AMD RDNA3 Specifications Discussion Thread

[DegustatoR]

AMD Radeon RX 7900 XTX/XT reviews go live on December 12th, a day before launch - VideoCardz.com

AMD RX 7900 reviews a day early The RDNA3 review embargo date has been confirmed. Radeon RX 7900 XTX vs. RTX 4080, Source: HardareLuxx According to HardwareLuxx, the results of Radeon 7900 GPUs performance evaluation will be shared on December 12th. This is a day before official launch, which is...

videocardz.com

 

[Deleted member 2197]

First AMD Radeon RX 7900 XTX/7900 XT 3DMark TimeSpy/FireStrike scores are in - VideoCardz.com

AMD Radeon RX 7900 in 3DMark We now have the very first confirmed 3DMark results of AMD’s upcoming Radeon 7900 series. The RX 7900 series cards perform just as expected and akin to what we know from RDNA2 series. This means better 3DMark FireStrike performance, which is a DirectX11 benchmark...

videocardz.com

 

[arandomguy]

Blender Benchmark (HIP) 3.4.0 -

Blender - Open Data

Blender Open Data is a platform to collect, display and query the results of hardware and software performance tests - provided by the public.

opendata.blender.org

7900 XTX	3783.31 (1 test only)	142%
7900 XT	3461.95 (1 test only)	130%
6950 XT	2806.75 (1 test only)	106%
6900 XT	2660.19 (median of 5 tests)	100%

 

[TopSpoiler]

https://twitter.com/x/status/1601213869284012033

https://twitter.com/x/status/1601218354978516992

 

[trinibwoy]

355w is not enough to hit target clocks?

 

[neckthrough]

355w is not enough to hit target clocks?

Certainly possible if there's hw/timing bug(s) affecting the V/f curve. If they are going for a re-tapeout it probably means they've isolated and fixed at least the most egregious ones.

It sucks but given that the industry is going through a demand trough (and Radeon's MSS is low anyway) the impact is probably not as bad as it could have been. Plus it looks like even the buggy chips are decent enough make lemonade, especially given the price headroom Nvidia has graciously given them to play with.

Very interested to see how the fixed V/f curve lines up vs. NVidia's now that they are in the same process node.

 

[Kaotik]

Very interested to see how the fixed V/f curve lines up vs. NVidia's now that they are in the same process node.

While all "5nm" and "4nm" are same node, there are pretty big differences between different variants and sadly we have no details on the process NV uses because they insist on using custom name for it too (all big players get to customize the process to some extent for their needs)

 

[arandomguy]

Just based on the current leaked information this doesn't seem that straight forward.

From the leaked benchmarks there's seemingly some scaling limitations and inconsistency from the 7900XT to 7900XTX. On paper the 7900XTX has roughly 15% more "hardware" and also a 15% higher power limit. If it's just a power issue as traditionally understood due issues with frequency/voltage that doesn't seem like it would explain what is happening here.

We'll need to see how things actually shake up once more information turns up at launch and post, but I'd wonder at this point if there is tradeoff/limitation/etc. to do with the chiplet design.

 

[Bondrewd]

there are pretty big differences between different variants

Nah, not really, N5 variants are very samey.
Unlike say, N3 and N3e which are basically two entirely different nodes altogether.

the impact is probably not as bad

The channel normalized earlier than expected at least according to Gigabyte so it kinda stings, but AMD's non-quite-so-fucked mainstream options aren't far off so it's not all doom and gloom either.

but I'd wonder at this point if there is tradeoff/limitation/etc. to do with the chiplet design.

The tradeoff is that you spend more watts moving stuff off-chip, add some (relatively minor) packaging costs and some extra validation steps.

 

[Digidi]

Because N31have two different clock domain im curiose if the issue is only on one side of the domains.

 

[digitalwanderer]

Because N31have two different clock domain im curiose if the issue is only on one side of the domains.

Do you mean between the GCD and the chiplets or is this an infinity fabric thing?

 

[Digidi]

I'm taling about frontend and shaider domain.

 

[digitalwanderer]

I'm taling about frontend and shaider domain.

Thanks, apologies for my stupid. This is a bit over my head but I'm trying like heck to figure it out/follow it along. There is always so much more to learn, thanks for edumacating me.

 

[Bondrewd]

Because N31have two different clock domain im curiose if the issue is only on one side of the domains.

Aren't you the smartest cookie out there.

 

[digitalwanderer]

Aren't you the smartest cookie out there.

DEETS!

PS - And then please explain the details to me and ramifications using small words and pictorial aids...

 

[Digidi]

Aren't you the smartest cookie out there.

Im totaly stupid about gpus. Just wilde guessing...

 

[Bondrewd]

Im totaly stupid about gpus. Just wilde guessing...

Nah, it's just very evident since they've printed the slideware where lower SIMD clock saved them power.

 

[Digidi]

But i cant belave that power is issue. They could go easyl 500W and nobody will care for high end. so why not go 500w with good aio cooling and get the 3ghz?

Also intresting even 4090 running in an power limit. If you recude the power on 4090 massily you don't get much losses in fps. Mabe issue with 5nm process?

 

[Bondrewd]

They could go easyl 500W

No they can not, ATX isn't designed around handling 500W pressure cookers in tight cases.

so why not go 500w with good aio cooling and get the 3ghz?

They ~can't~.
It's broken.

 

[eastmen]

But i cant belave that power is issue. They could go easyl 500W and nobody will care for high end. so why not go 500w with good aio cooling and get the 3ghz?

Also intresting even 4090 running in an power limit. If you recude the power on 4090 massily you don't get much losses in fps. Mabe issue with 5nm process?

Well if its true and they are goign to go for a respin with bug fixes why blow your load on a crazy power hungry card when you can swim in the $1000 end of the pool and have a bigger splash when the higher clocked fixed part comes out ?