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 My advice to everyone: treat the old APU rumours as being out of date.

 

[MadRat]

Rather than consumer chips with big...little cores, I'd rather see groupings of cores where the first core (or two) in each core group is built with caches focused on the lowest possible latencies. It probably focuses on brute clock speed. It should use a different cache ratio than other cores, to concentrate on raw performance with less dependence on memory access. The other cores in the group should focus on memory throughput. The OS driver should balance the workload to the appropriate cores. Gaming probably should focus on performance. Media would probably focus on memory throughput.

Memory throughput on media probably looks much different than for gaming. A game might benefit from one fast core with interleaving memory access focused across many channels, where buffers catch most of the accesses and keep memory latency to a minimum. Media probably benefits from brute memory throughput with little benefit from buffering but much benefit from sustained access. So maybe your first core (that is built on lower latency / brute clock speeds) lays in a way to gain maximum benefit from quicker access to buffers, and access to memory usage as the priority. The other cores in the group lay in a way to maximize large chunks of memory through each channel. While the first core would have priority, its interuptions would be shorter.

I look at it like a focus through a cone. Look through the big end of the cone you see micro details. Look through it from the small end you see macro details.

 

[TESKATLIPOKA]

Not a difficult task as TPU already gives the 680M (Rembrandt GPU) as practically equal to the RX470 (1080P).
New Architecture, new process, some more clocks, RX570 should be the bare minimum (it depends of course also on power limits).

Are you sure? It's not as easy as you think.

TPU didn't test 680M, just wrote a BS relative performance in their GPU database.
Notebookcheck tested them, Link, Link, Link
I also added a 55W 6500M for comparison.

Time Spy Graphics
680M: 2449 (100%)
470 Desktop: 3624 (148%)
Sapphire NITRO+ RX 570 8 GB: 3958 (162%)
RX 6500M: 4434 (181%)

Fire Strike Standard Graphics
680M: 7279 (100%)
470 Desktop: 11869 (163%)
Sapphire NITRO+ RX 570 8 GB: 13827 (190%)
RX 6500M: 13980 (192%)

edit:
PHX would need 162/1.15=140 about 40% higher clockspeed to be on par with 570 Desktop. So there are 3 possibilities:
1.) It doesn't perform as 570 or not desktop variant
2.) It can really clock 40% higher, that's 3.05-3.35GHz depending on U or H variant.
3.) architectural improvement is more than 15%

 

[leoneazzurro]

Yeah, I was remembering wrong, the 680M was outperforming the 460, not the 470. In any case, there is a good analisys with actual FPS at computerbase (syntetic benchmarks are often misleading):

AMD Radeon 680M: Die Ryzen-6000-iGPU im Test: Benchmarks der Radeon 680M in Spielen - ComputerBase

With the architecture/process advantages Phoenix should have, a 470/570 may not be too far, especially considering that in these benchmarks the CPU used with the 570 is way more powerful and that matters at the settings used.

 

[TESKATLIPOKA]

Yeah, I was remembering wrong, the 680M was outperforming the 460, not the 470. In any case, there is a good analysis with actual FPS at computerbase (syntetic benchmarks are often misleading):

AMD Radeon 680M: Die Ryzen-6000-iGPU im Test: Benchmarks der Radeon 680M in Spielen - ComputerBase

With the architecture/process advantages Phoenix should have, a 470/570 may not be too far, especially considering that in these benchmarks the CPU used with the 570 is way more powerful and that matters at the settings used.

Sorry, you were actually right. PHX will have an easy time to not just equal 570, but actually be faster, at least in Cyberpunk 2077.


P.S. Never buy a 4GB Vram GPU. I myself have GTX 1650M, but bought It 3-4 years ago, so I can be forgiven.

 

[SpudLobby]

You should read what you quote, because 6600U was mentioned by me.
I also mention how 4CU PHX2 should perform comparably against 6600U If It clocks 30% higher, which is 2.5GHz.

You meant 6 and 12 CU configurations.
They won't go down to 4CUs If IGP has physically 12CUs. 6 and 12 CU is not a bad option and performance difference is ~50%, the worse thing is that full IGP option doesn't exist for 6 cores.

Sure, you said “will have a weaker IGPU than Rembrandt” which paired with 12CU’s seems odd. I’d agree it’ll be weaker than a 12CU Rembrandt sku, sure. Maybe a misread on my part, but do tell me about reading something I fairly directly replied to.

 

[TESKATLIPOKA]

Sure, you said “will have a weaker IGPU than Rembrandt” which paired with 12CU’s seems odd. I’d agree it’ll be weaker than a 12CU Rembrandt sku, sure. Maybe a misread on my part, but do tell me about reading something I fairly directly replied to.

Ok, I should have replied: Not to focus on that single sentence in my post or just point out I was also comparing It to 6600U not just the full configuration(PHX2 vs 6800U). My mistake.

P.S. Edited my original post, to not be so confrontational.

 

[LightningZ71]

Yeah, I was remembering wrong, the 680M was outperforming the 460, not the 470. In any case, there is a good analisys with actual FPS at computerbase (syntetic benchmarks are often misleading):

AMD Radeon 680M: Die Ryzen-6000-iGPU im Test: Benchmarks der Radeon 680M in Spielen - ComputerBase

With the architecture/process advantages Phoenix should have, a 470/570 may not be too far, especially considering that in these benchmarks the CPU used with the 570 is way more powerful and that matters at the settings used.

That breakdown from ComputerBase is pretty interesting, but it does still demonstrate something quite profound. Even with all the grunt that the 680M has, it still often gets left behind by the 3050 mobile at 60W, sometimes by more than double the performance. Looking at laptops on the market, prices for 3050 and 3050ti laptops are often cheaper than any of the 6800U and 6800H+ laptops without dGPUs. The only place that the 3050 falls behind the 680m is in the 3dmark port royal benchmark which has a pretty big performance drop off when limited to only 4GB of VRAM. That's going to continue to be an issue until AMD and laptop makers develop some sort of price sanity. There's no reason that a CPU only laptop should cost more than one with a dGPU one from a total parts cost perspective. The board is simpler, the cooling is simpler, there's no need for dedicated VRAM or a second chip.

 

[jpiniero]

The board is simpler, the cooling is simpler, there's no need for dedicated VRAM or a second chip.

AMD probably charges a lot more for Rembrandt compared to even Cezanne... between that and DDR5, it's easy to see where the 3050 on a cheap Samsung node is affordable in comparison.

 

[Glo.]

3.) architectural improvement is more than 15%

Yes. It lacks full dual issue, and remains only on opportunistic dual issue, while maintaining larger low level caches. So It should be less memory starved than RMB.

Time Spy score for PHX should be around 4000 pts. Around - doesn't mean above, necessarily.

 

[TESKATLIPOKA]

That breakdown from ComputerBase is pretty interesting, but it does still demonstrate something quite profound. Even with all the grunt that the 680M has, it still often gets left behind by the 3050 mobile at 60W, sometimes by more than double the performance. Looking at laptops on the market, prices for 3050 and 3050ti laptops are often cheaper than any of the 6800U and 6800H+ laptops without dGPUs. The only place that the 3050 falls behind the 680m is in the 3dmark port royal benchmark which has a pretty big performance drop off when limited to only 4GB of VRAM. That's going to continue to be an issue until AMD and laptop makers develop some sort of price sanity. There's no reason that a CPU only laptop should cost more than one with a dGPU one from a total parts cost perspective. The board is simpler, the cooling is simpler, there's no need for dedicated VRAM or a second chip.

I don't think It's necessarily AMD's fault, don't even think Rembrandt is so much costlier than Cezanne.
Rembrandt is put into more premium laptops + DDR5, that's the problem.
I can buy ThinkBook 13s G4 with 6600U, 8GB, 512GB SSD for 1029 €.
I can buy IdeaPad Gaming 3 with 5600H, 8GB, 512GB SSD, RTX 3050 4GB 85W for 794,90 €.

P.S I hope you all are preparing the money as I do, because PHX won't be cheap either.

 

[Tigerick]

To act as a memory controller for a single CPU, you'd only need CXLmem, which is part of the original spec, so that is perfectly plausible. But saving on pins at the cost of power, bandwidth, and latency is a bad tradeoff for mobile.

CXL memory expansion is more likely to be first employed in the datacenter to support large in-memory databases etc. that benefit from tons of capacity, but can afford some performance hit. Similar to how Optane is used today. Eventually, CXL memory pooling (part of the 3.0 spec) + switching (from the 2.0 spec) will allow enormous (10s to 100s of TB) memory pools shared by an entire rack, if not more. You'd have dedicated blades solely for DRAM shared by blades of CPUs, accelerators, NICs, etc.

On the desktop, CXL memory expansion might be handy for more memory capacity without taking the performance hit to near memory demanded by 2 DIMMs per channel, and/or you could interleave the two for even more total bandwidth. But the software side might be tricky. Anyway, probably not a discussion relevant to these kind of chips, at least for the foreseeable future.

Thanks for detailed explanation of CXL.mem, any ideas how much power required for PCIe compared to parallel memory controller?

 

[Tigerick]

I don't think It's necessarily AMD's fault, don't even think Rembrandt is so much costlier than Cezanne.
Rembrandt is put into more premium laptops + DDR5, that's the problem.
I can buy ThinkBook 13s G4 with 6600U, 8GB, 512GB SSD for 1029 €.
I can buy IdeaPad Gaming 3 with 5600H, 8GB, 512GB SSD, RTX 3050 4GB 85W for 794,90 €.

P.S I hope you all are preparing the money as I do, because PHX won't be cheap either.

Yeah, not going to be cheap, at least initially. AMD targets PP notebook as MBA alternative with all the integration features into SoC. With N4 process, AMD has to create two dies of PP to cover full range of lineup. The most completed features might be available on Ryzen 9 only.

 

[Tigerick]

Regarding these rumored specs, obviously an unknown leaker should be taken with plenty of skepticism, but this sort of split would make a lot of sense. If nothing else, it would be very beneficial for AMD to add a cost-reduced part for the mainstream market. Your typical office or home PC user wouldn't really see a benefit from 8 big cores or a class-leading GPU, but they would benefit from better battery life and lower power envelopes enabled by big.LITTLE/hybrid/whatever you want to call it, and of course the cost savings of such an arrangement. Then they still have the beefed up config for light gaming, content creation, etc. Real or not, sounds like a great set of configs.

The only part that confuses me is that neither goes below 15W. I'm not sure why PHX2 wouldn't be able to scale down to ~10W, and that would let them fit it in fanless laptops. That's a pretty significant gap in AMD's lineup right now, but they should have everything they need to fill it.

I am afraid AMD has gone chiplets design on Phoenix Point SoC, thus comes with power penalty. The specs of PHX2 is similar to Mendocino with 4 Zen 2 cores and 4MB L3 cache which is made by N6 process with 100mm2 die sizeand 15W TDP. If the leaks are legit, the PHX2 is really only for future mainstream notebook..

Amend: PHX2 should come with 2 Zen4 and 4 Zen4c so above statement is not 100% correct

 

[TESKATLIPOKA]

I am afraid AMD has gone chiplets design on Phoenix Point SoC, thus comes with power penalty. The specs of PHX2 is similar to Mendocino with 4 Zen 2 cores and 4MB L3 cache which is made by N6 process with 100mm2 die sizeand 15W TDP. If the leaks are legit, the PHX2 is really only for future mainstream notebook..

What does chiplet design on PHX or PHX2 mean for you? CPU on one die and IO+IGP on the other or different combination? I personally don't know how PHX2 could combine 2 cores with significantly different characteristics on a single die, If we talk about two dies, then that's not a problem.

Actually, AMD really needs something smaller for mainstream, this PHX2 should be pretty small.
Although they released Mendocino, but 4 Zen2 cores and 2CU RDNA 2 IGP is simply weak compared to Rembrandt or PHX, although enough for light use.
This PHX2 has 1.5x more cores, significantly higher IPC, CPU clocks won't be lower, much better IGP. This could be an interesting chip for ultraportables, If It doesn't come too late.

 

[Tigerick]

What does chiplet design on PHX or PHX2 mean for you? CPU on one die and IO+IGP on the other or different combination? I personally don't know how PHX2 could combine 2 cores with significantly different characteristics on a single die, If we talk about two dies, then that's not a problem.

Actually, AMD really needs something smaller for mainstream, this PHX2 should be pretty small.
Although they released Mendocino, but 4 Zen2 cores and 2CU RDNA 2 IGP is simply weak compared to Rembrandt or PHX, although enough for light use.
This PHX2 has 1.5x more cores, significantly higher IPC, CPU clocks won't be lower, much better IGP. This could be an interesting chip for ultraportables, If It doesn't come too late.

We will know about PP designs next month, I am refrain from talking about it. I got two downvote and one upvote out of it.

I thought the leaks said PHX2 comes with 4 ZEN 4C only? Anyhow, AMD still need to clear Rembrandt-R and Barcelo-R before let the PHX2 hits the market. So for the moment PHX should be the one we focus on.

 

[TESKATLIPOKA]

We will know about PP designs next month, I am refrain from talking about it. I got two downvote and one upvote out of it.

I thought the leaks said PHX2 comes with 4 ZEN 4C only? Anyhow, AMD still need to clear Rembrandt-R and Barcelo-R before let the PHX2 hits the market. So for the moment PHX should be the one we should focus on.

Ignore the downvoting. What is important is a good discussion and a lot of likes.
From that leak PHX2 should be 2 Zen 4 cores and 4 Zen 4c cores.

 

[Tigerick]

Ignore the downvoting. What is important is a good discussion and a lot of likes.
From that leak PHX2 should be 2 Zen 4 cores and 4 Zen 4c cores.

Oh ok, PHX2 actually comes with 2P + 4E hybrid design, interesting...I mislook the specs, the L3 cache is not shared among all cores? Kinda wasting...

 

[Tigerick]

The leaker seems to know something, here is the Barcelo-R that I just mentioned. One more SKU with some changes on graphics core and speed.

 

[Exist50]

Mendocino is the current gen part for fanless designs afaik, and that'll have it's own successor.

Mendocino is a budget chip. But AMD could really use a premium (or at least mainstream...) low power chip to compete with Apple and eventually Lunar Lake.

 

[Glo.]

https://twitter.com/x/status/1598350848622469121

Ryzen 9 7940HS 8c 12 CU.
Ryzen 9 7840HS 8c 12 CU.
Ryzen 7 7740HS 8c 12 CU.
Ryzen 5 7640HS 6c 6 CU.
Ryzen 7 7740U 8c 12 CU
Ryzen 5 7640U 6c 6 CU
Max boost >5GHz Max base >3.5GHz.
ES around 2600 MHz max iGPU.
PHX2 low 5 & 3 later.

2600 MHz? 200 MHz more than RMB, and that big of a difference in performance? 680M - RX 570 is around 60% perf. increase.

 

[Exist50]

Thanks for detailed explanation of CXL.mem, any ideas how much power required for PCIe compared to parallel memory controller?

It would certainly be worse, normalized for bandwidth, but I don't have any exact numbers, sorry.

 

[Tigerick]

https://twitter.com/x/status/1598350848622469121

2600 MHz? 200 MHz more than RMB, and that big of a difference in performance? 680M - RX 570 is around 60% perf. increase.

RMB iGPU should be having 2.2GHz clock speed, so it is 400MHz faster, hmm..

 

[Glo.]

RMB iGPU should be having 2.2GHz clock speed, so it is 400MHz faster, hmm..

RMB has 2400 MHz iGPU clock.

P.S. I think now we know what variant will have RX 570 performance. And it is 35W, HS variant.

I'd say that is pretty darn good.

P.S. I wonder how high will it clock with higher TDPs, like 55W-65W...

 

[TESKATLIPOKA]

For ES 2.6GHz is very good. Actually, that is comparable to the highest clockspeed for RDNA2 based GPUs If I exclude RX 6700XT.
I don't see how 2.6-2.8GHz IGP can perform as RX 570.
What's the secret sauce? Frequency will provide ~15%, architecture another 15%, that's still 1*1.15*1.15= 1.32 or 32% better performance. So from where comes the rest?

Those new CPU(APU) names are pure insanity. I so want to slap someone.

 

[TESKATLIPOKA]

Do you know what's the difference between Rembrandt and PHX?
Try guessing.
R9 7935HS vs R9 7940HS
R9 7835HS vs R9 7840HS
R7 7735HS vs R7 7740HS
R5 7635HS vs R5 7640HS

Ryzen 7 7735U vs Ryzen 7 7740U
Ryzen 5 7635U vs Ryzen 5 7640U

I hereby declare AMD as a king of rebranding, because having 3 generations in one is nothing short of amazing marketing genius.
I think Intel marketing is having a strategic marketing meeting, to see how fast they can use the same scheme.