Question AMD Phoenix/Zen 4 APU Speculation and Discussion

[uzzi38]

I can finally make this thread.

https://twitter.com/x/status/1519669375283957760

Phoenix is indeed RDNA3. My advice to everyone: treat the old APU rumours as being out of date.

 

[Kaluan]

I agree that Polaris is very inefficient. It is also the case that the RX550 is quite undersized for the amount of bandwidth available. The RX560 is specced for the exact same total bandwidth and is markedly faster at roughly 40% in most cases, depending on which model you are testing.

The RX570 is slightly faster than the gtx 1650 and slightly slower than the 1650 super. If Rembrandt was as fast as the desktop 1650 today, then I would believe that PHX will blow past the RX570. Unfortunately, the desktop 1650 is, on average, 30+% faster than the 680m, with the RX570 being a bit faster. 30+% is a lot to overcome when you start getting to those levels. Is it possible? I suppose, but it's going to take more than bigger caches and 20% more memory bandwidth with a slightly higher iGPU clock to overcome that deficit.

Well 3400G, 4750G and 5700G share similar architectures and almost identical TFLOP grunt. But they still perform distinctly different and better gen-on-gen one fron each other.

4750G and 5700G in particular are bizarre, both are GCN v5.1 and same CU/TMU/ROP, 4750G even clocks higher. Yet somehow 5700G is consistently faster, (rarely, but) up to 20%.

And all of these desktop APUs get enough power to clock as per their spec. Where as on mobile, thermal headroom is always a factor.

So stranger things have happened.

680M vs "780M" @ 15W or 28W will be were the the real chip engineering prowess will shine.

Hopefully the Navi31 bugs are not present on Phoenix. lol

 

[LightningZ71]

Between 3400g and the 4x and 5x chips, there is a revision of the VEGA architecture. Raven ridge is based on VEGA64/56 CUs and Renoir up is based on VEGA VII. That may not seem like a big difference, but, there's a shader model update, some internal caches were changed, and the architecture was modified for higher clock speeds. When you overclock the Raven Ridge iGPU against the Renoir iGPU with a similar CU count and similar ram, Renoir still does better, though its not night and day.

Once you get to the 4750g vs the 5700g, you are facing a modified memory architecture. Cezanne has twice the L3, reducing dram bus contention. In addition, it has a single CCX as opposed to a pair of smaller ones, further simplifying internal data movement. Its not a huge difference again, but it does add up. I dare say that, for all of them, having a better CPU core running their drivers each generation doesn't hurt their performance either.

I suspect that Phoenix will see a notable performance increase due to the Zen 4 core improvements as well as the larger L2 that's available. I suspect that the larger internal caches of the iGPU won't hurt either. However, eventually, without enough VRAM bandwidth, the gpu will run out of things to do. I think that that is a big part of why rembrandt underperforms compared to its grunt, it just doesn't have enough memory bandwidth.

 

[uzzi38]

Remember there is a 64 bit version of the RX550 that is actually slower, so it is using that extra bandwidth, maybe not to full extend. The RX560 is petty much a double RX550, incluiding the L2 cache.

I still think there is something wrong with RMB that resulted in lower than expected performance and im not sure if memory bandwidth is the issue here. I think RDNA2 is build around IC and not having it here is creating some issues.

It's not memory bandwidth that's the issue, it's that the FCLK drops extremely low when the iGPU is at full tilt (a power saving feature) which badly hurts CPU performance.

It's the reason why games that are more CPU heavy are the ones that struggle the most, e.g. God Of War struggles to maintain over 40fps at all resolutions.

 

[Shivansps]

It's not memory bandwidth that's the issue, it's that the FCLK drops extremely low when the iGPU is at full tilt (a power saving feature) which badly hurts CPU performance.

It's the reason why games that are more CPU heavy are the ones that struggle the most, e.g. God Of War struggles to maintain over 40fps at all resolutions.

If thats the case then the IGP uses way too much power. And that is something they can actually fix.

Yeah it cant be a bandwidth issue, the 660M beats the desktop 5700G only by a hair, what likely means the 660M should perform the same with DDR4 while using less bandwidth than the Vega 8 on the 5700G, as RDNA2 should be more memory efficient than VEGA. The 680M should have plenty of bandwidth avalible using LPDDR5-6400 and it performs almost the same as DDR5-4800..

OR RDNA2 is not memory efficient AT ALL and the IC was created as a hotfix. Altrough i would not know why they would develop a APU with that if thats the case.

 

[LightningZ71]

Common GPU codebase. It makes maintaining the drivers MUCH easier if they are based around the same compute engines. Though, we know that RDNA1 was more memory efficient than VEGA, and there shouldn't be a regression on RDNA2. I do think, however, that RDNA2 was certainly conceived around the notion of the Infinity Cache. I don't think they would have bothered with the IC for the 6500/6400 if it wasn't essential to base performance for the platform.

 

[Glo.]

RDNA1 brought massive improvement in performance per CU, and in memory bandwidth efficiency.
RDNA2 was just improved RDNA1 with higher clock speeds, and Ray Tracing, but without increase in performance per CU, and also, again - improved memory efficiency thanks to Infinity Cache which alleviated a lot of memory bandwidth requirements.
RDNA3 is for RDNA1 and 2 the same thing as RDNA1 was for GCN.

Now, we have to ask a question. Would giving Infinity Cache to Radeon 680M bring performance higher than RX 6400(680M has higher clock speed than 6400, while having the same amount of CUs)? If it would - it bodes well for any APU iGPU with Infinity Cache, or System Level cache.

P.S. PHX is supposed to have performance of RX 570, yes? That is the performance level of RX 6400.

Without Infinity Cache - that is some architectural improvement, that AMD has achieved, if true.

 

[TESKATLIPOKA]

RDNA1 brought massive improvement in performance per CU, and in memory bandwidth efficiency.
RDNA2 was just improved RDNA1 with higher clock speeds, and Ray Tracing, but without increase in performance per CU, and also, again - improved memory efficiency thanks to Infinity Cache which alleviated a lot of memory bandwidth requirements.
RDNA3 is for RDNA1 and 2 the same thing as RDNA1 was for GCN.

Now, we have to ask a question. Would giving Infinity Cache to Radeon 680M bring performance higher than RX 6400(680M has higher clock speed than 6400, while having the same amount of CUs)? If it would - it bodes well for any APU iGPU with Infinity Cache, or System Level cache.

P.S. PHX is supposed to have performance of RX 570, yes? That is the performance level of RX 6400.

Without Infinity Cache - that is some architectural improvement, that AMD has achieved, if true.

I think we will have to wait some time before we can objectively say how good RDNA3 is compared to previous gen.
680M should be slower even with IC, It has only 1/2 of ROPs compared to RX 6400 and that one has Its own GDDR6 memory.

 

[Kaluan]

If memory serves me rights (pardon the pun), we still saw a ~8% (gaming) to ~15% (synthetics) gain LPDDR5-6400 (DR) vs DDR5-4800 (SR) on 2200MHz 680M. The R9 2400MHz 680M probably exacerbates that a few more %.
(in TDP scenarios were they can actually sustain those clocks). LPDDR/LPDDRX probably also shave off a few watts from the SoC/system that can go towards higher clocks on the chip itself too, but whatever.

 

[tamz_msc]

It's not memory bandwidth that's the issue, it's that the FCLK drops extremely low when the iGPU is at full tilt (a power saving feature) which badly hurts CPU performance.

It's the reason why games that are more CPU heavy are the ones that struggle the most, e.g. God Of War struggles to maintain over 40fps at all resolutions.

God of War ain't CPU-heavy. Scaling stops beyond 4C/8T and it is mostly GPU bound. APUs, especially in laptops, will always be power-limited, which is why people need to dial back on their expectations with Phoenix. We saw that with Rembrandt. 680M in the 6800U was only as fast as an MX450 GDDR6, and the latter was faster in most cases unless the game in question had issues with the 2 GB framebuffer. I expect a GTX 1650 to be the upper limit of the performance that Phoenix APUs can deliver in thin laptops.

 

[TESKATLIPOKA]

God of War ain't CPU-heavy. Scaling stops beyond 4C/8T and it is mostly GPU bound. APUs, especially in laptops, will always be power-limited, which is why people need to dial back on their expectations with Phoenix. We saw that with Rembrandt. 680M in the 6800U was only as fast as an MX450 GDDR6, and the latter was faster in most cases unless the game in question had issues with the 2 GB framebuffer. I expect a GTX 1650 to be the upper limit of the performance that Phoenix APUs can deliver in thin laptops.

I won't guess the final performance, but If someone wants Phoenix for casual gaming, then It's better to choose the higher TDP option.

6900HS even at 45W was capable to keep clocks at only ~2.2GHz. Link
Here you can compare It at 25W vs 45W

You loose 20% of performance just by limiting TDP to 25W instead of 45W.

A desktop GTX 1650 paired with 12900K is ~41% faster than 680M. Link

Everything depends on how much better is perf/W compared to Rembrandt. From specs we can't expect any improvement, only from architecture and the clocks.

 

[uzzi38]

God of War ain't CPU-heavy. Scaling stops beyond 4C/8T and it is mostly GPU bound. APUs, especially in laptops, will always be power-limited, which is why people need to dial back on their expectations with Phoenix. We saw that with Rembrandt. 680M in the 6800U was only as fast as an MX450 GDDR6, and the latter was faster in most cases unless the game in question had issues with the 2 GB framebuffer. I expect a GTX 1650 to be the upper limit of the performance that Phoenix APUs can deliver in thin laptops.

You said it yourself - scaling stops beyond 4c/8t.

The 6800U will constantly spike between 40 and 50% CPU utilisation (50% being the point where 4c/8t is fully saturated) once you lighten the load with FSR settings even at 1080p.

You can also see the GPU report that it's also maxed out on utilisation but somehow also sustaining a steady 1900MHz (the two of which combined don't make sense - power limits would cause issues with both the CPU and GPU so highly stressed). This indicates that the chip is idling while waiting on memory for most of the test run here.

That being said, I also don't expect Phoenix to exceed the 1650 by much (if at all) in thin and light notebooks either. As @TESKATLIPOKA pointed out, the 1650 is a good 40% higher performance than Rembrandt already, and I'm not expecting more than a 30% uplift to be frank.

 

[moinmoin]

You loose 20% of performance just by limiting TDP to 25W instead of 45W.

To me that seems like a great trade off, 20% less performance for 44% less power consumption. Just shows how the whole package is optimized for lower TDPs.

That in mind seems rather disproportionate to me to compare such with dGPUs or even desktop dGPUs.

 

[tamz_msc]

You said it yourself - scaling stops beyond 4c/8t.

The 6800U will constantly spike between 40 and 50% CPU utilisation (50% being the point where 4c/8t is fully saturated) once you lighten the load with FSR settings even at 1080p.

You can also see the GPU report that it's also maxed out on utilisation but somehow also sustaining a steady 1900MHz (the two of which combined don't make sense - power limits would cause issues with both the CPU and GPU so highly stressed). This indicates that the chip is idling while waiting on memory for most of the test run here.

That being said, I also don't expect Phoenix to exceed the 1650 by much (if at all) in thin and light notebooks either. As @TESKATLIPOKA pointed out, the 1650 is a good 40% higher performance than Rembrandt already, and I'm not expecting more than a 30% uplift to be frank.

That video has every game tested hitting a 30 W package power wall. Both CPU and GPU are power limited it would seem. The reason why the GPU is still at 1900 MHz is probably because its the frequency floor.

 

[Glo.]

RMB with 6000 MHz RAM - 2800 pts in 3dMark Time Spy, on stock clock speeds: 2400 MHz.
RMB with 4800 MHz RAM - 2500 pts in 3dMark Time Spy.

Bodes actually pretty well for Phoenix.

 

[tamz_msc]

I won't guess the final performance, but If someone wants Phoenix for casual gaming, then It's better to choose the higher TDP option.

6900HS even at 45W was capable to keep clocks at only ~2.2GHz. Link
Here you can compare It at 25W vs 45W
View attachment 72585
You loose 20% of performance just by limiting TDP to 25W instead of 45W.

A desktop GTX 1650 paired with 12900K is ~41% faster than 680M. Link

Everything depends on how much better is perf/W compared to Rembrandt. From specs we can't expect any improvement, only from architecture and the clocks.

Higher TDP options inevitably come with discrete GPUs, so it defeats the purpose of playing games with iGPU.

 

[tamz_msc]

RMB with 6000 MHz RAM - 2800 pts in 3dMark Time Spy, on stock clock speeds: 2400 MHz.
RMB with 4800 MHz RAM - 2500 pts in 3dMark Time Spy.

Bodes actually pretty well for Phoenix.

Time Spy isn't reflective of actual performance in games.

 

[TESKATLIPOKA]

To me that seems like a great trade off, 20% less performance for 44% less power consumption. Just shows how the whole package is optimized for lower TDPs.

That in mind seems rather disproportionate to me to compare such with dGPUs or even desktop dGPUs.

If you compare only the APU then yes, It's not a bad tradeoff, If the missing 20% is not making the game unplayable.
If you compare the whole laptop, then It's 33% less power consumption at best. Still better than what you loose in FPS.

Model	APU	TDP	Witcher 3 (1080p Ultra)	Power draw in Witcher 3
Asus Zenbook S 13 OLED	Ryzen 7 6800U	PL2: 30W ; PL1: 25W	23 FPS	Ø42.4 (40.2-44.5)
Lenovo Yoga 7-14ARB G7	Ryzen 7 6800U	PL2: 45W ; PL1: 25W	25.4 FPS	Ø46.3 (43.5-50.9)
Lenovo ThinkPad T14 G3	Ryzen 7 PRO 6850U	PL2: 30W ; PL1: 25W	23.9 FPS	Ø42.8 (40.7-45.7)
Lenovo ThinkPad Z13 G1	Ryzen 7 PRO 6860Z	PL2: 51W ; PL1: 27W	18.3 FPS	Ø48.1 (43.3-53.5)
HP EliteBook 845 G9	Ryzen 9 PRO 6950HS	PL2: 50W ; PL1: 40W	14.6 FPS (Single channel RAM)	Ø63.6 (61-67.9)

 

[TESKATLIPOKA]

Higher TDP options inevitably come with discrete GPUs, so it defeats the purpose of playing games with iGPU.

Business laptops like HP EliteBook will be without dGPU, but those notebooks will be pretty expensive and who knows how many models will be available.
A laptop with Phoenix + dGPU will likely end up cheaper than without a dGPU.
The question is what will OEMs put inside of gaming laptops, Dragon range or Phoenix.

Glo. -> Phoenix APU could be an interesting option for All in one systems.

 

[Glo.]

Business laptops like HP EliteBook will be without dGPU, but those notebooks will be pretty expensive and who know how many models will be available.
I am not sure If gaming laptops will be paired with Phoenix or Dragon Range or both.
For N33 even 8C16T Phoenix would be more than enough.
I don't expect a cheap laptop with Phoenix APU only.

We need to see those high performance 35W APUs in computers like: Dell Optiplex Micro, Lenovo Thinkcenter small form factor, etc.

 

[TESKATLIPOKA]

It's pretty bad that AMD is not planning a 50W GPU, I don't think N33 will go lower than 80W, the only barely usable thing is 6500M.
Nvidia will take that part of market with Ada107, which will be a pretty good upgrade over 3050(Ti) If they finally let go of 4GB Vram.

 

[Mopetar]

For a desktop APU I don't see any reason why AMD couldn't let it pull 105W or more, outside of not wanting to need to ship a more expensive cooler with it. AM5 had more than enough headroom to feed an APU like Phoenix with enough power to where it has really reached diminishing returns.

Even if it's just an option that users can enable, it'd be nice.

 

[LightningZ71]

Higher TDP options inevitably come with discrete GPUs, so it defeats the purpose of playing games with iGPU.

There are a few, and I mean FEW, laptops out there that come with H-class processors without dGPUs. They are usually more expensive than laptops that have a dGPU because those are typically sold as "pro" and "premium" models. They don't have to be though. Technically, the cost of producing a laptop with only an APU and providing it with sufficient cooling and power should be lower than any config that includes a dGPU, though, there are volume considerations and shared configurations that can come into play. They can use the same motherboard with just empty pads for the dGPU stuff, but, it would require a different cooler configuration than either lower end laptops that are APU only and higher end ones with a dGPU.

 

[Glo.]

For a desktop APU I don't see any reason why AMD couldn't let it pull 105W or more, outside of not wanting to need to ship a more expensive cooler with it. AM5 had more than enough headroom to feed an APU like Phoenix with enough power to where it has really reached diminishing returns.

Even if it's just an option that users can enable, it'd be nice.

And how much would you benefit from it in terms of performance?

Invest that additional power budget in memory bandwidth. But that is impossible without vast changes to AM5, and creating a new socket.

 

[Kaluan]

It's pretty bad that AMD is not planning a 50W GPU, I don't think N33 will go lower than 80W, the only barely usable thing is 6500M.
Nvidia will take that part of market with Ada107, which will be a pretty good upgrade over 3050(Ti) If they finally let go of 4GB Vram.

And why would they need a 50 or sub 50W mGPU? The gap between Phoenix IGP and 35-50W mGPUs will be tiny.

It's not like we know the TDP ranges of the N33 mobile SKUs either. Except for 7700M = up to 120W. Maybe.

They need a decent number of good designs (with competitive prices), not 10 for every niche segment, especially considering they probably won't have enough chips to go around for awhile anyway.

By the cycle after this one, they may as well have GFX1150/RDNA3+ Navi34 ready to flood the market with 50W SKUs.

 

[Kaluan]

TBH if they'd give us desktop Phoenix in next year's first half. I'd have half a mind to just get that and wait out Zen5VC and RDNA3+ or "Battlemage". 😂