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

 

[jpiniero]

If RDNA4 supports GDDR7, maybe they will save the busted IMC chips for then.

"You don't want 8 GB? Here's... 9."

 

[jpiniero]

As for the theoretical MSRP of the 9070 GRE... do note that the 5070, which there was a trickle of $549 cards available for a bit, seems to have ended for the most part. That might be because of tariffs tho. Now it's more like starting at $600 and only going up from there.

 

[Win2012R2]

If RDNA4 supports GDDR7

Why would AMD pay extra money for IP of controller that supports that when they had zero plans to use that memory and targeted cheap product, plus validation etc? There is zero chance 9000 series will support GDDR7, and even if it did AMD will keep this boost dry for RDNA5.

 

[marees]

As for the theoretical MSRP of the 9070 GRE... do note that the 5070, which there was a trickle of $549 cards available for a bit, seems to have ended for the most part. That might be because of tariffs tho. Now it's more like starting at $600 and only going up from there.

The 9070 gre is china only as of now (just like the 7900 gre was initially)

 

[GTracing]

A denialist, okay pso.

Edit. I really must add but how on earth can you state the 7600XT is less than 33% faster. The smallest delta is Spider Man 2 which is around 35%. Every other example is greater than that, some cases are > 100%.

Doom Eternal is 63.9 FPS Vs 17.7 at 1440p
F1 24 RT is 47.3 Vs 18.1 at 1080p
TLOU is 62.7 Vs 41.7 at 1080p
Spider Man 2 is 69.9 Vs 51.9 at 1080p
F1 24 is 59.4 Vs 23.9 at 4k (some people plug their PC into a TV)

Like how can I take anything you say seriously when it is so divorced from reality.

This is also not even looking at 1% lows where I bet the deltas are even larger as we see with 5060Ti 8GB Vs 16GB comparisons.

Just saw this edit now. I had thought that the screenshots you shared were comparing the 5060 ti 8GB to 5060 ti 16 GB. Apologies.

I'll give you TLOU and Spiderman, but most of your examples don't fit the criteria I gave.

Doom Eternal was tested at "Ultra Nightmare", which increases the texture pool size over "Ultra", this makes it run like crap on 8GB cards, but has practically no effect on graphical fidelity. That's why I specifically mentioned it in my original comment, I don't know why you chose to use it as an example.

 

[jpiniero]

Why would AMD pay extra money for IP of controller that supports that when they had zero plans to use that memory and targeted cheap product, plus validation etc? There is zero chance 9000 series will support GDDR7, and even if it did AMD will keep this boost dry for RDNA5.

The original plan for the higher end models probably intended to use GDDR7.

 

[Timorous]

Just saw this edit now. I had thought that the screenshots you shared were comparing the 5060 ti 8GB to 5060 ti 16 GB. Apologies.

I'll give you TLOU and Spiderman, but most of your examples don't fit the criteria I gave.

Doom Eternal was tested at "Ultra Nightmare", which increases the texture pool size over "Ultra", this makes it run like crap on 8GB cards, but has practically no effect on graphical fidelity. That's why I specifically mentioned it in my original comment, I don't know why you chose to use it as an example.

Ultra nightmare reduces popin, hard to show in a bar chart though, also the example was with RT enabled.

Your criteria was nonsense. Also why a 33% performance delta for a 20% price difference? It was just made up to make a ridiculous argument that is easily disproved.

Further more as stated 1% lows are often a bigger issue and I CBA to go through TPU, tech spot etc for the examples. Also plenty of titles suffer IQ degradation because they don't render the game at the chosen settings so you won't see those in bar charts either.

I provided 6 examples very quickly, I could get out space marine 2, hogwarts, forspoken, halo infinite just off of the top of my head of titles with worse IQ on 8GB models and I am not even looking very hard at all.

 

[GTracing]

Ultra nightmare reduces popin, hard to show in a bar chart though, also the example was with RT enabled.

Your criteria was nonsense. Also why a 33% performance delta for a 20% price difference? It was just made up to make a ridiculous argument that is easily disproved.

Further more as stated 1% lows are often a bigger issue and I CBA to go through TPU, tech spot etc for the examples. Also plenty of titles suffer IQ degradation because they don't render the game at the chosen settings so you won't see those in bar charts either.

I provided 6 examples very quickly, I could get out space marine 2, hogwarts, forspoken, halo infinite just off of the top of my head of titles with worse IQ on 8GB models and I am not even looking very hard at all.

You're the one picked 33% not me. I readily admit that you can find examples where 16GB cards are massively faster than 8GB cards. My original point was there are not "dozens" (your words) of games where budget 16GB cards are 33%+ faster than their 8GB counterparts at the settings where budget cards are actually used. If you don't agree with that criteria, then you could have just said so from the beginning rather than sending a bunch of examples that don't fit.

 

[Timorous]

You're the one picked 33% not me. I readily admit that you can find examples where 16GB cards are massively faster than 8GB cards. My original point was there are not "dozens" (your words) of games where budget 16GB cards are 33%+ faster than their 8GB counterparts at the settings where budget cards are actually used. If you don't agree with that criteria, then you could have just said so from the beginning rather than sending a bunch of examples that don't fit.

It was not me who mentioned 33%. It was another poster who made the absurd claim that a 25% cut to memBW would reduce performance across the board by 25% which we know is wrong because the 7900GRE has less memBW than the 7800XT and performs a better.

Point is though there are dozens where a 16GB (or 12GB part) is playable but 8GB is not. I linked to an article that shows this with the 5060Ti 8GB Vs 16GB which is in a similar performance tier to where the 9060XT is going to land. There were multiple cases where the performance difference was over 100% in favour of the 16GB card. Cases where a 12GB part, would bot suffer anywhere near as badly despite the memBW cut.

If you choose not to read then sure thing, go watch the HUB 8GB review video where he goes through lots of games and lots of settings in those games.

As for your contrived 'at settings people actually use' criteria, it is a nonsense way to discard examples you don't like.

Edit to add: we know the 25% memBW reduction won't lead to a 25% lower performance because of the 9070GRE. That has a 25% cut bus and 10% slower VRAM than the 9070 while also having less compute and it is around 18% lower performance.

A 12GB N44 part with 20gbps ram has a smaller BW deficit to the 9060XT than the 9070 GRE does to the 9070 so the performance drop off should be closer to 15% give or take. That would be between 3060Ti and 4060Ti performance which is actually faster than the 6700XT.

 

[GTracing]

As for your contrived 'at settings people actually use' criteria, it is a nonsense way to discard examples you don't like.

You don't think that people would rather have 117fps with RT off, than 47fps with it on? In a racing sim no less?

 

[adroc_thurston]

Why would AMD pay extra money for IP of controller that supports that when they had zero plans to use that memory and targeted cheap product, plus validation etc?

because RDNA4 was not supposed to be cheap.

There is zero chance 9000 series will support GDDR7, and even if it did AMD will keep this boost dry for RDNA5.

really funny right

 

[Timorous]

You don't think that people would rather have 117fps with RT off, than 47fps with it on? In a racing sim no less?

Calling F1 24 a racing SIM is laughable.

People who play it more seriously as an e-sport title will want 120fps +.

People who play it casually as a bit of fun and just do the career mode could very well settle with 47 FPS. Or they could reduce a couple of settings to hit 60. Far easier to turn on RT and tune things from a baseline or 47fps than 18fps like the 8GB version gets.

 

[GTracing]

Calling F1 24 a racing SIM is laughable.

People who play it more seriously as an e-sport title will want 120fps +.

People who play it casually as a bit of fun and just do the career mode could very well settle with 47 FPS. Or they could reduce a couple of settings to hit 60. Far easier to turn on RT and tune things from a baseline or 47fps than 18fps like the 8GB version gets.

You admit that people will turn down settings. Why is it "contrived" when I say it? I don't think you're arguing in good faith. Regardless, this conversation is going nowhere, so I'm probably done responding.

 

[dr1337]

You admit that people will turn down settings. Why is it "contrived" when I say it? I don't think you're arguing in good faith. Regardless, this conversation is going nowhere, so I'm probably done responding.

In a game where everyone turns down settings, the vram argument really isn't poignant. Except for the fact that texture quality will probably be lower due to dynamic compression effects. And if you wanna talk good faith, why aren't you considering their original argument? 7600xt 8gb gets 30fps in F1, 16gb gets 60fps. Even the difference between the 5060ti 8/16gb is over 30fps. Which one do you think will be faster at lower settings? What if all the important settings already are set to low in the TPU test?

but has practically no effect on graphical fidelity.

yes it does

 

[GTracing]

In a game where everyone turns down settings, the vram argument really isn't poignant. Except for the fact that texture quality will probably be lower due to dynamic compression effects. And if you wanna talk good faith, why aren't you considering their original argument? 7600xt 8gb gets 30fps in F1, 16gb gets 60fps. Even the difference between the 5060ti 8/16gb is over 30fps. Which one do you think will be faster at lower settings? What if all the important settings already are set to low in the TPU test?

The 16GB card is definitely faster than the 8GB, even with lower settings, but the hypothetical 12GB card would be slower than the 8GB card.

Spoiler

For reference, these are the original specs that we are arguing over.

9060(XT) 8GB -> 3.2GHz, 32CU, 32MB, 128-bit 20Gbit
9050XT 12GB -> 2.75GHz, 28CU, 24MB, 96-bit 20Gbit

Edit: Also, I'm pretty sure TPU tests all their games at max settings. Someone correct me if I'm wrong.

 

[Timorous]

The 16GB card is definitely faster than the 8GB, even with lower settings, but the hypothetical 12GB card would be slower than the 8GB card.

The entire argument is about balance. A 12GB 28CU part will be far better balanced at $280 than an 8GB 9060XT will be and the 8GB 9060XT won't be cheaper than that.

As for the quoted specs, those are not specs I came up with as I didn't specify clockspeeds.

You admit that people will turn down settings. Why is it "contrived" when I say it? I don't think you're arguing in good faith. Regardless, this conversation is going nowhere, so I'm probably done responding.

The change in settings to hit 60fps when you are at 47fps average at 1080p max + RT than when you are at 17fps is drastically different. The latter probably requires you to turn off RT entirely where as the former gives you an option to turn down a few particle effects to hit the desired frame rate and keep RT turned on if that is the preference of the user.

The 4K example also shows the 7600XT hitting pretty much 60fps and the 7600 down at 24fps. I know people will go 'the x600 is not a 4K part' which is fine but plenty of people will build a PS5 pro priced PC and plug it into their TV and to hit that kind of price point you are looking at x600 / x60 tier parts.

Given the fact you just want to sit there and discount scenarios for 'reasons' to accuse me of not arguing in good faith is very very pot kettle. I provided 6 direct examples of 7600 vs 7600XT and a link to an article with 15 examples of 5060Ti 8GB vs 16GB which is the best allegory we would have for a 9060XT 16GB vs 8GB matchup if AMD do cancel the 8GB part. I also pointed out how average frame rate is not the only indicator because often 1% lows are the part that is hit making for a smooth vs choppy experience or that IQ takes a hit because the game will suffer worse popin, dynamically reduce texture quality, have texture swapping issues or just outright crash. Rather than accept that evidence you are arguing at the margins over 1 example, the epitome of bad faith arguing.

I won't be doing any more legwork. The water is there, I can't make the horse drink it.

 

[marees]

Based on experience, AMD is unlikely to stop supplying or cancel a product before it is launched. After all, AIB partners have already stocked and produced it. Therefore, in the early stages of sales, both Radeon RX 9060 XT 16GB and Radeon RX 9060 XT 8GB will be available on the market at the same time. As for the follow-up, AIB partners will of course make adjustments based on market sales conditions. Perhaps the 16GB version of Radeon RX 9060 XT will be more likely to appear than the 8GB version.

https://benchlife-info.translate.go...y-1/?_x_tr_sl=auto&_x_tr_tl=en&_x_tr_hl=en-GB

 

[Timorous]

Based on experience, AMD is unlikely to stop supplying or cancel a product before it is launched. After all, AIB partners have already stocked and produced it. Therefore, in the early stages of sales, both Radeon RX 9060 XT 16GB and Radeon RX 9060 XT 8GB will be available on the market at the same time. As for the follow-up, AIB partners will of course make adjustments based on market sales conditions. Perhaps the 16GB version of Radeon RX 9060 XT will be more likely to appear than the 8GB version.

https://benchlife-info.translate.go...y-1/?_x_tr_sl=auto&_x_tr_tl=en&_x_tr_hl=en-GB

It is a wishful hope that they see sense.

Maybe at this late stage the best they can do is offer it at $250 or below due to the massive short comings of 8GB. Then it may be a decent pickup for specific cases. The problem with that is the 16GB card will then look super expensive in comparison because that thing will perform properly.

I don't see a way AMD get through with an 8GB variant that does not tarnish the 16GB one in some way, if the product is priced appropriately for its shortcomings the 16GB card will look like a money grab. If the 8GB card is priced according to the BOM cost difference the 8GB version will look terrible.

If I was AMD and it was too late to cancel the 8GB version then I would stop producing 8GB cards so there are no more and I would heavily encourage them to only be sold in bundle deals with other low end AM5 components. That way for those looking at a standalone product they only see a 16GB version at around $350 or something.