Discussion RDNA4 + CDNA3 Architectures Thread

[DisEnchantment]

With the GFX940 patches in full swing since first week of March, it is looking like MI300 is not far in the distant future!
Usually AMD takes around 3Qs to get the support in LLVM and amdgpu. Lately, since RDNA2 the window they push to add support for new devices is much reduced to prevent leaks.
But looking at the flurry of code in LLVM, it is a lot of commits. Maybe because US Govt is starting to prepare the SW environment for El Capitan (Maybe to avoid slow bring up situation like Frontier for example)

See here for the GFX940 specific commits

History for llvm/lib/Target/AMDGPU - llvm/llvm-project

The LLVM Project is a collection of modular and reusable compiler and toolchain technologies. - History for llvm/lib/Target/AMDGPU - llvm/llvm-project

github.com

Or Phoronix

More AMD "GFX940" Enablement Work Landing In LLVM - Phoronix

www.phoronix.com

There is a lot more if you know whom to follow in LLVM review chains (before getting merged to github), but I am not going to link AMD employees.

I am starting to think MI300 will launch around the same time like Hopper probably only a couple of months later!
Although I believe Hopper had problems not having a host CPU capable of doing PCIe 5 in the very near future therefore it might have gotten pushed back a bit until SPR and Genoa arrives later in 2022.
If PVC slips again I believe MI300 could launch before it

This is nuts, MI100/200/300 cadence is impressive.



Previous thread on CDNA2 and RDNA3 here

Question - Speculation: RDNA3 + CDNA2 Architectures Thread

Man I have been dying to make this one for a while now. First rumours for RDNA3 are here so new thread time! Just going to start off with this one for now: kopite7kimi on Twitter: "@VideoCardz Ah, I mean a simple mcm design with 10240 cores is not enough. Because the lift from RDNA2 to RDNA3...

forums.anandtech.com

 

[Hans Gruber]

You're not reading what people wrote you:


The top RDNA4 card is going to be slower than the 7900xtx (I would bet - slower than 7900xt). Rumors state that it's a 240mm^2 GPU.

I read what they wrote. I was questioning the RDNA3 flaw that wiped out 25-30% of the chip. Meaning part of the GPU was deactivated and there is performance missing. If RDNA4 is an upgrade over RDNA3, does that mean they will have a fully functional GPU? I get that the new top is a 7800xt but in RDNA4 form. There is a 15% improvement just by using N4P vs. N5 in performance and even more in efficiency. The 4080 Super is as much as a brute in performance as the 4090. AMD needs to sell the RDNA4 flagship for $400 or less. This is getting stupid absurd.

The reason people still use and talk about the RX 5700 and 5700xt. It was an outstanding value compared to Nvidia. The performance per dollar was great. The RX 580 8GB cards are everywhere still because they went for $165-$185 all the time before the crypto craze.

If AMD fixed the design flaw with RDNA4 vs RDNA3, how could the new architecture not equal even the 7900xtx? Architectural gain, silicon gains on power/efficiency/performance and uplift with RDNA4 + the missing performance for RDNA3 design flaw.

 

[adroc_thurston]

I was questioning the RDNA3 flaw that wiped out 25-30% of the chip

It just clocks lower.
You can clock it the intended way, just that you won't like the power.

AMD needs to sell the RDNA4 flagship for $400 or less. This is getting stupid absurd.

You don't get to command anything.

If AMD fixed the design flaw with RDNA4 vs RDNA3

They're wholly different microarchitectures.

 

[Joe NYC]

I read what they wrote. I was questioning the RDNA3 flaw that wiped out 25-30% of the chip. Meaning part of the GPU was deactivated and there is performance missing. If RDNA4 is an upgrade over RDNA3, does that mean they will have a fully functional GPU? I get that the new top is a 7800xt but in RDNA4 form. There is a 15% improvement just by using N4P vs. N5 in performance and even more in efficiency. The 4080 Super is as much as a brute in performance as the 4090. AMD needs to sell the RDNA4 flagship for $400 or less. This is getting stupid absurd.

Resulted in lower performance per Watt, which lead to lower clock speed to stay in the intended power envelope.

It is not that difficult to understand.

If AMD fixed the design flaw with RDNA4 vs RDNA3, how could the new architecture not equal even the 7900xtx? Architectural gain, silicon gains on power/efficiency/performance and uplift with RDNA4 + the missing performance for RDNA3 design flaw.

Much smaller die size of Navi48 vs. combined die size Navi 31.

If AMD comes close in performance to Navi 31 with much lower die size and much lower BOM, that would make it a winner (in its BOM class).

 

[Hans Gruber]

Resulted in lower performance per Watt, which lead to lower clock speed to stay in the intended power envelope.

It is not that difficult to understand.



Much smaller die size of Navi48 vs. combined die size Navi 31.

If AMD comes close in performance to Navi 31 with much lower die size and much lower BOM, that would make it a winner (in its BOM class).

What people do not seem to understand is silicon. A few months ago forum members believed TSMC 5nm silicon was all the same. It's not and they released newer versions of silicon based on the 5nm platform that have vastly different performance/power differences.

In October 2021, TSMC introduced a new member of its "5 nm" process family: N4P. Compared to N5, the node offered 11% higher performance (6% higher vs N4), 22% higher power efficiency, 6% higher transistor density and lower mask count. TSMC expected first tapeouts by the second half of 2022

There is no clock regression with RDNA3 vs RDNA4 because both processes are on 5nm silicon. If anything the newer 5nm process will clock better and not worse.

Now to the broken RDNA3 architecture where 20-30% of the GPU was non functioning. When RDNA4 was developed, are they basing the performance metrics on what RDNA3 was supposed to be or what the real world performance of the flawed silicon had. Again, not factoring in specific GPU's 7900xtx/7900xt/7800xt but questioning the uplift in gen vs gen performance gains in power and efficiency. Even if you halve the GPU's, you would still have the same ratio of performance based on the GPU metrics at a reduced level based on GPU die size.

The 7800xt uses 252-268 watts of power under load while gaming heavy. If you take 22% off that alone, you would have power consumption of 202.8w if you had a 7800xt on N4P. I am using the middle number between 252 and 268w to get 260w minus 22%. That is still a bit higher than the 4070 but it's in the ballpark for efficiency. Then you add 11% of free performance based on N4P according to TSMC performance metrics of N4P. If AMD had a refresh of 7800xt with N4P, above would be what you should expect. It's for comparison purposes only.

So My point is even with a stripped down RDNA4. Should we not expect the 8800xtx to be close to if not equal or better than the 7900xtx? I am not accounting for any performance uplift between RDNA3 and RDNA or power efficiency that AMD is advertising in their RDNA4 GPU slides.

Finally I will add. I am speculating and not saying that I am right or wrong about RDNA4. I am just trying to figure out the future of AMD GPU's. I will add there is speculation that RDNA4 GPU core clocks will boost up to 3.2ghz. Remember when RDNA3 was announced? They said the GPU would clock up to 3ghz and it didn't. The reason has been speculated that RDNA3 was a broken but functional architecture. That is why I am speculating that a RDNA4 GPU could/should perform much better than RDNA3 if the architecture is fully functional and sound.

 

[blckgrffn]

What people do not seem to understand is silicon. A few months ago forum members believed TSMC 5nm silicon was all the same. It's not and they released newer versions of silicon based on the 5nm platform that have vastly different performance/power differences.

In October 2021, TSMC introduced a new member of its "5 nm" process family: N4P. Compared to N5, the node offered 11% higher performance (6% higher vs N4), 22% higher power efficiency, 6% higher transistor density and lower mask count. TSMC expected first tapeouts by the second half of 2022

There is no clock regression with RDNA3 vs RDNA4 because both processes are on 5nm silicon. If anything the newer 5nm process will clock better and not worse.

Now to the broken RDNA3 architecture where 20-30% of the GPU was non functioning. When RDNA4 was developed, are they basing the performance metrics on what RDNA3 was supposed to be or what the real world performance of the flawed silicon had. Again, not factoring in specific GPU's 7900xtx/7900xt/7800xt but questioning the uplift in gen vs gen performance gains in power and efficiency. Even if you halve the GPU's, you would still have the same ratio of performance based on the GPU metrics at a reduced level based on GPU die size.

The 7800xt uses 252-268 watts of power under load while gaming heavy. If you take 22% off that alone, you would have power consumption of 202.8w if you had a 7800xt on N4P. I am using the middle number between 252 and 268w to get 260w minus 22%. That is still a bit higher than the 4070 but it's in the ballpark for efficiency. Then you add 11% of free performance based on N4P according to TSMC performance metrics of N4P. If AMD had a refresh of 7800xt with N4P, above would be what you should expect. It's for comparison purposes only.

So My point is even with a stripped down RDNA4. Should we not expect the 8800xtx to be close to if not equal or better than the 7900xtx? I am not accounting for any performance uplift between RDNA3 and RDNA or power efficiency that AMD is advertising in their RDNA4 GPU slides.

Finally I will add. I am speculating and not saying that I am right or wrong about RDNA4. I am just trying to figure out the future of AMD GPU's. I will add there is speculation that RDNA4 GPU core clocks will boost up to 3.2ghz. Remember when RDNA3 was announced? They said the GPU would clock up to 3ghz and it didn't. The reason has been speculated that RDNA3 was a broken but functional architecture. That is why I am speculating that a RDNA4 GPU could/should perform much better than RDNA3 if the architecture is fully functional and sound.

You are presuming that the RDNA4 dies are the same size relative to the RDNA3 dies. Like its a die shrink of the 7800XT.

They aren't.

 

[linkgoron]

If AMD fixed the design flaw with RDNA4 vs RDNA3, how could the new architecture not equal even the 7900xtx? Architectural gain, silicon gains on power/efficiency/performance and uplift with RDNA4 + the missing performance for RDNA3 design flaw.

The card is rumored to be 240mm^2 in 4nm. For the record AD104 (aka Geforce 4070TI) is ~290mm^2 on whatever 4nm/5nm variant Nvidia is using.

Navi 31's GCD size alone is 300nm^2, without the MCDs. Navi 32 has a 200mm^2 GCD without MCDs. Navi 48 is supposedly 240mm^2 monolithic.

 

[Joe NYC]

There is no clock regression with RDNA3 vs RDNA4 because both processes are on 5nm silicon. If anything the newer 5nm process will clock better and not worse.

Clock "regression" was between RDNA3 anticipated and RDNA3 actual

Now to the broken RDNA3 architecture where 20-30% of the GPU was non functioning.

Where are you getting this stuff? All of the GPU was functioning, just running at lower clock speeds.

When RDNA4 was developed, are they basing the performance metrics on what RDNA3 was supposed to be or what the real world performance of the flawed silicon had. Again, not factoring in specific GPU's 7900xtx/7900xt/7800xt but questioning the uplift in gen vs gen performance gains in power and efficiency. Even if you halve the GPU's, you would still have the same ratio of performance based on the GPU metrics at a reduced level based on GPU die size.

The 7800xt uses 252-268 watts of power under load while gaming heavy. If you take 22% off that alone, you would have power consumption of 202.8w if you had a 7800xt on N4P. I am using the middle number between 252 and 268w to get 260w minus 22%. That is still a bit higher than the 4070 but it's in the ballpark for efficiency. Then you add 11% of free performance based on N4P according to TSMC performance metrics of N4P. If AMD had a refresh of 7800xt with N4P, above would be what you should expect. It's for comparison purposes only.

So My point is even with a stripped down RDNA4. Should we not expect the 8800xtx to be close to if not equal or better than the 7900xtx? I am not accounting for any performance uplift between RDNA3 and RDNA or power efficiency that AMD is advertising in their RDNA4 GPU slides.

Assuming RDNA4 ha greater area efficiency, it still have much smaller die. We will see how these two offset each other.

Finally I will add. I am speculating and not saying that I am right or wrong about RDNA4. I am just trying to figure out the future of AMD GPU's. I will add there is speculation that RDNA4 GPU core clocks will boost up to 3.2ghz. Remember when RDNA3 was announced? They said the GPU would clock up to 3ghz and it didn't. The reason has been speculated that RDNA3 was a broken but functional architecture. That is why I am speculating that a RDNA4 GPU could/should perform much better than RDNA3 if the architecture is fully functional and sound.

I think a good yardstick for now is Navi 44 will have lower cost than Navi 32 and higher performance. By a decent margin.

AMD is just not targeting top performance with large die. That ship has sailed with cancellation of chiplet based RDNA 4, and there is not going back. AMD will just not have a Halo part, but may offer good performance / price in the mainstream, $400-$500 price range market segment.

 

[Hans Gruber]

You are presuming that the RDNA4 dies are the same size relative to the RDNA3 dies. Like its a die shrink of the 7800XT.

They aren't.

I did not assume anything. I am just trying to figure out all the performance metrics separate from the nice silicon that many outlets have speculated AMD will use. TSMC has been encouraging their customers to use N4P and I think they offered financial incentives to push them in that direction. N3 3nm sucks other than impressive efficiency improvements. The next iteration or N3P will give a performance uplift and increased effc

The card is rumored to be 240mm^2 in 4nm. For the record AD104 (aka Geforce 4070TI) is ~290mm^2 on whatever 4nm/5nm variant Nvidia is using.

Navi 31's GCD size alone is 300nm^2, without the MCDs. Navi 32 has a 200mm^2 GCD without MCDs. Navi 48 is supposedly 240mm^2 monolithic.

The rumors are the big one for RDNA4 will be 300-350mm. The thing people have to understand. Nvidia is going to knock it out of the park. It's not a good idea to compare RDNA4 to the 4070ti. It's the 5070 that will be a comparable card to the AMD top (RDNA4) GPU.

 

[Saylick]

Where are you getting this stuff? All of the GPU was functioning, just running at lower clock speeds.

I thought there was something borked in RDNA 3 that got disabled for the launch... I think it was the OREO feature that Angstronomics alluded to? Did the Scaler ALU get implemented as well?

 

[marees]

The rumors are the big one for RDNA4 will be 300-350mm.

You seem to have your own source of rumours

Just going by numbering,

for RDNA 2 we had navi 21, navi 22, navi 23 & navi 24

For RDNA 4 we have navi 41 (scrapped), navi 42 (scrapped), navi 43 (respun as navi 48), navi 44

Basically navi 44 is the RDNA 4 version of 6500 xt & navi 48 is the RDNA 4 version of 6600 xt
(Or equivalently polaris 14 & polaris 11)

So the chip size is pretty low.

Navi 44 should be equal to 4060 ti super
Navi 48 double that probably 4070 ti super

 

[Kepler_L2]

I thought there was something borked in RDNA 3 that got disabled for the launch... I think it was the OREO feature that Angstronomics alluded to? Did the Scaler ALU get implemented as well?

There were some issues with OREO but it wasn't disabled. Scalar FPU was never part of RDNA3.

 

[Hans Gruber]

I thought there was something borked in RDNA 3 that got disabled for the launch... I think it was the OREO feature that Angstronomics alluded to? Did the Scaler ALU get implemented as well?

AMD has never officially admitted RDNA3 was broken from launch. Part of the GPU was disabled and AMD said part of the GPU was disabled for a future card that would have the disabled part of the card enabled.

Before RDNA3 launched AMD promised 3ghz boost speeds that never came to be. This reinforces the belief that RDNA3 was broken and could not be fixed. If RDNA4 launches with 3-3.2Ghz boost speeds. It would reinforce the speculation regarding RDNA3. There was talk of RDNA 3.5 which would have addressed/fixed the issues with RDNA3.

I still do not even know how much performance was lost/missing because of the boost clocks falling 500mhz short or part of the GPU being disabled because it was not able to function correctly.

You seem to have your own source of rumours

Just going by numbering,

for RDNA 2 we had navi 21, navi 22, navi 23 & navi 24

For RDNA 4 we have navi 41 (scrapped), navi 42 (scrapped), navi 43 (respun as navi 48), navi 44

Basically navi 44 is the RDNA 4 version of 6500 xt & navi 48 is the RDNA 4 version of 6600 xt
(Or equivalently polaris 14 & polaris 11)

So the chip size is pretty low.

Navi 44 should be equal to 4060 ti super
Navi 48 double that probably 4070 ti super

If your numbers are correct. AMD will be going head to head with Intel Battlemage. Both AMD and Intel should be on N4P. I think Nvidia will be using N4X or something close to it. N4X has 6% better performance than N4P.

 

[blckgrffn]

"I did not assume anything. I am just trying to figure out all the performance metrics separate from the nice silicon that many outlets have speculated AMD will use. "

"The rumors are the big one for RDNA4 will be 300-350mm."

Those two sentences contradict each other.

Maybe this whole thread is blowing smoke but it seems like the information is increasingly well sourced.

 

[Joe NYC]

I thought there was something borked in RDNA 3 that got disabled for the launch... I think it was the OREO feature that Angstronomics alluded to? Did the Scaler ALU get implemented as well?

My take is that they had to increase voltage to maintain stability. Which erased all of the potential power efficiency gains.

 

[linkgoron]

The rumors are the big one for RDNA4 will be 300-350mm. The thing people have to understand. Nvidia is going to knock it out of the park. It's not a good idea to compare RDNA4 to the 4070ti. It's the 5070 that will be a comparable card to the AMD top (RDNA4) GPU.

More AMD RDNA 4 GPU Lineup info surfaces: Navi 48 and Navi 44 Details

While the official launch of AMD's RDNA 4 GPU architecture might not occur within 2024, leaked specifications suggest the lineup will include only two models: Navi 48 and Navi 44.

www.guru3d.com

per source:

Navi 48 model is reported to feature 64 Compute Units, a 256-bit memory interface, and a memory bandwidth of 693 GB/s, potentially reaching up to an effective 2770GB/s. This GPU's die size is expected to be around 240mm². For comparison, the Radeon RX 7800 XT, powered by the Navi 32 GPU, occupies a larger area of 346mm², with 60 Compute Units and a slightly lower memory bandwidth of 624.1GB/s

On the other hand, the Navi 44 GPU appears designed to succeed the Radeon RX 7600 series, boasting 32 Compute Units and a 128-bit memory interface, along with a memory bandwidth of 288GB/s. The estimated die size for Navi 44 is 130mm², presenting a more compact option compared to its predecessor, the Navi 33, which also has 32 Compute Units but a larger die size of 204mm².

 

[marees]

Before RDNA3 launched AMD promised 3ghz boost speeds

I believe RDNA 3 can clock as high as 3.75ghz but power consumption in typical gaming loads restricts max boost

The cards & arch are designed to boost as high as possible (the limit is power)

 

[reaperrr3]

What people do not seem to understand is silicon.

No offense, but neither do you seem to.
- Those +11% perf/-22% power numbers are either/or numbers, you get one or the other, not both. It's either 11% more perf aka clockspeed at same power budget, or 22% less power at same clocks. Or ~5% higher clocks at 10-12% less power, if you look for the middle ground.
- Those N4P numbers are vs. N5, but RDNA3 might already be using N5P (AMD usually doesn't officially state which exact process variant they use), so the difference could be even less.

> Now to the broken RDNA3 architecture where 20-30% of the GPU was non functioning.

As you've already been told, there is no "20-30% of the GPU not functioning". There's just some architectural flaw(s) that makes RDNA3 extremely power-hungry even at relatively modest voltages and clocks, preventing the N3x GPUs from hitting the 3+ GHz they were intended for.
IPC might've landed a little below target, too, but that doesn't mean there's some magic fix for that.

So My point is even with a stripped down RDNA4. Should we not expect the 8800xtx to be close to if not equal or better than the 7900xtx?

In RT, probably. In raster, most likely not.
All signs are currently pointing to N48 featuring only 2/3 of the CU/WGP/ROP/IF$ that N31 has, meaning even at perfect scaling, it would need 50% higher clocks to match N31, if the "IPC" per WGP stays about the same.
And even if N48 hits somewhere between 3 to 3.2 GHz and has somewhat higher "IPC" per WGP (unknown, since they might've optimized the WGPs for similar IPC at less area and power + higher clocks instead), the smaller IF$ and lower memory bandwidth may still hold it back versus the 79XTX.
Plus, the 7700XT/7800XT only got those high TDPs because that was the only way to make them competitive enough.
If N48 is as small and fast as suggested, AMD will probably want to lower the TDP, as moving away 250+W of heat from such a small die isn't easy.

Which is probably fine for AMD, since judging by the specs, N48 seems to be mostly aiming to fight salvage GB205 (5070) and the 5060 Ti (GB206) at similar (a.k.a. higher than for N32) margins compared to Nvidia.

 

[dhruvdh]

My 7900XTX runs at 3GHz for compute workloads (pytorch) even when capped at 330W.

 

[Abwx]

In October 2021, TSMC introduced a new member of its "5 nm" process family: N4P. Compared to N5, the node offered 11% higher performance (6% higher vs N4), 22% higher power efficiency, 6% higher transistor density and lower mask count. TSMC expected first tapeouts by the second half of 2022

It s either 11% higher perf or 22% higher efficency, the two possibilities are exclusive, you cant have both at the same time, although there s something wrong in the efficency number, it s rather 22% lower power, wich amount actually to 28% better efficency.

 

[Hans Gruber]

It s either 11% higher perf or 22% higher efficency, the two possibilities are exclusive, you cant have both at the same time, although there s something wrong in the efficency number, it s rather 22% lower power, wich amount actually to 28% better efficency.

From what I read it means clock for clock the performance increase is 11% N5 vs. N4P and a 22% reduction in power/efficiency N5 vs N4P. I believe the numbers are right considering Nvidia was on a version of silicon similar to N4P in power and efficiency. Take the 4070 N4--something vs 7800XT that was on N5. 22% higher efficiency (N4P) means less power to achieve the same output of N5 silicon.

The 11% higher performance is the performance boost provided to the GPU by being on the superior silicon N4P vs. N5. According to the numbers from TSMC that means if RDNA4 is on N4, it would be a 5% performance gain. The simplest way to test the results would be if AMD made a 7800XT on N4P and tested the performance of a standard 7800XT on N5 silicon. I am sure they do this already but never share the results.

It will be very tough for AMD to sell the top variant of RDNA4 for $550 if the performance is equal to a 4070/Super/Ti. The Blackwell 5070 is coming out in early 2025 that will smoke the 4070 variants. If Nvidia prices the 5070 around $550. Where does that leave RDNA4? $350-400. Add to the competition Intel's Battlemage top offering is supposed to be 4070ti performance. Which would mean Intel would be the main competition to AMD next year.

I personally think it would make a lot of sense to have a 7900xtx on N4P. That would fill the gap of not offering a high end card with RDNA4.

 

[Joe NYC]

From what I read it means clock for clock the performance increase is 11% N5 vs. N4P and a 22% reduction in power/efficiency N5 vs N4P.

There is no clock for clock gain (in other words IPC) gain from process node. Only from changing architecture.

From process node, you can either clock higher, or use less power at the same clock speed.

It will be very tough for AMD to sell the top variant of RDNA4 for $550

The top RDNA 4 parts will be cheaper to make that N32, which sells from $500, at which price, AMD would have higher margins than on N32.

I personally think it would make a lot of sense to have a 7900xtx on N4P. That would fill the gap of not offering a high end card with RDNA4.

The RDNA 4 architects have already moved on to RDNA 5.

And your suggestion is for them to go to RDNA 3? Definitely not happening.

NVidia enthusiasts are just going to get raped by NVidia buying 5080, 5090 cards. Make peace with that.

 

[linkgoron]

NVidia enthusiasts are just going to get raped by NVidia buying 5080, 5090 cards. Make peace with that.

Not only Nvidia enthusiasts, people who want to upgrade from Navi31 or 6800/6900xt or older etc that didn't think that Navi 31 was good enough for them to upgrade. It's either that, or wait at least two more years in the hopes that AMD will actually deliver a proper product, with meaningfully better prices than Nvidia, and doesn't actually cancel again - none of which is certain.

 

[Joe NYC]

Not only Nvidia enthusiasts, people who want to upgrade from Navi31 or 6800/6900xt or older etc that didn't think that Navi 31 was good enough for them to upgrade. It's either that, or wait at least two more years in the hopes that AMD will actually deliver a proper product, with meaningfully better prices than Nvidia, and doesn't actually cancel again - none of which is certain.

There is an option to pay $1,800 for 4090 for people who think $900 Radeon 7900 XTX is not good enough for them to upgrade.

Very simple choice there (at my local MicroCenter prices).

 

[linkgoron]

There is an option to pay $1,800 for 4090 for people who think $900 Radeon 7900 XTX is not good enough for them to upgrade.

Very simple choice there (at my local MicroCenter prices).

In the not too distant future they will probably have a much cheaper option from Nvidia with that level of performance, while AMD will have zero until at least 2026.

 

[branch_suggestion]

From what I read it means clock for clock the performance increase is 11% N5 vs. N4P and a 22% reduction in power/efficiency N5 vs N4P. I believe the numbers are right considering Nvidia was on a version of silicon similar to N4P in power and efficiency. Take the 4070 N4--something vs 7800XT that was on N5. 22% higher efficiency (N4P) means less power to achieve the same output of N5 silicon.

The 11% higher performance is the performance boost provided to the GPU by being on the superior silicon N4P vs. N5. According to the numbers from TSMC that means if RDNA4 is on N4, it would be a 5% performance gain. The simplest way to test the results would be if AMD made a 7800XT on N4P and tested the performance of a standard 7800XT on N5 silicon. I am sure they do this already but never share the results.

It will be very tough for AMD to sell the top variant of RDNA4 for $550 if the performance is equal to a 4070/Super/Ti. The Blackwell 5070 is coming out in early 2025 that will smoke the 4070 variants. If Nvidia prices the 5070 around $550. Where does that leave RDNA4? $350-400. Add to the competition Intel's Battlemage top offering is supposed to be 4070ti performance. Which would mean Intel would be the main competition to AMD next year.

I personally think it would make a lot of sense to have a 7900xtx on N4P. That would fill the gap of not offering a high end card with RDNA4.

N4P wasn't ready for a big die launch in 2022, A16 was the lead N4P part and that was September 2022.
It was N4 vs N5P basically, pretty much irrelevant.
Keep ignoring the evidence if you must, but there was seemingly a physdes issue with RDNA3 in gaming workloads, this was true on discrete and integrated versions, across N6/5/4.
Probably BEOL/FEOL. Less than 2 days for RDNA3.5 to prove that was the case.