Discussion RDNA4 + CDNA3 Architectures Thread

[DisEnchantment]

With the GFX940 patches in full swing since first week of March, it is looking like MI300 is not far in the distant future!
Usually AMD takes around 3Qs to get the support in LLVM and amdgpu. Lately, since RDNA2 the window they push to add support for new devices is much reduced to prevent leaks.
But looking at the flurry of code in LLVM, it is a lot of commits. Maybe because US Govt is starting to prepare the SW environment for El Capitan (Maybe to avoid slow bring up situation like Frontier for example)

See here for the GFX940 specific commits

History for llvm/lib/Target/AMDGPU - llvm/llvm-project

The LLVM Project is a collection of modular and reusable compiler and toolchain technologies. - History for llvm/lib/Target/AMDGPU - llvm/llvm-project

github.com

Or Phoronix

More AMD "GFX940" Enablement Work Landing In LLVM - Phoronix

www.phoronix.com

There is a lot more if you know whom to follow in LLVM review chains (before getting merged to github), but I am not going to link AMD employees.

I am starting to think MI300 will launch around the same time like Hopper probably only a couple of months later!
Although I believe Hopper had problems not having a host CPU capable of doing PCIe 5 in the very near future therefore it might have gotten pushed back a bit until SPR and Genoa arrives later in 2022.
If PVC slips again I believe MI300 could launch before it

This is nuts, MI100/200/300 cadence is impressive.



Previous thread on CDNA2 and RDNA3 here

Question - Speculation: RDNA3 + CDNA2 Architectures Thread

Man I have been dying to make this one for a while now. First rumours for RDNA3 are here so new thread time! Just going to start off with this one for now: kopite7kimi on Twitter: "@VideoCardz Ah, I mean a simple mcm design with 10240 cores is not enough. Because the lift from RDNA2 to RDNA3...

forums.anandtech.com

 

[poke01]

https://www.reddit.com/r/radeon/comments/1j6qja7/amd_fsr4_outperformed_dlss4_in_quality_while_the

Very nice

 

[poke01]

Iv also been hearing alot of people have stutter and black screen issues. Im on windows10 its super smooth

Interesting, tho I think its very rare compared to blackwell black screen

 

[Keller_TT]

I was able to get a ASUS 9070 Prime for 530 CHF from a retailer buddy, which is about $600, and it's into a build for the kids with a Zen 4 7700 in it, and it's looking great! The 9070 with UV and clock tweaks, power limited to 250W is pretty much 9070 XT stock level. I spent the most of yesterday testing stability and performance and it is well under limits and runs silently.
I'm a sucker for efficiency and this gave me great joy. And now I'm up early installing games. lol.

The PC didn't have a graphics card in it, the kids have their PS5 and I had decided to bite on the PS5 Pro and sell the original for 350 bucks. The Pro and Disc drive would cost me 500 Francs more and I was Ok as i wouldn't be looking around until 2027 Holiday season.
But the 9070 efficiency really impressed, I couldn't pass up on one of AMD's most impressive cards ever, and getting hardware has never been a problem. I am lucky to have a couple of good friends in the business. All worked out well in the end.

 

[Racan]

GitHub - FakeMichau/fakenvapi: NVAPI library for spoofing an Nvidia GPU. With support for AntiLag 2 and LatencyFlex

NVAPI library for spoofing an Nvidia GPU. With support for AntiLag 2 and LatencyFlex - FakeMichau/fakenvapi

github.com

Unfortunately native AntiLag 2 support in games is still abysmal.

But it looks like you can drop this .dll file together with OptiScaler to replace Nvidia Reflex with AL 2.

 

[SolidQ]

Interesting, tho I think its very rare compared to blackwell black screen

yeah Rare, but i saw few posts, and that was cable problem

 

[gaav87]

Anyway this taichi is kinda crazy (i did ml translation from polish thats why its gargled)
I can go up to 400W or 220W with almost same perf...
I matched 4090 or even beat some 4090 in monster hunter 1080p ultra
matched oced 4080S in forza horizon 4k maxed out RT while only 36% slower than 4090 on mora...
10% over 5070Ti in RT cyberpunk
This card can go +-15% from base depending how much power u wana throw at it.

 

[TESKATLIPOKA]

I don't know about you guys, but I am kind of more excited about N44 than N48, but of course I am talking about the 16GB version.
Hopefully AMD will stick with 20gbps GDDR6 at least for the 16GB version, which would provide 11% higher BW compared to predecessor.

Known models:

Model	GPU specs (CU, Shaders, TMU, ROPs)	Frequency (Boost)	Infinity Cache	Memory (Bandwidth)	Vram	Total Board Power	Performance (TPU 4K)
RX 7600	32:2048:128:64	1720 MHz (2655 MHz)	32 MB	128-bit 18gbps (288 GB/s)	8 GB	165 W (62.7 %)	41.8 %
RX 7600 XT	32:2048:128:64	1720 MHz (2755 MHz)	32 MB	128-bit 18gbps (288 GB/s)	16 GB	190 W (72.2 %)	56.7 %
RX 7700 XT	54:3456:216:96	1900 MHz (2544 MHz)	48 MB	192-bit 18gbps (432 GB/s)	12 GB	245 W (93.2 %)	82.1 %
RX 7800 XT	60:3840:240:96	1800 MHz (2430 MHz)	64 MB	256-bit 19.5gbps (624 GB/s)	16 GB	263 W (100 %)	100 %
RX 9700	56:3584:224:128	2070 MHz ( 2520 MHz)	64 MB	256-bit 20gbps (640 GB/s)	16 GB	220 W (83.7 %)	130 %
RX 9700 XT	64:4096:256:128	2400 MHz ( 2970 MHz)	64 MB	256-bit 20gbps (640 GB/s)	16 GB	304 W (115.6 %)	149.3 %

N44 could look like this ->

Model	GPU specs (CU, Shaders, TMU, ROPs)	Frequency (Boost)	Infinity Cache	Memory (Bandwidth)	Vram	Total Board Power	Performance (TPU 4K)
RX 9600	32:2048:128:64	2070 MHz ( 2520 MHz)	32 MB	128-bit 18gbps (288 GB/s)	8 GB	125 W	?
RX 9600 XT	32:2048:128:64	2400 MHz ( 2970 MHz)	32 MB	128-bit 20gbps (320 GB/s)	16 GB	165 W	75% ?

Performance is pretty questionable. At these "conservative" clocks It won't hit 7700XT level of performance. I think 75% of RX 7800XT is more realistic or 32% over RX 7600XT for the 16GB version.
Of course AMD will push clocks hard for the extra performance, even 3.2-3.3GHz shouldn't be out of the question and at those clocks It should hit 7700XT level of performance, but TBP could hit 200W and efficiency will take a nosedive, although It will still be better than 7700XT.

P.S. I kinda wonder why no one is using GDDR6W at least in laptops. It saves space, because per chip you have double capacity and the height is also smaller.

 

[GaiaHunter]

I'm testing stability at -100mV. It does hit 3400 MHz with just the undervolt.

 

[marees]

I don't know about you guys, but I am kind of more excited about N44 than N48, but of course I am talking about the 16GB version.
Hopefully AMD will stick with 20gbps GDDR6 at least for the 16GB version, which would provide 11% higher BW compared to predecessor.

Known models:

Model	GPU specs (CU, Shaders, TMU, ROPs)	Frequency (Boost)	Infinity Cache	Memory (Bandwidth)	Vram	Total Board Power	Performance (TPU 4K)
RX 7600	32:2048:128:64	1720 MHz (2655 MHz)	32 MB	128-bit 18gbps (288 GB/s)	8 GB	165 W (62.7 %)	41.8 %
RX 7600 XT	32:2048:128:64	1720 MHz (2755 MHz)	32 MB	128-bit 18gbps (288 GB/s)	16 GB	190 W (72.2 %)	56.7 %
RX 7700 XT	54:3456:216:96	1900 MHz (2544 MHz)	48 MB	192-bit 18gbps (432 GB/s)	12 GB	245 W (93.2 %)	82.1 %
RX 7800 XT	60:3840:240:96	1800 MHz (2430 MHz)	64 MB	256-bit 19.5gbps (624 GB/s)	16 GB	263 W (100 %)	100 %
RX 9700	56:3584:224:128	2070 MHz ( 2520 MHz)	64 MB	256-bit 20gbps (640 GB/s)	16 GB	220 W (83.7 %)	130 %
RX 9700 XT	64:4096:256:128	2400 MHz ( 2970 MHz)	64 MB	256-bit 20gbps (640 GB/s)	16 GB	304 W (115.6 %)	149.3 %

N44 could look like this ->

Model	GPU specs (CU, Shaders, TMU, ROPs)	Frequency (Boost)	Infinity Cache	Memory (Bandwidth)	Vram	Total Board Power	Performance (TPU 4K)
RX 9600	32:2048:128:64	2070 MHz ( 2520 MHz)	32 MB	128-bit 18gbps (288 GB/s)	8 GB	125 W	?
RX 9600 XT	32:2048:128:64	2400 MHz ( 2970 MHz)	32 MB	128-bit 20gbps (320 GB/s)	16 GB	165 W	75% ?

Performance is pretty questionable. At these "conservative" clocks It won't hit 7700XT level of performance. I think 75% of RX 7800XT is more realistic or 32% over RX 7600XT for the 16GB version.
Of course AMD will push clocks hard for the extra performance, even 3.2-3.3GHz shouldn't be out of the question and at those clocks It should hit 7700XT level of performance, but TBP could hit 200W and efficiency will take a nosedive, although It will still be better than 7700XT.

P.S. I kinda wonder why no one is using GDDR6W at least in laptops. It saves space, because per chip you have double capacity and the height is also smaller.

I guess it beats the 4060 ti in both raster & RT

That means it is between the 5060 & 5060 ti

 

[SolidQ]

I am kind of more excited about N44 than N48

are we sure 9060 is Navi44, not 9050?

 

[TESKATLIPOKA]

are we sure 9060 is Navi44, not 9050?

Why would It be 9050? AMD stopped using *500 in the previous generation.

 

[jpiniero]

are we sure 9060 is Navi44, not 9050?

9050 feels like more likely to be a rebrand.

 

[SolidQ]

Why would It be 9050? AMD stopped using *500 in the previous generation.

NVIDIA also doesn't have DS 4050, but now want to do RTX 5050

 

[jpiniero]

NVIDIA also doesn't have DS 4050, but now want to do RTX 5050

And it's gonna be hella slow. A rebrand of the 7600 would be fine enough competitor.

 

[TESKATLIPOKA]

NVIDIA also doesn't have DS 4050, but now want to do RTX 5050

But we are talking about AMD here not Nvidia.
And you still didn't say why It would be called 9050?

BTW Nvidia has RTX 4050 Laptop, they just didn't release a cutdown AD107 for desktop.

 

[SolidQ]

And you still didn't say why It would be called 9050?

Maybe for bottom perf? With 9060 AMD need 7700XT and 7800XT perf, for below perf should be 9050
5060ti should be like ~4070 perf, 5060 ~4060ti

 

[TESKATLIPOKA]

I guess it beats the 4060 ti in both raster & RT

That means it is between the 5060 & 5060 ti

5060Ti is likely a full GB206 with 36SM, but what is this RTX 5060?

 

[Josh128]

I had never realized that 7600XT (204mm) is so much smaller than the 6700XT (354mm) and is only ~10-15% slower. Thats a huge die size reduction on just a tweak of the same 7nm process for very little performance loss. I dont know if they'll do it, but a ~ 150-175mm RDNA4 die, clocked to the hilt, would probably be ~10-20% faster than the 6700XT and quite a potent little entry level GPU that could also be pretty affordable. If they really want some market share, that would be the way to go, IMO.

And it's gonna be hella slow. A rebrand of the 7600 would be fine enough competitor

 

[TESKATLIPOKA]

I had never realized that 7600XT (204mm) is so much smaller than the 6700XT (354mm) and is only ~10-15% slower. Thats a huge die size reduction on just a tweak of the same 7nm process for very little performance loss. I dont know if they'll do it, but a ~ 150-175mm RDNA4 die, clocked to the hilt, would probably be ~10-20% faster than the 6700XT and quite a potent little entry level GPU that could also be pretty affordable. If they really want some market share, that would be the way to go, IMO.

N22 had too much BW(192-bit controller) and Infinity cache(96MB) compared to Its raw power.
N44 shouldn't be only ~ 150-175mm, but >180mm2 in my opinion.

 

[Josh128]

N22 had too much BW(192-bit controller) and Infinity cache(96MB) compared to Its raw power.
N44 shouldn't be only ~ 150-175mm, but >180mm2 in my opinion.

I would think 180mm would be the limit, unless they are planning a yet smaller N42 (which I doubt). 6700XT/2080Ti perf is still very capable even today at 1440p and 1080p. Plus you'd get better RT, FSR 4, and presumably the new video engine. 180mm or less would allow for some fantastic yields and # of die per wafer, and thus price. The last thing they want to do is get upended by Nvidia on 50/60 class pricing due to silicon/production costs. Slap a $279 price tag on it and watch gamers gobble it up like Puck Man.

Adroc previously seemed to indicate that N44 would be a massive disappointment, Im not sure if what I am envisioning is way beyond what it would actually be or not, but it doesnt seem like a very high bar for AMD. 🤷‍♀️

 

[gaav87]

I'm testing stability at -100mV. It does hit 3400 MHz with just the undervolt.

I did extensive testing since morning.
Mem sweet spot is 2764mhz fast timings (any higher or lower u lose perf)
-65mV is rock stable at least on 4 cards
-100mv vs 65mV gives 2% perf uplift but its not stable for me. 90mV is stable but not furmark/3dmark stable
+450mhz offset for core is sweetspot
+6-8% power draw is sweet spot

Full monster OC vs this above settings is 1.5% diff
380W vs ~340W full load with RT or (288-300w in light or no rt games)

Try this settings i already gave them to multiple people one guy gained 16% from stock hellhound with this settings.

 

[jpiniero]

https://videocardz.com/newz/amd-adeon-rx-9600-xt-confirmed-with-16gb-and-8gb-gddr6-memory-sticking-to-128-bit-memory-bus

Does seem like AMD is going with the 9060 XT only to start (with 16 and 8 GB options)

 

[TESKATLIPOKA]

I would think 180mm would be the limit, unless they are planning a yet smaller N42 (which I doubt). 6700XT/2080Ti perf is still very capable even today at 1440p and 1080p. Plus you'd get better RT, FSR 4, and presumably the new video engine. 180mm or less would allow for some fantastic yields and # of die per wafer, and thus price. The last thing they want to do is get upended by Nvidia on 50/60 class pricing due to silicon/production costs. Slap a $279 price tag on it and watch gamers gobble it up like Puck Man.

Navi 48 is 356.5 mm2.
Half of that is already 178.25mm2 and don't forget not everything is halved(PCIe, video engine etc.), that's why >180mm2. I think over 190mm2 is very likely.

It will perform a lot better than RX 6700XT, which is only 13% faster than 7600XT 16GB at QHD or 4K.

Even a 200mm2 chip is pretty cheap to produce, but the 16GB version will certainly cost more than $279, more like $379-399 MRSP.

 

[GaiaHunter]

I did extensive testing since morning.
Mem sweet spot is 2764mhz fast timings (any higher or lower u lose perf)
-65mV is rock stable at least on 4 cards
-100mv vs 65mV gives 2% perf uplift but its not stable for me. 90mV is stable but not furmark/3dmark stable
+450mhz offset for core is sweetspot
+6-8% power draw is sweet spot

Full monster OC vs this above settings is 1.5% diff
380W vs ~340W full load with RT or (288-300w in light or no rt games)

Try this settings i already gave them to multiple people one guy gained 16% from stock hellhound with this settings.

It hasn't crashed in the games I played yet at -100mv (Guild Wars 2, Borderlands 3, MW5 and RDR2).

In RDR2 it hit 3400 but didn't stay there long, mostly it hovers around 3300 MHz.

I'll play with memory tomorrow. Don't have more time today, wife is starting to give me the stink eye...

 

[maddie]

N22 had too much BW(192-bit controller) and Infinity cache(96MB) compared to Its raw power.
N44 shouldn't be only ~ 150-175mm, but >180mm2 in my opinion.

Agree. Roughly 1/2 (N48 - video display/encoder engines) + N48 video display/encoders.

Performance of XT model could approach RX 6800 (=1.33 RX 7600 XT) as the smaller die normally clock higher than bigger models.