMILY_VGH:
      identify_chip(VANGOGH);
      break;
   }