Discussion RDNA4 + CDNA3 Architectures Thread

DisEnchantment · Mar 23, 2022

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

ToTTenTranz · Mar 3, 2025

SolidQ said:
Interesting what is AIB with those high memory temps
P.S I don't think it's with Honeywell PTM7950

Is that on N48 cards?

Perhaps because the memory chips aren't in contact with the heatsinks?

SolidQ · Mar 3, 2025

ToTTenTranz said:
Is that on N48 cards?

Yeah N48. on 9070 GPU/Hotspot very good. We need to see XFX/Sapphire/Powercolor temps

PJVol · Mar 3, 2025

SolidQ said:
I don't think it's with Honeywell PTM7950

Afaik PTM is for gpu only. Vram chips TIM is used to be a regular thermal pads ca. 1 mm thick. Wondering what's the highest DPM VID for VDD_MEM ?
Are these temps normal for 20gbs chips?

SolidQ · Mar 3, 2025

PJVol said:
Are these temps normal for 20gbs chips?

Dunno.
I'm looking at 21GB GDDR6X and here is better temps

at least we need looking into different AIB models. 2 day left

Tup3x · Mar 3, 2025

When the AIBs finally fix the hotspot temp issue using PTM, they screw up the memory cooling. That would be something and hopefully that's not actually the case.

MrTeal · Mar 3, 2025

Win2012R2 said:
Demand is likely to be much higher

Way higher. Summer 2023 it was not hard to find a 4070, 6800 XT's were still being firesaled out, and used mining Ampere was everywhere.
It's a desert out there right now.

PJVol · Mar 3, 2025

SolidQ said:
Dunno.
I'm looking at 21GB GDDR6X and here is better temps

TPU has not provided data on VRAM hotspot t° on radeons, so I had to do a bit of searching and found it in a PCGH review. The temperatures are very similar to those on the rdna 3 cards with the same 20gbs chips

UPD: This is why I'll never go back to air-cooled gpus (7 years with EK full cover wb on vega-rdna1-rdna2 )

MrTeal · Mar 3, 2025

Abwx said:
I dont see how it could be otherwise for redundancy purposes and cost, controling each cable current would increase cost too much and wouldnt solve anything since
the problem for Nvidia FI is that the connectors have a loosy grip on the cards pins.

Only thing that could be done at very low cost is probing the currents by measuring the voltage drops between the PCB individual pins and the commoned track, this allow to know if your connector has the required contact torque for each pin.

You don't explicitly control each conductor current, but you route different pins to different phases. It's the same setup as they would use for a 3x8pin card, except that instead of running each 8 pin to 1/3rd of the phases you run a pair from the 12 pin to 1/3rd of the phases.
You could get fancy with a crossbar system that allows any phase (or group of phases) to be fed from any of the 12V rails for additional cost and complexity and a slight efficiency hit, but that's not necessary. Just hardwiring it has 0 cost over pooling it other than some duplicated input filtering. As you say there's some loss of redundancy as if you have a couple bad contacts on the pair the feeds one set of phases you're either going to have to lower the draw of those phases, shut them down or just put the GPU into limp mode. That's the same effect as if you lose one of your 8 pin PCIe cables though, and realistically having to shut down the GPU to fix your cable even if it's on something like a 5070 Ti or 9070 XT where it would happily run with 3 or 4 good pairs of wire isn't necessarily a bad thing.

SolidQ · Mar 3, 2025

PJVol said:
The temperatures are very similar to those on the rdna 3 cards with the same 20gbs chips

yeah, but too big difference now between GPU/Hotspot and Memory

Abwx · Mar 3, 2025

MrTeal said:
You don't explicitly control each conductor current, but you route different pins to different phases. It's the same setup as they would use for a 3x8pin card, except that instead of running each 8 pin to 1/3rd of the phases you run a pair from the 12 pin to 1/3rd of the phases.
You could get fancy with a crossbar system that allows any phase (or group of phases) to be fed from any of the 12V rails for additional cost and complexity and a slight efficiency hit, but that's not necessary. Just hardwiring it has 0 cost over pooling it other than some duplicated input filtering. As you say there's some loss of redundancy as if you have a couple bad contacts on the pair the feeds one set of phases you're either going to have to lower the draw of those phases, shut them down or just put the GPU into limp mode. That's the same effect as if you lose one of your 8 pin PCIe cables though, and realistically having to shut down the GPU to fix your cable even if it's on something like a 5070 Ti or 9070 XT where it would happily run with 3 or 4 good pairs of wire isn't necessarily a bad thing.

So if half the cables have loosy contact not only half the other cables but also half the phases count have to provide most of the power to the GPU, so each phase has to be overdimensioned in case there s intermitent bad contacts in one or worse, in several pins, that wouldnt make sense cost and reliability wise.

gaav87 · Mar 3, 2025

SolidQ said:
yeah, but too big difference now between GPU/Hotspot and Memory

I think its the acer 9070xt. Its 3100mhz 340W with 2slot cooler
Or its asus prime cause napoleon had that
Or its gigabyte cause i think they used thermal putty instead of pads for mem cooling
I know when i browsed all cards to gather size data i saw one use putty for vram instead of pads.

Hail The Brain Slug · Mar 3, 2025

gaav87 said:
I think its the acer 9070xt. Its 3100mhz 340W with 2slot cooler
Or its asus prime cause napoleon had that
Or its gigabyte cause i think they used thermal putty instead of pads for mem cooling
I know when i browsed all cards to gather size data i saw one use putty for vram instead of pads.

Thermal putty should perform better than pads.

Hans Gruber · Mar 3, 2025

Will the 9070xt equal the 7900xtx in performance?

gdansk · Mar 3, 2025

Hans Gruber said:
Will the 9070xt equal the 7900xtx in performance?

No, it'll be a bit worse. But pretty close. Maybe they are "perceptually equal" outside of RT where it stands to be ahead of last gen.

Hans Gruber · Mar 3, 2025

gdansk said:
No, it'll be a bit worse. But pretty close. Maybe they are "perceptually equal" outside of RT where it stands to be ahead of last gen.

The theory was the 7900xtx had a design flaw that made rougly 20% of the processor gimped/non functional. Is RDNA 4 based on the same architecture of RDNA 3 but with the non working part of the silicon fully functioning? I know RDNA 4 has FSR 4 and other enhancements. Or is it an all new design? So in theory the 9070xt could gain close to 20% of free performance even though it's a weaker GPU than the 7900xtx.

I saw this because AMD said they were not pushing the high end GPU market with RDNA 4. It seems they are confident that RDNA 4 will exceed their expectations and I was wondering if the deactivated silicon that AMD denied was ever an issue could give the 9070xt some free performance upgrades with fully functioning silicon.

I am not a conspiracy theoriest. The 7900xtx was supposed to be a monster and internet sleuths brought up the missing performance a few years back. It makes me wonder because of new found hype for 9070xt performance.

gdansk · Mar 3, 2025

Hans Gruber said:
Is RDNA 4 based on the same architecture of RDNA 3

It's quite different. They have sent out slides to the media going over the differences:

AMD Radeon RX 9070 Series Technical Deep Dive

AMD's next-generation of GPUs is launching very soon. We have all the details on the architecture and the new models. Our preview article also talks about new new FSR 4 upscaling algorithm, AMD's new naming scheme, Radeon Software and the improvements for HYPR-RX.

www.techpowerup.com

eek2121 · Mar 3, 2025

GTracing said:
I wouldn't pair a 9070 XT with a 650W unless the PSU was a really nice one. They suggest higher wattage PSUs for a reason. Transient spikes can be way higher than the TDP of the GPU. The RTX 3000 series was where this issue first got some attention.

I ran a 4090 on a 650W PSU when it was in my ITX case. Torture testing got the entire 7950X system up to 644W, but normal scenarios never approached even 600W.

I only moved cases because the GPU was mounted vertically and caused the chipset fan to run full speed. In my big case I have a 1000W PSU.

MrTeal · Mar 3, 2025

Abwx said:
So if half the cables have loosy contact not only half the other cables but also half the phases count have to provide most of the power to the GPU, so each phase has to be overdimensioned in case there s intermitent bad contacts in one or worse, in several pins, that wouldnt make sense cost and reliability wise.

That's how the Ampere cards were configured, and it's also how cards with multiple 6/8pin connectors are configured. You don't overprovision your output stages to handle losing a 12V input, you just don't run.

Either way, you haven't said anything to support this.

Abwx said:
There s no current balancing as such, in a 2 x 6 pin connector the 6 pins that handle the positive rail are connected in parralel in the PCB, and the same for the 6 of the negative rail, if the currents are unbalanced it just mean that there s some pins of the card that are not in tight contact with the cable connector pins, hence most of the current flow in the pins that are connected the tightest.

Do you have any evidence that all the pins are connected in parallel on the Nitro+ PCB? I'm not saying they aren't since it's quite possible they are, you just stated it as fact and I'd like to know how they implemented it.

SolidQ · Mar 3, 2025

Fabio, mentioned AMD improve RDNA3 RT perfomance in new driver

HuesToo4 · Mar 3, 2025

SolidQ said:
Fabio, mentioned AMD improve RDNA3 RT perfomance in new driver

Anything on FSR 4 support for RDNA 3?

SolidQ · Mar 3, 2025

HuesToo4 said:
Anything on FSR 4 support for RDNA 3?

still in progress (c)

Hail The Brain Slug · Mar 3, 2025

MrTeal said:
I'd like to know how they implemented it.

Thats what we'd all like to know yet people keep answering it without having the answer.

Josh128 · Mar 3, 2025

Hans Gruber said:
Will the 9070xt equal the 7900xtx in performance?

Gonna be a bit less (probably ~10% or so) in raster, hopefully a bit faster in RT.

carrotmania · Mar 3, 2025

VERY rough numbers, but a hypothetical 96CU 9080XT could looked like this:

Imagine a full fat 128CU 9090XT

Josh128 · Mar 3, 2025

Hans Gruber said:
The theory was the 7900xtx had a design flaw that made rougly 20% of the processor gimped/non functional. Is RDNA 4 based on the same architecture of RDNA 3 but with the non working part of the silicon fully functioning? I know RDNA 4 has FSR 4 and other enhancements. Or is it an all new design? So in theory the 9070xt could gain close to 20% of free performance even though it's a weaker GPU than the 7900xtx.

I saw this because AMD said they were not pushing the high end GPU market with RDNA 4. It seems they are confident that RDNA 4 will exceed their expectations and I was wondering if the deactivated silicon that AMD denied was ever an issue could give the 9070xt some free performance upgrades with fully functioning silicon.

I am not a conspiracy theoriest. The 7900xtx was supposed to be a monster and internet sleuths brought up the missing performance a few years back. It makes me wonder because of new found hype for 9070xt performance.

The only thing we can be fairly sure of regarding 7900XTX is that they missed their published "architected for" frequency target (~3GHZ) for the RDNA3 architecture by a good margin. It released with a boost clock of 2.5GHz, 20% below 3GHZ. RDNA 4 has achieved 3GHz+. Whether that was related to MCM vs monolithic design or something else, is unclear.

Discussion RDNA4 + CDNA3 Architectures Thread

Golden Member

Senior member

Golden Member

Senior member

Golden Member

Golden Member

Diamond Member

Senior member

Diamond Member

Golden Member

Lifer

Senior member

Diamond Member

Platinum Member

Diamond Member

Platinum Member

Diamond Member

Diamond Member

Diamond Member

Golden Member

Junior Member

Golden Member

Diamond Member

Senior member

Member

Senior member