Around a 4080S in performance.
Not bad.

Why is the RAM amount 16304 MB and not 16384 MB ? Little edit went wrong ?

Battler624Around a 4080S in performance.
Not bad.

I hope so too BUT...taking into account core count and ignoring memory bandwidth, it needs ~30% performance bump per core to match 7900 XT, not 4080(S). It's indeed a better node than 7000 series, not like the ADA refresh we got from Nvidia, but still a very optimistic wish. We 'll know soon.

Hard to be excited about this as a 7000 series owner. Maybe Intel could fill the void.

RavenasHard to be excited about this as a 7000 series owner. Maybe Intel could fill the void.

Unless you're using a card at least two tiers below from one generation earlier (either Ada or RDNA3), or same tier from two generations (Ampere/RDNA2), there should be little to no need in upgrading now.

I would like to see someone use a 7900XTX with the same settings and different FSR Quality settings and see the results.

dir_dI would like to see someone use a 7900XTX with the same settings and different FSR Quality settings and see the results.

You'd need it to be someone using it paired with a 285K as well.

wNotyarDYou'd need it to be someone using it paired with a 285K as well.

There's a couple people around here with them. The 7900XTX with a 285K though, i don't know how many people around here have that combination.

dir_dThere's a couple people around here with them. The 7900XTX with a 285K though, i don't know how many people around here have that combination.

I can at least test my XTX with the "1080p high profile" later today and give you the numbers. Any profile on the benchmark activates the vendor upscaler (DLSS for NV, FSR for AMD and XeSS for Intel), but not the FG, that must be set by hand.
The 5700X3D will give numbers a bit lower than a 285K would, but at least there'll be an idea.

This is supposed to be 1080p, ultra (FSR quality) + frame generator? Still, that's pretty high considering how CPU-limited this game is.

Is it some kind of insider joke/meme to leak benchmarks on something non-standard and/or only marginally meaningful?

Older leaks have indicated that the first wave of RDNA 4 cards will make do with PCI-Express 4.0 x16 interfaces, but the GPU-Z screenshot shows a PCI-Express 5.0 x16 bus interface (detection could be bugged).

or it might actually be PCIe 5.0 because this is a production board and not some soldered together engineering sample some rando leaker saw running on x570 board.

we still need the price, until then...

oh and independent benchmark results, don't fall for manufacturer lies

dir_dI would like to see someone use a 7900XTX with the same settings and different FSR Quality settings and see the results.

Would this even tell us anything? 1080p FSR sounds like a massive CPU bottleneck.

Won't tell us much but we could probably guess the real performance of the card compare to AMD's best offering right now. Plus the leaked Canadian price we can guess within 15% where it's going to land and if it's actually worth buying.

All speculation of course off one game. Just fun trying to if you can figure it out before the launch i guess.

The card was allegedly tested in 1080p with the high preset, where the benchmark uses upscaling on balanced. Frame generation was enabled. This is my result with a stock MBA 7900xtx:



Rendering resolution is 626p. With TPU's game suite the 285K from the leak was 6% faster than the 5800X3D in 720p. With that assumption, a stock 7900xtx would be some 13% faster than a supposed top 9070xt here.

AMD basically told us if you have a 7900 series GPU, save your money for the next card.

Hmm 64 ROPs with 256 TMUs? So I'm assuming the ROPs are wider as thats 4 TMUs per ROP.

7900 XTX is 192 ROPs with 384 TMUs, so thats 2 TMUs per ROP.

A theoretical 192 ROP 9090 XTX (or whatever it may be called) would be a massive 768 TMUs, which should be able to destroy the RTX 5090 in pure rasterization horsepower but would require more energy to do so.

@dir_d got some numbers. Adrenalin 24.9.1.
1080p High preset implies FSR Balanced. FG on: 37984 points with 223.00 fps average.
1080p "High" with FSR Quality, FG on: 37335 pts with 219.06 fps avg.
1080p "High" with FSR Native AA, FG on: 33981 pts with 199.10 fps avg.
1080p "High" with no FSR and FG: 38800 pts with 113.51 fps avg.

wNotyarD@dir_d got some numbers. Adrenalin 24.9.1.
1080p High preset implies FSR Balanced. FG on: 37984 points with 223.00 fps average.
1080p "High" with FSR Quality, FG on: 37335 pts with 219.06 fps avg.
1080p "High" with FSR Native AA, FG on: 33981 pts with 199.10 fps avg.
1080p "High" with no FSR and FG: 38800 pts with 113.51 fps avg.

That's tough to swallow. This card is going to above the 7900XT and maybe match the 7900XTX at max. It REALLy needs to be priced well. We are talking $550, if not, what's the point of releasing the same card a couple years later. The prices for the XT and XTX were not that bad before all the stock dried up. I had high hopes but now i am not very confident in this launch specially after seeing the Canadian pricing leak.

These specs have already been confirmed for the 9070XT

Real news would be confirmation of the vanilla 9070. Most leaks and rumours suggest a 3584 shader config, but a news article here last week suggested that both the XT and vanilla were getting 4096 shaders which was rather controversial.

Hi, just registered an account real quick to correct this, the screenshot actually said very high (極高) preset not high preset.

CheeseballHmm 64 ROPs with 256 TMUs? So I'm assuming the ROPs are wider as thats 4 TMUs per ROP.

7900 XTX is 192 ROPs with 384 TMUs, so thats 2 TMUs per ROP.

A theoretical 192 ROP 9090 XTX (or whatever it may be called) would be a massive 768 TMUs, which should be able to destroy the RTX 5090 in pure rasterization horsepower but would require more energy to do so.

Ok, so...this is because they're bandwidth-limited and shader ops limited to around ~1878sp per 16 ROPs in raster. This is just how rasterization works (on all archs).
This is why most clusters are 1792-2048sp; nVIDIA's are currently 1536sp bc reasons (buff up certain other ops to look good, but I would argue we don't need them at such high ratio).
IMO 1920sp is optimal, but it's asynchronous (15x128 instead of the more symmetrical 14 or 16). 7800xt was this configuration. N48 is likely 2048 (for the extra TMUs/rt ratio, most-likely).
Given it can clock high, and bandwidth limited, no reason to have the extra ROPs.
This was made apparent by the stock clock. As I've said, 8192 (if not rop limited) would be limited to ~2720mhz @ 20gbps.
If limited by 64 ROPs, or around low ~75xxsp, you get a clock not unlike the one of 9070xt (2970mhz). Don't worry about it that much (ROP limitation), as extra compute can be used for things outside raster.

TMUs can be used for many things...in this case it is clearly for BVH/RT. One does not need more than 1RT unit per core, but it can be buffed up through the use of a TMU. I think AMD said ~30% iirc.
If I'm not mistaken, this is already used on the XSX (4 ops/CU). FWIW, nvidia does this (what the extra TMUs will do) with fixed-function hw (which is probably not unlike a TAU).

Hope I explained that well-enough to that make sense. You can read more about it here, if it helps. I wouldn't expect miracles on this chip (as it's mid-range), but probably good for 'low' RT like 7900xtx.
Essentially 2/3 chip of N31, but ~30% better RT through extra TMUs with clockspeed differance making up the equalization. Yeah...yeah...Pretty much exactly that...in theory.

As for the 'ideal' for RT, you can clearly see with nVIDIA it is 8 clusters (where 4080/5080 are 7), and they are very-purposely withholding this (probably until Rubin). AMD likely until the 7900xtx successor.
Think 'high-RT' 1080p->4k upscaled, or 'normal (in some new games)/low (in games with a low/high mode)' RT 1440p->4k upscaled.
This is why you shouldn't invest in RT rn, unless you want to upscale games that require it and/or have a 'low' mode from ~1080p...and even that might be iffy moving foward (bc of 12/16GB ram limitations).
With AMD we won't know until we (meaning probably I) throughly scrutinize the actual real-world capability of the TMUs on N48 wrt RT. Probably safe to assume a ~7900xtx with 2x TMU would also be good.
I am very curious if we can get really good RT on (3nm) 192-bit, or rather it will require 256-bit. For nVIDIA it should be possible w/ high clocks (6 clusters), but they may again inch it along to force an upsell.
All the more reason to wait/buy the 3nm 192-bit/256-bit chips, where I would assume if AMD doesn't match nVIDIA per clock (and units), they will likely make up for it (if not exceed it) w/ clockspeed.

Ultron1337Why is the RAM amount 16304 MB and not 16384 MB ? Little edit went wrong ?

More likely that the driver is misreporting, or there is some sort of massaging of the data it returns that GPU-Z isn't yet aware needs to be done.

docnorthI hope so too BUT...taking into account core count and ignoring memory bandwidth, it needs ~30% performance bump per core to match 7900 XT, not 4080(S). It's indeed a better node than 7000 series, not like the ADA refresh we got from Nvidia, but still a very optimistic wish. We 'll know soon.

The clock speed is about 33% higher than a 7800XT with 4 more CUs and slightly higher memory bandwidth. There’s your 30% and then some.