Radeon Thread

[Keyrock]

2 minutes ago, Sarex said:

Then again how much do you care for ray tracing.

Nada. Zip. Zilch. Zero.

[Zoraptor]

That's some pretty aggressive pricing. Won't be faster than a 4090 (hence the lack of benchmarks, one strongly suspects) as realistically you'd have needed the previously rumoured (but debunked a while ago) extra chip on the MCM for that, but should be a fair bit faster than a 4080 and at a significant discount.

[Sarex]

How are the Radeon drivers nowadays?

[Zoraptor]

Fine. Every problem I've personally had with their drivers has been external, ie Microsoft update installing old drivers over the top of new ones despite being told not to. And updates deleting the workgroup policy that hard enforces that. They've even fixed DX9/ OpenGl performance recently, late, but done.

5700XT was the last problem release, and that was a new arch.

[Keyrock]

First very brief look at high-end Sapphire RDNA3 cards:

Vapor chamber, three 8-pin connectors, and a 20 phase VRM will likely mean substantially higher clocks. The rumor is that RDNA3 is capable of hitting 3000 MHz clocks on air cooling. If that's true then AMD left quite a bit of overclocking headroom for partner cards. Of course, said partner cards will undoubtedly also feature a significantly higher price tag.

 

Edited December 1, 2022 by Keyrock

[Keyrock]

tl;dw Can't hang with the 4090 but it beats the more expensive 4080 more often than not in rasterization. Somewhat competitive with mild ray tracing but falls behind with heavy ray tracing.

[Bartimaeus]

Forget Singin' in the Rain, we're onto benchmarking in the rain. Bizarrely no musical numbers, though - somewhat disappointing for long term fans of the franchise.

(e): While I wouldn't ever buy something as ridiculous as the 7900 XTX, I am curious to see how the rest of the series shakes out. Perhaps it's time to finally upgrade...

Edited December 12, 2022 by Bartimaeus

[Keyrock]

I can't possibly consider a $1000 GPU as a good value, but at least it's a slightly less terrible value than the RTX 4080.

As far as upgrading goes, it comes down to what is your goal and does what you currently have achieve that goal satisfactorily.

Edited December 12, 2022 by Keyrock

[Zoraptor]

Kind of a weird release which makes me wonder if they've had something Go Wrong somewhere; at AMD/ radeon specifically, not the reviewers. From the hardware it ought to be close to half way between 4080 and 4090 (raster) performance wise, and it's a fair bit behind that. It is still better value than a 4080, but it is also a bit disappointing it isn't better value/ performance.

Will be interesting to see if they can claw back that 'missing' performance with drivers or AIB cards, though in terms of marketing it would largely be irrelevant. Also interesting to see if the speculated 7970/90 eventuate.

[Keyrock]

What could they do with a 7950 XTX besides raise the clocks? AIB cards are already going to bring higher clocks. This is already the full 96 CUs of Navi31, they can't add any more CUs. I suppose HBM instead of GDDR, but that would involve a significant price hike.

Still, the 7900 XTX, and for the most part the 7900 XT can already handle 4K comfortably. Changing the second number from the right isn't going to get it anywhere near playable 8K, not to mention that 8K is currently a pipe dream anyway. So what's even the point? Games aren't going to push far beyond where we are currently for several years because the lowest common denominator are the MSony consoles. We're looking at at least 4 more years before a PS6 or xXxBxOxXxxxxxxxxxxxxxxxXx.

Edit: I'll answer my own question. The point is that some dumbasses will buy a 7950 XTX that's 4% faster for $200 more.

Edited December 13, 2022 by Keyrock

[Zoraptor]

There's a decent amount they could do, in theory.

Even the 96CU (will WGPs ever stick?) could be got around*. OTOH, there's plenty of scope for potential memory improvement without HBM- go from 6x16MB to 6x32MB, and/or bring in vCache. The latter surely has to happen for GPUs at some point, and you could draw a parallel with the vanilla cache 7000 CPUs being released first and 3d cache versions later.

I don't think we've got anything even approaching semi official on cards above the 7900XTX even existing yet though, so it's all doubly in theory.

*7990 was a dual GPU card last time and that approach fundamentally makes more sense if the design is 'proper' MCM. Likely... no, especially for Navi31. Outside, outside, possibility of 2x64 from Navi 32 though.

Edited December 13, 2022 by Zoraptor

[Keyrock]

I agree on V-Cache, but I wouldn't do it on RDNA3, that's asking for trouble. They just introduced chiplets to GPUs with RDNA3 and there will surely be kinks to work out. I think it's a bad idea to add more than 1 new technology in a given generation. I know both of these have been done before on CPUs, but GPUs bring different challenges. In my opinion the smarter move would be to save V-Cache for RDNA4.

[Humanoid]

An AI could probably write the script of how this generation's launch went quite easily. See you all again for more of the same in 2024?

Actually that might not be strictly true. nVidia would have gotten a big one-off boost this generation by switching from Samsung back to TSMC, so hindsight tells us that such a leap might be difficult to replicate. At least, that's what AMD might be hoping for.

 

As for me? Well, when does The Witcher 4 launch?

(I can easily see myself still using this 6800 non-XT through the next generation too, so maybe 2026? The only thing I play that challenges it is Cyberpunk, though it also "only" does 100-110fps in Euro Truck Sim 2. At 7680x1440, that is.)

[Azdeus]

I'm a bit surprised that I haven't heard anywhere near asmuch outrage for the frankly quite anti-consumer move of having two such widely performing cards be separated by one X. That's even worse than the nvidia kerfuffle.

[Humanoid]

44 minutes ago, Azdeus said:

That's even worse than the nvidia kerfuffle.

Is it though? It's certainly not good, but an X doesn't carry the same implication as specifying a certain amount of VRAM. An X could mean anything, same as a 'Ti', and at least invites some investigation. The nVidia thing is more like discovering that a 600mL bottle of Coke not only has less Coke in it than a 1.25L bottle as one would expect, but it also tastes worse.

[Keyrock]

Nomenclature issues aside, the RX 7900 XT is a terribly priced GPU, much like the RTX 4080, maybe even more so. I mean, the prices on all of these cards are horrific, but the 7900 XT is particularly egregiously priced when compared to the XTX. I wouldn't be at all surprised to see a price cut on it fairly quickly, much like what happened with the Ryzen 7xxx CPUs. The RX 7900 XTX, on the other hand, will probably stay at the current price for quite a while, as it's a "good value" (I'm speaking in relative terms here), will likely remain in short supply for at least a couple months, and is already being scalped by scumbags.

[Zoraptor]

Trying to scalp a 7900XTX --> nice idiot detector since there isn't as much demand and a lot more supply.

I don't think there's anything wrong with the naming of the 7900XT, but its pricing is certainly bad since price/ performance is nearly linear to the xtx. Maybe should have been the 7900 and 7900XT, but that's still just a couple of letters different. OTOH if it's anything like the 6800/XT vs 6900 it will end up almost discontinued once the manufacturing kinks have been worked out.

(and unfortunately I'm now in the market for a card. Not, one suspects, a 7900XT/X though since 800NZD is about the most I'm willing to pay)

[Azdeus]

4 hours ago, Humanoid said:

Is it though? It's certainly not good, but an X doesn't carry the same implication as specifying a certain amount of VRAM. An X could mean anything, same as a 'Ti', and at least invites some investigation. The nVidia thing is more like discovering that a 600mL bottle of Coke not only has less Coke in it than a 1.25L bottle as one would expect, but it also tastes worse.

Atleast nvidia clearly stated that you got less RAM, the fact that it was an entirely different chip in it is awful, but AMD's naming implies that it at worst would be a non overclocked version or similar to how their other products such as CPUs are handled, but it has 17% less memory, cache, bandwidth and 10% less compute units.

[Zoraptor]

Yeah, going to change my mind after checking out the aggregate benchmark results. Big enough difference in performance that it needs a different model number. Also even worse value, somehow.

[Keyrock]

The 7900 XT should have been $799.

[ComradeYellow]

RX 7900 XTX looks promising but seriously, if you have 1000 USD to blow on a card alone it makes literally zero sense not to just find another few hundred and go with the 4090, which could probably last you almost a decade and has boatloads more features.  AMD goofed here and people aren't gonna suddenly change their ways because AMD outflanked the ridiculously over priced 4080 with the also ridiculously over priced 7900 XTX.  I really think they're overestimating themselves in the GPU arena, Jensen is simply too powerful and universally well respected at this point to attempt to dislodge in extreme high end.

 

Something tells me AMD WANTED to make their latest cards cheaper but production was simply too costly and ambitious to do so and decided to play it safe rather than risk a profit loss and just force a low price with a hail mary.  Oh well, maybe next time.

[Humanoid]

10 hours ago, Azdeus said:

AMD's naming implies that it at worst would be a non overclocked version

So a 3060 Ti vs 3060 situation then.

To be clear, I absolutely do think they should never have brought back the XTX naming scheme and that the XT should have been the 7800 XT. But I do think it's no more egregious than what the industry has already been doing for years.

2 hours ago, ComradeYellow said:

if you have 1000 USD to blow on a card alone it makes literally zero sense not to just find another few hundred and go with the 4090

I get the sentiment, but I think $600 (or 60%) is more than "another few hundred". On the other hand, I have seen retailers here cut the 4080 price such that it was half the MSRP difference. It was a one-day special so it's gone now, but the equivalent of $1100 for the 4080 vs $1000 for the just-launched XTX did deflate AMD's balloon somewhat there.

[Azdeus]

2 hours ago, Humanoid said:

So a 3060 Ti vs 3060 situation then.

To be clear, I absolutely do think they should never have brought back the XTX naming scheme and that the XT should have been the 7800 XT. But I do think it's no more egregious than what the industry has already been doing for years.

To paraphrase Hardware Unboxed, this is anticonsumer horse****, there are plenty of numbers to pick from, do better AMD. To be fair, they did somewhat the same thing with the 6000 series, but it wasn't quite as egregious from what I rembmer as what they're doing now. IIRC their non-XT versions had cut dies but were otherwise the same, while the 6*50s were overclocked?

[Humanoid]

RDNA 2 was reasonably clean, Navi 21 for the 6800 series and above, Navi 22 for the 6700 series, and Navi 23 for the 6600 series. And yes, the refreshes were simply clock boosts.

For Ampere, it was GA102 for the 3080 and above, fine. GA104 though problematically started from the 3060 Ti through the 3070 series. Which left GA106 to serve as the plain 3060 and 3050.

(None of these count mobile GPUs which are a complete mess that I don't even want to look at, let alone disentangle)

[Keyrock]

@ComradeYellowhas a rather different definition of "a few hundred" than I do, it seems.

Anyway, I'm looking forward to the eventual Zen4/RDNA3 APUs. There's been a whole bunch of handhelds recently released utilizing the 6800U APU with 8 Zen3+ cores and 12 CUs of RDNA2. They're actually quite capable by all accounts. In theory, an eventual 7800U, or whatever they wind up calling it, should be a 1080p champion.