*** The AMD RDNA 4 Rumour Mill ***

[humbug]

i reckon amd too will be eventually embracing a superscalar architecture like nvidia, all basic functional units bunched together, its been a great success for nvidia, not only have they utterly destroyed amd in raw performance but are currently 1.5 generations ahead of them in efficiency

Supescaler is simply executing multiple instructions per clock cycle, or what we now call and measure as IPC, this was cracked in the 1970's along with the move away from RISC to programable logic.
For decades everything has been Superscaler.

Also, the 7900 XTX despite being on 5nm vs the 4090 4nm is 85% the performance at 85% the power consumption.

 

[humbug]

amd has vector units, nvidia has scalar bunched up functional units called cuda/tensor/rt cores etc... amd has to package instructions using a more complicated scheduler to maintain high utilisation of their vector units, nvidia just has to segregate instructions by the type of operation and then push it in the pipeline..

a very simplistic way to describe both architectures could be in terms of lets say visa application:
amd: runs a single window to process a single visa application. to process an application documents A, B, C will be processed by a single window per applicant
nvidia: maintains separate dedicated windows for processing documents A, B, C.. everybody, submits their documents in a common queue, the documents are then separated and processed by dedicated windows which specialise in processing a specific document




perf numbers below for timespy extreme
rtx 4090: 19261
7900 xtx: 14688
and rtx 4090 doesnt touch max rated TDPs

https://www.kitguru.net/wp-content/uploads/2022/12/3D-TSE-3.png

and the performance gulf will be even greater for RT apps

None of this has anything to do with Superscaling.

8FP.

H100: 3.96 TeraFlops
MI 300X: 5.23 TeraFlops

You're cherry picking a synthetic benchmark.

Multiple Gamers overall performance 4K.

RX 7900 XTX: 100%
RTX 4090 : 122%

Power consumption in gaming.

7900 XTX: 356 watts
RTX 4090: 411 watts (116%)

They are near enough the same performance per watt, again despite the 4090 being on a better lithography node, TSMC 4nm vs TSNC 5nm.

AMD Radeon RX 7900 XTX Review - Disrupting the GeForce RTX 4080

Navi 31 is here! The new $999 Radeon RX 7900 XTX in this review is AMD's new flagship card based on the wonderful chiplet technology that made the Ryzen Effect possible. In our testing we can confirm that the new RX 7900 XTX is indeed faster than the GeForce RTX 4080, but only with RT disabled.

www.techpowerup.com

 

[humbug]

short of reading all this is there any actual solid specs yet?

No.... wont be for a while yet.

 

[humbug]

Yeah but listen... pineapple belongs on pizza.

Absolutely not!

 

[humbug]

Mid tier for the next generation, the thinking is:

7900 XT > 8800 XT
7800 XT > 8700 XT

 

[humbug]

Pretty underwhelming really! Wasn’t the 6700xt 25% faster than the 5700xt. This means the 7900xtx should be 8800xt and and 7900 GRE should be 8700xt BUT these companies are loving it right now and will do as they please for profit.

Depends on price.

At 1440P the 7900 XT is 28% faster than the 7800 XT and 17% faster than the 7900 GRE

If it lands at £500 with 2X the RT throughput i think it will be a hit.

Sapphire Radeon RX 7900 GRE Pulse Review

The Sapphire Radeon RX 7900 GRE Pulse emerges as the best GRE model released today, offering exceptional value at AMD's MSRP of $550. Outperforming the RX 4070 Super in raster, it operates whisper-quiet, catering to gamers and enthusiasts seeking top-tier performance without exceeding their budget.

www.techpowerup.com

 

[humbug]

Tbf there's nothing 'wrong' with AMD's RT tech as long as it's programmed for agnostically. RTX is what kills it, purely optimised for Nvidia and a lot of black box **** going on that hinders AMD more than helps (this is Nvidia we're talking about remember, done some scummy tricks in the past and they haven't really changed from this!).

If rumours hold true its going to be quite interesting.

Sony Said "Ray Tracing Performance 2X to 4X faster" its where the 3X comes from, but some of that will also be down to the core count difference, so drop it down to 2X.

Ok so RT performance is like bandwidth, the more of it you have the less FPS you will lose by running RT.

Cyberpunk.
7900 XTX 125 FPS
4080 119 FPS

Cyberpunk with RT.
7900 XTX 40 FPS (-68%)
4080 59 FPS (-50%)

The 4080 has 36% more bandwidth.

If we take a doubling of bandwidth as read then an RDNA 4 GPU would only lose 34%.

Cyberpunk with RT.
RDNA 4 82 FPS (-34%)
4080 59 FPS (-50%)

RDNA 4 has 47% more bandwidth

I almost didn't want to write this as it seems too incredible, i have no idea how this would work, it just makes sense to me, it would be fun if this is how it works out, but, well lets see....

Sapphire Radeon RX 7900 GRE Pulse Review

The Sapphire Radeon RX 7900 GRE Pulse emerges as the best GRE model released today, offering exceptional value at AMD's MSRP of $550. Outperforming the RX 4070 Super in raster, it operates whisper-quiet, catering to gamers and enthusiasts seeking top-tier performance without exceeding their budget.

www.techpowerup.com

 

[humbug]

I don't know where this reasoning that Nvidia has "dedicated RT cores" and AMD doesn't, it is not correct, it is not how this works, they both have "dedicated RT hardware" there is nothing emulated about AMD's RT, it is physical.

The difference is in how Nvidia and AMD build BVH, this is really over simplified because i'm not going to write a wall of text to explain this, AMD Construct BVH over many branches, Nvidia do it over a very wide tree, this is a bit like 8 slow cores vs 4 fast cores, both can be equally as fast but not by the same method.

The advantage of the wide approach is it doesn't really matter which BVH construction you code for you will always get the most out of being wide, the disadvantage wide requires more caching, its why Ada has so much L2 cache, the advantage of the branch approach is you don't need so much cache, but unless you're specifically going to code for that its going to be slower.

Now i's sure AMD's thinking was keep the die size down, it doesn't matter as we own consoles they are going to code for us, hmm... well they don't have to and if the studio is packed with Nvidia cards they aren't going to.

Also, and game that AMD does RT well in must be fake RT, no, not necessarily.

 

[humbug]

^^^^

I don't want to drag this off topic given the complaint already.... AMD's shaders double up as RT cores, Nvidia have specific, or "dedicated" RT cores, so in terms of wording yes it is accurate to say that but in terms of functionality AMD's is still hardware with specific hardware extension functionality, IE "not emulated"

That's the last i will say on it here

 

[humbug]

Ah, the AMD defence force arrives.

I was keeping my previous post very simple and even clarified that AMD has some dedicated hardware. What they leave out is hardware dedicated to BVH traversal.

And I don't need you to write a wall of text explaining it. Because I can sum up your wall of text right now. It will be load of technical sounding gibberish basically giving a bunch of excuses for why AMD doesn't perform as well as Nvidia in Ray Tracing.

I'm sorry but this isn't going to descend into a silly argument. Because if you actually know what you say know, you know that there is no getting around the hardware limitation that AMD's solution has compared to Nvidia's. Nvidia's will always be faster. Hence the reason why it's changing for RDNA 4.

And I never said Fake RT either. Where did you even get that from?

Read first line, ignores rest... goes back to reading something interesting.

 

[humbug]

AMD’s new GPU can’t even outperform Radeon RX 7900 XTX, says leak

The best AMD RDNA 4 graphics card will reportedly be better at ray tracing than RDNA 3, but will still be slower than the current flagship.

www.pcgamesn.com

Navi 48 RX 8800 will not be faster than Navi 31 RX 7900 XTX, it will be between RX 7900 XT and XTX performance.

What a weird click bait headline, its a choice they made and we already know about it, they are about 6 months late...

The only thing that matters is price, in the grand scheme of things no one cares about GPU's that cost more than £500.

 

[humbug]

They have a professional class cards that are not gaming cards and have different drivers.

This, CDNA is a different architecture than RDNA.

 

[humbug]

if we looking at 7900xt like performance, would have to be lower end of the price ladder, have seen them around the £600 mark, so realistically 8800xt for £600 isnt such a great deal. would have to be £500 to get my wallet open and i suspect many others

Considering the RX 7800 XT launched $500 it had better not be $600, it needs to be $500, at that its good, just like the RX 7800 XT was, Is.

 

[humbug]

The RX 7900 XT is 90% the performance of a 4080, if its between them its 95%, one would call that a 4080.

At $500 that's a smash hit, AMD, are you listing AMD? $500.

 

[humbug]

4080 performance my arse!! I'd bet on it being just a little bit faster than a GRE in a 'real review', AMD slides are just a joke unto themselves at this point!

Vs 7800 XT 7% more CU's, 2.9Ghz is 21% higher than the 2.4Ghz of the RX 7800 XT, its 28% more GPU, its possible.

Edit: at 2.9Ghz the 7800 XT is pretty much in line with the 4070 Ti.

 

[humbug]

They really benefit from tuning.

 

[humbug]

This is another way of arriving at the same take I have. I just looked at previous gen updates at similar GPU tiers (not prices) for AMD and that has been ~25% - 30% each gen. That puts the MLID “leak” as laughable because it’s hardly “news”.

The level of typical performance uplift over the 7800 XT equals between 7900 XT and RTX 4080 raster performance.

If they don’t release at the original 7800 XT prices then AMD deserve all the crap they will get.

The problem with RDNA 3 is a memory bandwidth issue, with about 650 GB/s of bandwidth they don't scale beyond a 4070 Ti, mine scales almost linier with clocks all the way up to about 2.9Ghz, if i add what little i can to the memory clocks, about 100Mhz i can match a 4070Ti easily, but that's it, i can get it running at 3.1Ghz but the increase in FPS is literally 0, from almost perfect scaling to nothing.

The 7900 GRE has a lot more CU's, 80 vs my 60, but it has the same memory architecture there fore its little better, about 13%, the 7900 XT only has 5% more CU's, 84 and yet its WAY... faster, it has a 320Bit Bus, with that it has 800 GB/s of Bandwidth.

So that's the problem with them, if AMD have solves that... even just slightly.

 

[humbug]

My own experience with my 7900XT is that almost 15% extra performance is possible with an undervolt and OC. That put it on par with my RTX 4080 which only gave a 5% uplift after an overclock.

Yeah, you have some more memory bandwidth

 

[humbug]

I'm running my GPU at touching distance of a 4070 Ti, its effectively a 16GB 4070 Ti in Raster and even for the same power consumption, £480.

 

[humbug]

Well what does that tell you about the rumoured performance and price then? It is a joke!

Not that I would even take a 7900xtx over my 4070 Ti. But still.

Oh and I play most games 120-140w with my tuned profile.

Is the 4070 Ti, £800, still, a joke? 4080 level performance at half the money, you go in to the Intel GPU thread what you will find there is people hyping up Battlemage to exactly that level and going nutty at the idea of it, desperately wishing it in to being.

Only when its branded AMD can this be bad. that's the joke, and its a crap one, its not funny.