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 My advice to everyone: treat the old APU rumours as being out of date.

 

[TESKATLIPOKA]

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.

I am still very skeptical about 4000 points for 780M, but I will use that value.

	680M	780M	6500M 55W	3050TI 50W	6800S (Full N23) 90-105W
Time Spy Graphics	~2500	~4000	4434	~5000	~9000
Performance in %	100%	160%	177%	200%	360%

3dmark should be the best case for IGPs.

I think even the weakest N33 should at least manage 9000 points, so It would be 2.25x faster than Phoenix and AMD won't have anything between them.

3050ti would be only 25% faster, but when the first Phoenix laptops are out, I expect Ada107 to be released too.

Ada107 could be 25-50% faster than 3050Ti within the same TDP, and that would mean 56-88% faster than Phoenix IGP.

 

[Kaluan]

I am still very skeptical about 4000 points for 780M, but I will use that value.

	680M	780M	6500M 55W	3050TI 50W	6800S (Full N23) 90-105W
Time Spy Graphics	~2500	~4000	4434	~5000	~9000
Performance in %	100%	160%	177%	200%	360%

3dmark should be the best case for IGPs.

I think even the weakest N33 should at least manage 9000 points, so It would be 2.25x faster than Phoenix and AMD won't have anything between them.

3050ti would be only 25% faster, but when the first Phoenix laptops are out, I expect Ada107 to be released too.

Ada107 could be 25-50% faster than 3050Ti within the same TDP, and that would mean 56-88% faster than Phoenix IGP.

Unlikely IMHO. What's a low TGP (40-50W) 3050 in TS graphics? 3300-4300 points. Ti's not much higher than that either.
So not exactly earth shattering more performance over Rembrandt. Just 1 tier-ish higher.
Phoenix may be the same with AD107.


Cut down N23 designs hover 7500-8500 points. I doubt lower stack N33 could only manage 9000.

But anyway, not sure why any of these matter. TGP is configurable for Radeon mGPUs. There's a bunch of laptops with 6600M configured for lower. AMD just doesn't advertise it and ODM don't seem to bother often either.

I have a hunch we won't see MX650 in nVidia's lineup next year.

 

[TESKATLIPOKA]

Unlikely IMHO. What's a low TGP (40-50W) 3050 in TS graphics? 3300-4300 points. Ti's not much higher than that either.
So not exactly earth shattering more performance over Rembrandt. Just 1 tier-ish higher.
Phoenix may be the same with AD107.

RTX 3050 50W is 3931-4503, so that's 57-80% higher performance than Rembrandt.
This is a significant difference which could allow you to play at higher details or playable frame rates. I certainly wouldn't dismiss It.

Not sure about mobile but in desktop RTX 4080(320W) managed ~60% higher perf/W against RTX 3080(320W).
So I wouldn't be surprised If AD107 managed at least 7000(+40% over 3050Ti) and that is 75% more performance than Phoenix.

If Phoenix manages 50% more performance than Rembrandt, then It will only compare to a cutdown 50W GA107 at best, It will be far behind ADA107.
I have to say for an IGP that would be a very good result and you also save on power draw, but there are also significant disadvantages.
Phoenix without dGPU will be put inside premium laptops, so It will likely cost more than a laptop with GA107 or even AD107. Then you also have to sacrifice a part of your main memory for IGP and the lower performance compared to dGPU.

Cut down N23 designs hover 7500-8500 points. I doubt lower stack N33 could only manage 9000.

But anyway, not sure why any of these matter. TGP is configurable for Radeon mGPUs. There's a bunch of laptops with 6600M configured for lower. AMD just doesn't advertise it and ODM don't seem to bother often either.

I have a hunch we won't see MX650 in nVidia's lineup next year.

I set It to a very conservative 9000, because I don't know how low TDP will go. Even with this conservative estimate, you have a big gap, which will Nvidia take.

I don't see why MX650 wouldn't be released.
MX570(GA107) already has similar performance to Phoenix, although It has a huge disadvantage in only 2GB Vram.
They could rebrand It to MX650 but with 4GB Vram and It will sell very well. The reason being Intel has by far the highest volume and a much weaker IGP.

 

[Kaluan]

First sighting of N33 was in a 2023 Rembrandt refresh, I'm guessing they are trying to phase out N7 RDNA2 mobile parts. Which means they can make a lot of these and likely, for cheaper.
So far N33 looks like it can be a drop-in upgrade on older chassis.

Not many concrete leaks on 4000 mobile yet, which is weird for how loose lips nVidia usually is.

It would be smart if AMD offered official TGP configuration for N33 @ <80W, but at the end of the day it's up to the ODMs more than AMD.

I just hope we get to see at least 1 good PHX-U dGPUless and one PHX-HS + N33 reviewed as well by the end of Q1. Personally might be in the market for one.

 

[TESKATLIPOKA]

In my opinion Nvidia still has a sizeable inventory of mobile chips to sell. I hope we will see some mobile ones at CES 2023.

Not sure If It would be a great idea to go even lower than 80W for N33, but GA106 also has 60-115W configurable TGP.

I can imagine we will see reviews of PHX-H or HS by the end of Q1, possibly with N33 If AMD announces It at CES.
PHX-U I expect in Q2.

 

[Glo.]

I think in this competitive discussion we forget that there will be 6 GB 3050 mobile which supposedly scores around 6500 PTS in TS.

 

[TESKATLIPOKA]

I think in this competitive discussion we forget that there will be 6 GB 3050 mobile which supposedly scores around 6500 PTS in TS.

First time hearing about this. For that performance, I can only think about a cutdown 24SM GA106 with 192bit bus 6gb Vram.

 

[Glo.]

First time hearing about this. For that performance, I can only think about a cutdown 24SM GA106 with 192bit bus 6gb Vram.

It will be a cut down AD107 with 96 bit bus, most likely. Its supposed to be RTX 2050 replacement.

 

[TESKATLIPOKA]

It will be a cut down AD107 with 96 bit bus, most likely. Its supposed to be RTX 2050 replacement.

If It's AD107 based then shouldn't It be 4050 series? I know Nvidia made RTX 2050 based on GA107, but that was 1/2 year after releasing RTX 3050 laptop series.

RTX 4050 -> 20SM(2560 Cuda), 96-bit 6GB Vram
RTX 4050ti -> 24SM(3072 Cuda), 128-bit 8GB Vram

This would look pretty good and finally more Vram for lowend.
Not that good for Phoenix unless It's paired with It.

I am starting to like the combination of Phoenix + RTX 4050Ti If It will be for a reasonable price. RDNA3 doesn't look that hot to me, unless they fix It.

Tigerick: Thanks for the info. TGP seems a bit high in my opinion, but It could happen with increased clockspeed.

 

[Tigerick]

If It's AD107 based then shouldn't It be 4050 series? I know Nvidia made RTX 2050 based on GA107, but that was 1/2 year after releasing RTX 3050 laptop series.

RTX 4050 -> 20SM(2560 Cuda), 96-bit 6GB Vram
RTX 4050ti -> 24SM(3072 Cuda), 128-bit 8GB Vram

This would look pretty good and finally more Vram for lowend.
Not that good for Phoenix unless It's paired with It.

I am starting to like the combination of Phoenix + RTX 4050Ti If It will be for a reasonable price. RDNA3 doesn't look that hot to me, unless they fix It.

From the draft copy of upcoming TUL for China market, AD107 comes with 2 versions:

RTX4050 2304SP with 6GB 96-bit 16Gbps 70W TS 7300
RTX4050Ti 3072SP with 8GB 128-bit 16Gbps 95W TS 9600

Of course, changes in model name might happen but the specs should be correct.

 

[firewolfsm]

3050 Ti is 35-75W. I don't see why 4050 Ti would increase that so much, as it's going to be a design free swap for most manufacturers.

 

[TESKATLIPOKA]

firewolfsm:
I moved my reply to the correct thread. Ada/'Lovelace'? Next gen Nvidia gaming architecture speculation
If anyone wants to continue discussing Ada, let's discuss It there.

 

[Shivansps]

As i say, a 6CU 660M is slightly better than a 5700G Vega 8 that has about 36-38GB/s, i would be very suprised that DDR5-6400 is still not enoght for 12CU we are talking north of 80GB/s here. Something is up here, as the 660M should, in theory, perform better than Vega 8 on the same bandwidth. Either RDNA2 is performs poorly without the IC or there is a huge bottleneck issue in RMB, if the IF downclocks that is likely the issue.

 

[uzzi38]

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.

What? Frequency floor for all RDNA2 products under load is 500MHz, even the desktop cards. When idle it reports 0MHz frequencies, and if it spends time idle whilst occasionally enabling the core for minor workloads (e.g. a menu screen) it will report <500MHz frequencies.

AMD's boost algorithm for GPU is extremely well tuned, it'll downclock even when the GPU is under load if target framerate is hit. Conversely, it will boost as much as it can in order to try and hit said target framerate when not.

 

[FlameTail]

How does the iGPUs in Apple M chips comparw in raw performance to Radeon 660M/680M and potential Phoenix with RDNA 3 ?

Now, i understand that this is not an ideal comparison as Apple GPUs are not optimised for gaming ( thanks Metal API), which is why i said raw performance. That can be measured by benchmsrks.

 

[maddie]

From the draft copy of upcoming TUL for China market, AD107 comes with 2 versions:

RTX4050 2304SP with 6GB 96-bit 16Gbps 70W TS 7300
RTX4050Ti 3072SP with 8GB 128-bit 16Gbps 95W TS 9600

Of course, changes in model name might happen but the specs should be correct.

Now how can I take anything seriously with that join date? Cheers.

 

[tamz_msc]

What? Frequency floor for all RDNA2 products under load is 500MHz, even the desktop cards. When idle it reports 0MHz frequencies, and if it spends time idle whilst occasionally enabling the core for minor workloads (e.g. a menu screen) it will report <500MHz frequencies.

AMD's boost algorithm for GPU is extremely well tuned, it'll downclock even when the GPU is under load if target framerate is hit. Conversely, it will boost as much as it can in order to try and hit said target framerate when not.

I doubt it's 500 MHz in your typical gaming load. By typical I mean your modern 3D games that are taxing on the GPU.

 

[Kaluan]

I doubt it's 500 MHz in your typical gaming load. By typical I mean your modern 3D games that are taxing on the GPU. And lower when idle or 2D/desktop minor workloads.

I think they mean 3D load frequency *floor* (think for example having 60fps frame capped eSports game, GPU will not clock lower than 500MHz even if it only needs 300 for that), you're thinking of frequency *ceiling* (fmax?)
So nothing to do with "typical" scenarios.

I think in this competitive discussion we forget that there will be 6 GB 3050 mobile which supposedly scores around 6500 PTS in TS.

At what power envelope? 3060 at 65W does < 6500.

Navi34, like it's predecessor, may take a year to make a showing. At which point it might as well be RDNA3+ based, on N4. Likely monolithic <100mm2 die. But will post some insane P/W values.
Tho at that point it may have to contend with Arc "B3xxM" as well.

 

[tamz_msc]

I think they mean 3D load frequency *floor* (think for example having 60fps frame capped eSports game, GPU will not clock lower than 500MHz even if it only needs 300 for that), you're thinking of frequency *ceiling* (fmax?)
So nothing to do with "typical" scenarios.

I'm not including V-sync or capped FPS when I say typical 3D load. I'm also not including older titles that are easy to run and do not need full GPU boost frequency to give playable FPS.

 

[Glo.]

Navi34, like it's predecessor, may take a year to make a showing. At which point it might as well be RDNA3+ based, on N4. Likely monolithic <100mm2 die. But will post some insane P/W values.
Tho at that point it may have to contend with Arc "B3xxM" as well.

If it is 2024 release its not worth releasing it when we will have Strix Point APU, which will have Infinity Cache.

 

[BorisTheBlade82]

@Glo.
Has IC for Strix Point been mentioned by someone, anywhere, at all?
For comparison: The mostly disregarded rumour of Zen5+Zen4c/D has at least been mentioned in an internal AMD roadmap: https://www.tomshardware.com/news/amd-zen-5-cpus-apus-3nm-process-node

 

[FlameTail]

AMD is rumoured to feature a Xilinx based AI accelarator in their upcoming CPUs.

One thing i am curious about - is AI computation more bandwidth sensitive (like GPUs) or more latency sensitive (like CPUs) ?

 

[SpudLobby]

You did not at all read what I wrote, did you?

Your average consumer literally does not care about what processor is inside. It doesn't matter what you name them for that market of people, because those people won't recognise nor care about it either way. They don't read ANY of the numbers at all.

The number of people that read anything at all are a very small set of the consumer base. And they only make it to the i3/i5/i7/r3/r5/r7 names. not the rest of it. How many times have you seen people just call their CPU "an i7 processor" or something to that effect?

Why stop there? Why not further complicate it? The average consumer is also more likely to pick up e.g. Mendocino. I did read what you wrote — I don’t appreciate the silly condescension as if assuming your position a null hypothesis at the absolute limit. I’m happy to escalate this from here but frankly it seems unnecessary.
There *is* a tier of people that aren’t on Anandtech but have a basic interest in mind — you’re also underestimating how many laptops are sold via the holidays by gift, and some may actually want something fairly specific. I’ve watched it happen wrt youth & a gaming laptop.

 

[SpudLobby]

How does the iGPUs in Apple M chips comparw in raw performance to Radeon 660M/680M and potential Phoenix with RDNA 3 ?

Now, i understand that this is not an ideal comparison as Apple GPUs are not optimised for gaming ( thanks Metal API), which is why i said raw performance. That can be measured by benchmsrks.

I think 2.6 TFLOPS is an M1 8-core @ about 10W of GPU power.

 

[SpudLobby]

AMD's customers (ie: OEMs) are asking for this.

Probably the best take on this. I’ll accept that.