AMD RX 7900XTX and RX 7900XT Reviews

[DavidGraham]

What is causing this?

https://twitter.com/x/status/1602364484907315200

https://twitter.com/x/status/1602375589050159107

 

[Bondrewd]

What is causing this?

I assume WGPs still hog a fuckton of current even at Vmin.

 

[DegustatoR]

For the $200 more the 4080 is still asking in most markets, although it's far from cut and dried, I can see the argument for a 7900XTX.

I don't.
If we would be talking about some products which are completely comparable otherwise and are manufactured by companies each of which holds roughly 50% of market share then yes, this would make sense.
In the situation where AMD GPUs are right now being just a little bit more attractive in price won't help to attract any Nv buyers to their products.
Even the +20% markup on 4080 should look fine from Nv's perspective - and it's not really a flat +20% either, this is mostly a US case I think with the gap being smaller in other regions.
So as unfortunate as it sounds it looks like Nvidia knew exactly where N31 cards will land and their pricing for both 4080 and 4090 reflect that knowledge very well.
Recent rumors of 4070Ti "relaunch" happening at the exact same price as was planned for "4080 12GB" also make sense in this view - w/o RT it should land a bit below than 7900XT - which has the exact same price.

 

[Jawed]

That doesn't sound right to me.
Games which scale the best on RDNA3 seem to be games made for old GCN consoles where Wave64 was the only mode of execution.

GCN has long been characterised as a "compute-centric" architecture, which is why it has lived on in CDNA, which is a compute-centric product.

Also the cards are showing the lowest clocks when doing RT (so Wave32 mostly) which implies that this is when they are power limited the most.

Nice bit of data there! Sometimes the clock when doing RT is higher than when not, though...

XT, in that table, has higher average clocks than XTX in many games.

Fundamentally, clock speed measured versus clock speed maximum is a way to indirectly measure the amount of work per second that the GPU is doing - the larger that delta, the greater the work per second.

Ray tracing in RDNA 3 benefits from the following when compared with RDNA 2:

• 50% more register file space, allowing more hardware threads in flight, hiding more latency
• ray accelerators that can be tuned for the type of ray query they are doing
• almost double the off-die bandwidth (rasterisation theoretical rates are only ~50% improved)

so these are all possible reasons why RDNA 3 ray tracing has improved more than rasterisation. So it's not possible to say that wave32 is the reason why ray tracing has benefitted in this way. At the same time, it can't be eliminated.

The real problem is to characterise the tech in these games, to determine what proportion of frame time is spent running pixel shaders (wave64) versus compute (wave32, though in some games the workgroup size may actually be 128 or 256, in theory).

We can say that any game that uses a forward renderer is definitely high in pixel shading - but that doesn't eliminate the use of compute passes (post-processing). Some post-processing techniques are implemented with pixel shading as far as I can tell (screen-filling quad!) so that's why it's really difficult to characterise games.

The shading pass of deferred rendering is also usually a pixel shader...

Maybe we'll find out when AMD fixes it and tells us how the new GPU solves "X", whatever the hell X turns out to be.

 

[Bondrewd]

which is why it has lived on in CDNA

Not for THAT much longer tho.

 

[sir doris]

So what is the consensus on the poor performance in games that AMD usually excel in, such as Forza Horizon? Is it drivers or a hardware limitation?

The performance of the 7900 XTX seems all over the place; expected gains over the 6950 XT and soundly beating the 4080, through loosing to the 4080 while still seeing decent increases over the 6950 XT, to losing to the 4080 and only displaying relativley small increases over the 6950 XT.

 

[Bondrewd]

Is it drivers or a hardware limitation?

Mostly the latter.

 

[DegustatoR]

GCN has long been characterised as a "compute-centric" architecture, which is why it has lived on in CDNA, which is a compute-centric product.

While true I believe that this is mostly about GCN's execution pipeline being very compute friendly in contrast to its graphics pipeline having all sorts of issues and less about suitability of the architecture to modern (gaming) compute tasks - which tend to be more "single threaded" these days (which is what RDNA has "fixed" in comparison to GCN by going from 64 wide 4 tick execution to a 32 wide 1 tick). Presumably this means that compute generally runs better in wave32 mode than in wave64. I also vaguely remember seeing something about compute shaders being almost exclusively wave32 in RDNA compiler output?

Sometimes the clock when doing RT is higher than when not, though...

These probably hit CPU limitations without RT and the GPU downclocks in the absence of workload.

XT, in that table, has higher average clocks than XTX in many games.

Less power and/or temperature limited maybe?

 

[Bondrewd]

Less power and/or temperature limited maybe?

Lower power limits and smaller cooler too so probably not.

 

[digitalwanderer]

Has anyone seen any UE5 benchmarks? I'm really curious about it and haven't seen anything.

EDITED BITS: Nevermind, found it. Should have looked in the RDNA3 thread first. Sorry!

 

[Frenetic Pony]

So, register bug, power leak, maybe ought else.

Well it's price competitive with the 4080, and hey a fixed respin could see a 50% performance bump. Not what was wanted but, oh well

 

[Scott_Arm]

I wonder if under-volting will help a lot more than overclocking if this thing is basically power-limited all of the time. I'm still shocked that they lost their power efficiency advantage, and by a lot. I actually think the performance profile and the price are pretty good when you look at the overall market.

 

[Bondrewd]

I wonder if under-volting will help a lot more than overclocking if this thing is basically power-limited all of the time.

nnnope, it's draws silly currents instead then.

 

[chris1515]

Basically it looks like better to wait the respin and N32 or a 7950 for people wanting the best AMD offer,

 

[Bondrewd]

Basically it looks like better to wait the respin and N32 or a 7950 for people wanting the best AMD offer,

Yea, don't touch the fucked up one.

 

[Scott_Arm]

nnnope, it's draws silly currents instead then.

What? Normally when you under-volt current doesn't really increase, which is why overall power goes down.

 

[Bondrewd]

Normally when you under-volt current doesn't really increase

The thing isn't normal.

 

[tsa1]

What? Normally when you under-volt current doesn't really increase, which is why overall power goes down.

Depends on how the PL is configured. Typically less voltage you allow chip to use, more room is left for the current consumption to go up (as in P = I * V). That's why simply increasing power limit without increasing current limit does not amount to proper performance increase (Vegas were notorious for that because there was no SPPT at that time so current limit was at fixed 300A or 200A which wasn't that much)

 

[Scott_Arm]

Depends on how the PL is configured. Typically less voltage you allow chip to use, more room is left for the current consumption to go up (as in P = I * V). That's why simply increasing power limit without increasing current limit does not amount to proper performance increase (Vegas were notorious for that because there was no SPPT at that time so current limit was at fixed 300A or 200A which wasn't that much)

I would expect a gpu to have some pretty strict current-limiting for safety reasons, which is why under-volting is actually successful at reducing power. Oh well. I guess we'll see what people can do when they get their hands on the card. Maybe Optimum Tech will check it out. He's the only person I know of that reviews under-volting because he looks at ITX builds a lot.

 

[DegustatoR]

I wonder if under-volting will help a lot more than overclocking if this thing is basically power-limited all of the time.

That's what TPU has reported to be the case at least:

Surprisingly, changing the GPU clock either does nothing for performance, or it results in a loss of performance or crashes the card when set too high. I haven't been able to find any setting that actually increases performance.
I found the best gains for overclocking in general can be found when setting the power limit to maximum (+15%) and combining that with an undervolt that reduces heat output and GPU temperature considerably, which lets the boost algorithm boost higher, for longer.

I'm still shocked that they lost their power efficiency advantage, and by a lot.

They never had any advantage. Ampere was on par with RDNA2 on a considerably worse Samsung's production process. N31 is using a mix of N5 and N6 too while Ada is pure N5.