• Competitor rules

    Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.

RDNA 3 rumours Q3/4 2022

Status
Not open for further replies.
Sounds silly, but is it about "beating nvidia"? It's about making sure their company maintains steady, profitable growth isn't it? I'm which case, why bother making something brilliant and cheap to deter people upgrading later? Just a thought.
 
gpuerrilla said:
No the rumours I seen were referring to the x700 tier. From RDNA2 the 6700 was navi22. Navi23 is the x600 tier so I think your tiers are mixed up. Much more likely a 7700XT being on par with a 6900XT, smaller vram but improved in ray tracing.
Click to expand...

All rumours have been using chip numbering not SKU names. In reality AMD can call N31, N32 and N33 whatever they want just like NV just called AD104 a 4080.

N33 is the 128bit 8GB part and is the one with the rumoured 6900XT performance at 1080p and 1440p. With 4k shaders and higher clocks than N23 it should be doable provided there is no memory bandwidth or other bottleneck.

N32 is an MCD design so you can bolt on 4 MCDs for a 256bit design for the 7800 series or make some cuts and use 3MCDs for the 7700 series.

N31 is also MCD and there is likely to be a 6 MCD 384 bit 24GB part and a 5 MCD 320bit 20GB part. Also there may be a version with 3D stacked infinity cache on those MCDs.

Since N33 is 128 bit with 8GB of ram it will not be the x700 tier sonce I highly doubt they will regress on the memory amount for a 1440p/entry 4k tier of product. This means that it will likely be x600 tier for 6900XT like performance.

For more info read these articles.
Angstronomics RDNA3 leak article
SemiAnalysis BOM cost article

From this it look like full N32 will have a very similar BOM to the 4080 12GB yet it will be a 7800 XT performance class and is very likely to be substantially faster than the 12GB 4080 and maybe even the 16GB one as well.

The N32 that will be used in the 7700XT will actually have a lower BOM due to 1 less MCD and 2 fewer memory chips. I actually expect the 7700XT to outperform the 4080 12GB by a margin and come in at a lower price.
 
Interesting. I very much hope your explanation comes true.

Generally speaking and looking to recent past, although the newer gen improves enough and trickles down, its going to be mighty generous if AMD presents a 7600 card that matches the 6900XT. Although this is the key part "provided there is no memory bandwidth or other bottleneck".

Should a 7700XT beat the 4070 4080 12GB by a margin, even the hardest of nvidia champions will find that one hard to swallow! :p
 
Purgatory said:
We have to hope for something good coming out of this GPU generation for gamers, Nvidia well gave us all the middle finger sadly. Nvidia sadly still has the pro market and CUDA market at their grubby hands.

Lets hope AMD gives gamers what they need and has the ability to provide it. Nvidia needs a kick in the rear this time and we as consumers of this tech need to let them know we have had enough and I hope AMD has been watching how people and their future customers have reacted so far to the 40 series and learned from it.
Click to expand...

Totally agree mate. But I like a lot of us who have been around the generational releases over the years have seen both companies drop the ball so many times. Let’s hope AMD buck the trend as NVidia have given them the golden opportunity!

Still not getting on the hype train though!
 
If memory serves me right semianalysis is on the “serious” end of the leaking game a.k.a they pay people off to get info.

They confirmed a number of interesting things.
1. only Nvidia is suffering from an over supply of previous generation
2. Angstronomics leaks is quite accurate.
3. The insane cost advantage AMD have over Nvidia.

One other thing to note is that they seemed confident that AMD could raise their market share to 30-35%. If true then it means AMD has more wafers for GPUs than last time.
 
Stanners said:
Totally agree mate. But I like a lot of us who have been around the generational releases over the years have seen both companies drop the ball so many times. Let’s hope AMD buck the trend as NVidia have given them the golden opportunity!

Still not getting on the hype train though!
Click to expand...

AMD have bragged about low production costs, good yields comes as a granted from chiplets, they could release at the same pricepoint (inc infaltion) as last gen - from what I gather - make RTX3000 obselete and steal the show.. or more profit.
 
Last edited:
Chuk_Chuk said:
Missed is an understatement. I can accept missing, it happens. But they went on to score an own goal because someone decided to boast about having lots of stock when they didn’t.
Click to expand...

Yeah I remember it well, the guy tweeted a gloat and the lack of stock on launch was laughable. He went on to buy one for himself but the UK didnt have them outlets for buying at.
 
gpuerrilla said:
We got excited last gen launch and AMD missed the open goal.
Click to expand...
AMD were using the wafers for Zen 3 and Milan. They didn't have enough to go around.

This time they have both N5 class and N7 class parts and many more wafers so have much more capacity to supply the market.

gpuerrilla said:
Interesting. I very much hope your explanation comes true.

Generally speaking and looking to recent past, although the newer gen improves enough and trickles down, its going to be mighty generous if AMD presents a 7600 card that matches the 6900XT. Although this is the key part "provided there is no memory bandwidth or other bottleneck".

Should a 7700XT beat the 4070 4080 12GB by a margin, even the hardest of nvidia champions will find that one hard to swallow! :p
Click to expand...

Well from what NV showed that 4080 12GB is ~= 3080Ti which is where the 6950XT sits in performance so we have 7,680 shaders @ ? clockspeeds being about equal to 5120 shaders @ 2.3Ghz or there abouts.

Full N32 will be 7680 shaders and I expect the cut version in the 7700XT will use at least 5120 (that is cutting 10 WGPs and 1 whole shader engine) but will be running at > 3 Ghz. So unless there is a large drop in IPC going to RDNA3 (I expect there to be some drop in IPC because each shader is smaller but I expect it to be more than made up for by the increased clockspeed) the 7700XT should have a good 20% or so more performance than the 6950XT which will beat the 4080 12GB in raster quite comfortably. RT I don't really know, hard to get an actual estimate at the moment given NVs obfuscation of native RT performance.
 
timorous said:
AMD were using the wafers for Zen 3 and Milan. They didn't have enough to go around.

This time they have both N5 class and N7 class parts and many more wafers so have much more capacity to supply the market.



Well from what NV showed that 4080 12GB is ~= 3080Ti which is where the 6950XT sits in performance so we have 7,680 shaders @ ? clockspeeds being about equal to 5120 shaders @ 2.3Ghz or there abouts.

Full N32 will be 7680 shaders and I expect the cut version in the 7700XT will use at least 5120 (that is cutting 10 WGPs and 1 whole shader engine) but will be running at > 3 Ghz. So unless there is a large drop in IPC going to RDNA3 (I expect there to be some drop in IPC because each shader is smaller but I expect it to be more than made up for by the increased clockspeed) the 7700XT should have a good 20% or so more performance than the 6950XT which will beat the 4080 12GB in raster quite comfortably. RT I don't really know, hard to get an actual estimate at the moment given NVs obfuscation of native RT performance.
Click to expand...

Until we get reviews, the only sort of apples to apples RT numbers who have for RTx4000 is using Nvidias published numbers that it uses for its RT cores

4080 16gb has 76 RT cores, each is rated for 1.5tflop = 114 tflops of RT performance

4090 has 128 RT cores, each is rated for 1.5tflop = 192tflops of RT performance

RTX3090ti has 84 RT cores, each is rated for 0.92tflops = 77tflops of RT performance

You'll notice that each RT core for RTX4000 has a large improvement in tflop performance output and much of this is because RTX4000 GPU clock speeds are about 50% to 60% higher than RTX3000

These numbers aren't perfect, the problem is that RT effects don't make up 100% of time processing a new frame, so you can double, triple, quadruple RT performance while only having small framerate improvement if RT effects only make up small portion of the frametime.

What you can say from this, the new 4080 and 4090 have greatly more powerful RT output, but games are not made up of 100% RT affects so the actual improvement in gaming framerate will not be known until review day
 
Last edited:
Grim5 said:
Until we get reviews, the only sort of apples to apples RT numbers who have for RTx4000 is using Nvidias published numbers that it uses for its RT cores

4080 16gb has 76 RT cores, each is rated for 1.5tflop = 114 tflops of RT performance

4090 has 128 RT cores, each is rated for 1.5tflop = 192tflops of RT performance

RTX3090ti has has 84 RT cores, each so rated for 0.92tflops = 77tflops of RT performance

You'll notice that each RT core for RTX4000 has a large improvement in tflop performance output and much of this is because RTX400 GPU clock speeds are about 50% higher than RTX3000

These numbers aren't perfect, the problem is that RT effects don't make up 100% of time processing a new frame, so you can double, triple, quadruple RT performance while only having small framerate improvement of RT effects only make up small portion of the frametime.

What you can say from this, the new 4080 and 4090 have greatly more powerful RT output, but games are not made up of 100% RT affects so the actual improvement in gaming framerate will not be known until review day
Click to expand...

Yea. I find it much better to show describe RT performance as the relative hit to fps vs a raster only image. For Ampere that was 40-50% and for RDNA2 that was a 60-70% hit. I expect Ada to be 30-40% perhaps maybe even less but if RDNA 3 can just match ampere if it does have a raw raster advantage to begin with even after taking a large hit for RT it can still pump out more frames.

You are right that we will need to wait for reviews to get a proper idea of RT performance.
 
Last edited:
Status
Not open for further replies.
Back
Top Bottom