AMD RX 7900XTX and RX 7900XT Reviews

[Scott_Arm]

@T2098 Yah, it's weird. They lost their efficiency advantage.

 

[Bondrewd]

So the ray tracing performance gain over 6900XT is higher than the rasterisation gain?

Yea, it's the least broken bit of it all.

 

[troyan]

And why would it be? Raytracing means more workload, which helps the compiler to find more usable instructions for the ILP.

 

[Bondrewd]

Raytracing means more workload, which helps the compiler to find more usable instructions for the ILP.

Wrong, the thing grinds to a halt in pixel pushing synthetics a-la TSE which does not happen when it throws rays around.

 

[homerdog]

I was ready for AMD to demolish the $1200 4080. What a terrible time to deliver broken silicon.

 

[Scott_Arm]

Really curious how COD Modern Warfare 2 and Warzone 2 play so nice with AMD. Has to be something about their renderer that really plays to AMDs strengths. Visibility buffer? No idea.

 

[del42sa]

https://wccftech.com/amd-radeon-rx-...3-7-ghz-front-end-clock-3-5-ghz-shader-clock/

Twitter fellow, 0x22h, has been testing around with an AMD Radeon RX 7900 XT graphics card and it looks like he has managed to obtain the fastest overclocking speeds with the Navi 31 chip. The RX 7900 XT was overclocked to a maximum 3.5 GHz shader clock and a 3.7 GHz Front End Clock. That's a 1.2 GHz increase in the Front End and Shader Clocks which is very impressive. The card peaked at a maximum board power of 400W, 45W higher than its 355W TGP and the temperatures stabilized around 62C and 85C for the Hot Spot temps. With this impressive overclock, the GPU delivered FP 32 compute performance rated at up to 75 TFLOPs, a 44% increase over the stock clocked graphics card.

 

[Krteq]

"Fixed" HW already in channels?

 

[homerdog]

https://wccftech.com/amd-radeon-rx-...3-7-ghz-front-end-clock-3-5-ghz-shader-clock/

Twitter fellow, 0x22h, has been testing around with an AMD Radeon RX 7900 XT graphics card and it looks like he has managed to obtain the fastest overclocking speeds with the Navi 31 chip. The RX 7900 XT was overclocked to a maximum 3.5 GHz shader clock and a 3.7 GHz Front End Clock. That's a 1.2 GHz increase in the Front End and Shader Clocks which is very impressive. The card peaked at a maximum board power of 400W, 45W higher than its 355W TGP and the temperatures stabilized around 62C and 85C for the Hot Spot temps. With this impressive overclock, the GPU delivered FP 32 compute performance rated at up to 75 TFLOPs, a 44% increase over the stock clocked graphics card.

The comments section of that article is amazing.

 

[Bondrewd]

"Fixed" HW already in channels?

no it won't ever run those clks in video games.

 

[homerdog]

"Fixed" HW already in channels?

If we ever get a truly fixed N31, it will be a separate SKU and it will be a while.

 

[Granath]

"Fixed" HW already in channels?

LN2.
Retape takes much longer.

 

[Bondrewd]

it will be a separate SKU and it will be a while.

Yea and it's not *that* far off.

LN2.

Nope, it's on air.
Just in light OCL load.

 

[trinibwoy]

I was expecting worse with RT tbh. They have now around 3090 performances (I just watched Gamer Nexus review), which is still a monster. So, while they're behind 4080, it seems at least that RT is exploitable, which was not always the case with rdna2.

Well done.

Yeah that seems to be the most positive aspect of this launch. RT performance has gotten a nice boost. Once again though AMD hasn’t offered something very compelling to non-diehard fans.

 

[DegustatoR]

So from whatever benchmarks I've seen it's about +40% perf/watt over RDNA2 at best (+15% at worst).

 

[pjbliverpool]

We really need someone to benchmark it against Ada and Ampere in Fortnight Season 4. If it performs better with software Lumen relative to it's hardware RT capabilities then that would be a big win against NV. It's also be interesting to see how it fares in Lumen HWRT.

 

[troyan]

Kitguru gets between 30% and 35% improvement over the 6900XT and 6950XT: https://www.kitguru.net/components/graphic-cards/dominic-moass/amd-rx-7900-xtx-review/32/

 

[Jawed]

Theory: dual-issue to both parts of a VALU lane uses huge amounts of power.

Games that are heavy on pixel shading, which uses the dual-issue capability due to wave64 mode, scale worse on RNDA 3 than games that are compute heavy, where shaders are mostly running as wave32. With wave32 there would generally be less use of the co-issue capability (simply because it's hard to co-issue from a single thread), so there's less overall power usage.

My understanding is that ray tracing is generally executed in wave32 mode.

Pretty weak theory. Sounds plausible. Could be tested, I suppose. Does anyone care?

 

[Bondrewd]

Does anyone care?

Not until N32/N33 are out.

 

[troyan]

Theory: dual-issue to both parts of a VALU lane uses huge amounts of power.

Games that are heavy on pixel shading, which uses the dual-issue capability due to wave64 mode, scale worse on RNDA 3 than games that are compute heavy, where shaders are mostly running as wave32. With wave32 there would generally be less use of the co-issue capability (simply because it's hard to co-issue from a single thread), so there's less overall power usage.

My understanding is that ray tracing is generally executed in wave32 mode.

Pretty weak theory. Sounds plausible. Could be tested, I suppose. Does anyone care?

French site comptoir-hardware uses a OpenGL test to measure pixel shader performance: https://www.comptoir-hardware.com/a...-radeon-rx-7900-xt-a-rx-7900-xtx.html?start=5

7900XTX is 20% faster than the 6950XT - so the performance increase comes only from 20% more CUs.