It's trickier than that: running a many-threaded workload, you can't run all the cores at their top clocks anyway, because of power and cooling. The Zen 4C cores can (at least on Cinebench) add almost as much oomph as Zen 4 cores would, but may not be able to hit the higher frequencies used in single-thread workloads at all. It can all work out well on your workload (which is the point the slide is trying to make) but it doesn't mean that the Zen 4C cores are able to hit the same speeds as Zen 4 cores.
If anything, this will further cement AMD is the value proposition for mobile, as U-series chips are seldom used in budget Intel laptops over the less expensive and power hungry P-series chips, both which scale down terribly to 15-25w from their TDPup of 35-45w compared to Zen4.
So while no cheap laptop (and even many mid-range models upward of $1000) seem to allow modern mobile CPU's to consistently maintain a 25+ watt power profile, at that ceiling, AMD performs better. Intel's current architecture seems to be a spiritual reincarnation of Netburst where the only way to edge out an extra 10-20% performance and stay competitive is to double (or triple) power consumption. It's ridiculous.
The M2 uses 20W+ when running Cinebench 23 multithreaded, according to the Notebookcheck test. The M2 gets a score of 8500+, while a 7640U gets a score of 10500+. Although I haven't seen a test of the 7640U at 20W, it still seems to me like it will beat the M2. Therefore, AMD is already more efficient.
And sure, Apple is now at 3nm and so might have caught up with AMD's 4nm CPUs, but I wouldn't bet on it.
" Apple Silicon doesn't even require cooling. When will AMD actually make efficient mobile chips that can match Apple Silicon?"
right after intel does the same ? as hard as you try lemur, apple isnt the best out there... for anything.. regardless of the BS and misinformation you keep posting... either post links to what you claim, or stop posting...
Please don't try to act as a gatekeeper for everyone else's ability to make comments you may disagree with. Instead of trying to control the actions of others that you obviously will not be able to do, consider controlling your own reaction to disagreeable information. You will likely feel a lot happier if you take that approach to life.
All ARM processors does not require cooling. AMD is x86 which is far more complex than ARM. You can even put Apple ARM cpu to a phone but never an x86. Comparison is far off.
Add: x86 CPUs are generally more powerful than ARM CPUs, especially for demanding tasks like gaming, video editing, and scientific computing. Let's try your Apple Silicon match with AMD Ryzen on COD if you can win or can even play at all.
Everything above the Macbook Air with the base M2 requires cooling. Even the base M3 Macbook Pro has a heatsink and fan to sustain the performance, where the Macbook Air loses 30% over time when it's warm.
I think someone could probably take these "C" cores and make a take on the fanless Macbook Air though, I'd like to see it
Lot of speculation here and slamming AMD, without knowing full details. Zen 4 is designed for 5+ GHz clockspeeds, as shown by the up to 5.70GHz it reaches. How do we really know Zen 4c can't clock up much past 3GHz? Because of the server clockspeeds? That makes no sense - Zen 4, in its server versions, also barely gets above 4GHz in most parts - a full 1.5GHz below desktop chips. My point is: Zen 4c could very well easily reach 4+ GHz ranges, making the all-turbo clockspeeds on Phoenix 2 no different than Phoenix 1 processors. Someone should be able to verify this, somewhat easily, on the two different ROG Ally configurations.
Actually, AMD said in some interviews that the optimizations they made to the process technology to make this dense & efficient came at the expense of clock speeds. So we know for sure that clocks are limited for the "c" cores. Clocking these cores higher will put them way past their efficiency goals.
But, I personally don't mind that, as long as it brings performance and efficiency to the chip, in fact, I don't mind having a desktop Ryzen with 8xZen4 and 16xZen4c chiplets, but I guess AMD doesn't see this happening now, maybe in Zen5 or Zen6 with even more cores.
At least there it seems like the zen4c cores were capped at 3.5 Ghz. It might well be possible they can go a bit higher (the voltage is already quite a bit higher there than zen 4 for the same clock, but not sky high and certainly lower than the maximum voltage of zen 4). (Also keep in mind zen4c will use less power if it's running at the same voltage and clock as zen 4, so not sure at which clock exactly it will be less efficient.) Based on the voltage / frequency graph though I'd guess there's no way these cores can go past 4 Ghz.
@ Xajel: Oh no, I definitely agree that these are designed to clock lower - I was just wondering how MUCH lower. Keeping in mind that both the Ryzen 5 and 3 have lower top turbos anyway, I was just curious if Zen 4c could keep up in that environment. Zen 4c should definitely not be in a the Ryzen 7 or 9 series, as people expect all-core turbos there well above 4GHz or 5GHz.
@mczak - Thank you for that link! So yeah, it does appear Zen 4c tops out at 3.5GHz then. And it does confirm that its IPC is identical to Zen 4. However, what I would really like to see is a ROG Ally chart, one with Z1 Extreme and another with Z1, running a benchmark capped at 6 cores (12 threads), running the same TDP levels. That would show how a typical multi-core run is between Zen 4/4c. If a lot of the Zen 4 cores on Z1 Extreme clock down to 3.5GHz or less anyway during a benchmark, then Zen 4c isn't holding back anything.
The die size comparison fails to mention another cut (apart from cpu and gpu changes, and missing AI accelerator) - Phoenix 2 has only 14 instead of 20 pcie lanes. Shouldn't really matter in the mobile devices this will appear in, although if it ever comes as AM5 version, it very well might.
Judging by the original specs of the 7540U and 7440U, seeing how they both fit within the capabilities of this new chip, I'm guessing that this was a calculated play by AMD to wait and see how the specifics played out - if there was early demand, they'd launch full Zen4 versions, but if there wasn't, they'd hold off until Phoenix 2 was ready. And clearly there was early demand for the 7540U (as demonstrated by laptops with this chip being on the market as we speak), while as far as I can tell no product has ever reached the market with the 7430U, allowing AMD to keep that name for the Phoenix 2 chip only.
As for these chips, I agree that it's pretty bad to not disclose 4c core clockspeeds, but as long as these chips are kept in low-power designs it won't matter much anyhow - all-core clock speeds in 15W laptops tend to be very low, not to mention if the iGPU is active (easily bringing many U-series chips into the 1GHz range for the CPU). So, if Zen 4c is actually more efficient per clock, this would actually result in a net performance increase in these scenarios. This also somewhat explains the decision to put this chip in the cheaper ROG Ally - while at first I thought it was rather silly due to the 4c iGPU, it makes a lot of sense for a low power casual/emulation handheld. It's just kind of a shame that the price differential to the Z1 Extreme version is relatively small.
All cores by default hit differing clockspeeds, only final performance matters. If the IPC is identical, cache is identical, definition of base and turbo clocks hasn't changed, and MOST importantly performance is BETTER - what exactly is the issue?
How is AMD doing a disservice to consumers by making a better product?
It's not about better or worse, it's about predictability and transparency. This chip is indeed most likely a tad better in low power MT scenarios, but it might be worse in others. Without transparency on the behaviour of various cores, and especially without clearly disclosing how many of each core the SKUs have, it becomes difficult to predict as their relative scaling across tasks will be non-linear. Is this a huge issue? No, especially not for low end parts. But it is nonetheless valid criticism and something AMD should do better on.
If you put 2x Zen 4 and 4x Zen 4c on marketing material doesn't that end up making things more unclear and opaque? How do you teach consumers that Zen 4c is not the same as other efficiency cores?
Isn't just saying 6 cores better for the consumer? I would be surprised if there is much difference between Zen 4 and Zen 4c clocks for the non-prefered cores on a laptop anyway.
The specifications on AMD's product page tells you the configuration if you are interested.
Because checking reviews isn't always possible, especially for low end products that rarely get reviewed at all. Reviews of other products then provide data for rough estimates of performance, assuming accurate specs are available - but without accurate specs, those guesses become extremely unreliable. And yes, these cores will outperform Intel's e cores by quite a bit - but they'll also behave differently from regular Zen 4 cores. And because there are both pure Zen 4 and Zen 4+4c designs in the same product stack, internal comparisons are the most important, hence a distinction is necessary. Saying '6c (2+4, 12t)' really isn't very hard, nor is listing peak boost clocks for both types of cores. Intel manages, and they can barely make a cpu that doesn't cook itself these days, so why should amd be unable to do this? Have they run out of rows in their table budget for the year?
>and MOST importantly performance is BETTER Assuming that's true, that also means that the performance of the Phoenix 1 variants is worse, so that's enough reason to be transparent just there.
Exactly. Different behaviour needs to be highlighted. Doesn't matter if it's faster or slower or a mix of both, the point is transparency and accurate information.
Say what you will about Nvidia (and I have many negative things to say), but CUDA is performant and consistent across the entire lineup, especially post RTX 2060.
Meanwhile, AMD has split the server/client GPU arch, is asking if the NPUs are even worth supporting on linux, and now is excluding the NPUs from some new SoCs... This is not how to drive software adoption.
Intel SoC/dGPU support is sneaking into all sorts of "local AI" projects, even with Intel's massive current hardware disadvantage. ROCm is here and there, but I haven't seen a single implementation of Ryzen AI outside the quick AMD demo.
What kind of speeds can an existing 7540u hit on 15w for the CineBench MT test? According to their testing the new chip is faster at that point. That would have to mean either the existing chip is stuck pretty close to it's base 3.2ghz speeds or the 4c cores can still turbo up a fair bit beyond what we've seen in server parts. I do agree that it would be nice if AMD would come out and provide a more details boost speed than just the one fastest single core.
I wish AMD had made a version of Phoenix for PC gaming handhelds with 6-8x Zen4c cores and a wider 8-12CU GPU, with perhaps some 8-16MB Infinity Cache.
Size might have been close to the original Phoenix but performance at 15-20W could be so much better.
I think you're... eh, let's say optimistic about AMD's budget for taping out new chip designs. Yes, they did make the Steam Deck chip, but only after Valve pre-ordered likely hundreds of thousands of them. Taping out a new design costs millions of dollars at the minimum. Handheld gaming is still a niche market, so it would take someone willing to take on that risk for AMD to do something like this, and the only real candidate there would be an SD sequel, which Valve has confirmed isn't coming for a while yet. I wouldn't be surprised to see a 6-8c Zen *c +large-ish GPU design for that when it arrives though, but I wouldn't expect that until... late next year, or 2025?
Can a 7840U/HS run accelerated torch, tensorflow, onnx etc via rocm? On linux only? or windows also? IS the performance any good? (I don't think these have decidated vram)
Is the AI a separate NPU in addition to the GPU or are they just using the GPU?
The NPU in Phoenix is a separate functional block. It is not reusing the GPU.
As for what libraries are supported under what OSes, I'm afraid I don't have that information handy. I know that Ryzen AI is accessible via ONNX, but I don't have any further details than that.
ERROR : "The Chips: Ryzen 5 7545U and Ryzen 3 7440U" section, first table "AMD Ryzen 7040U series"
The 7545U with 2+4c Ryzen chips must have a 12MB L3 cache, not 16. That makes sense. The 7640U has 16MB cache. The 7545 changes 4 cores to 'c' cores so it must save at least 4MB of cache. There is no way that the 7545U and 7640U have the same cache size. The whole point of 'c' cores is to save on cache, and if both cores are 16MB cache, that would be stupid.
Hi. The specs in that table are correct. The L3 cache is shared over the entire CCX; it's not a per-core arrangement. Even though AMD has synthesized a CCX with just 6 CPU cores, they put a full 16MB of L3 cache in there.
More broadly speaking, Zen 4c in servers exists in part to save on L2 cache. But this isn't true for the mobile parts. The server parts had 2MB of L2 per core with Zen 4, and 1MB of L2 with Zen 4c. But the mobile parts always had 1MB of L2 per core.
As odd as it may seem, in this case the answer is no. I do mean L2 cache.
Server Zen4c parts have less private L2 cache per core; it's part of how they managed to shave off 35% of the area of a CPU core.
As for L3 cache, AMD's chiplet CCDs have the same 64MB of shared L3 cache each, regardless of the core type used. It just happens to be that with 16 cores per Zen 4c CCD, you have fewer CCDs altogether than with Genoa (8x16 instead of 12x8). Which is moot for mobile, because the mobile CCX designs all are built with 16MB of L3.
Monolithic APU Tape-outs are Bespoke affairs so can not really be compared to CCD/Chiplet based processors from AMD. And the Die Shots of some Zen-4/Zen-4C based APU designs all show the cores arranged around the same Ring Bus like topology on that monolithic APU die so the L3 is maybe just another stop along that Ring Bus(?) there and all the cores since Zen-3 have a unified L3 cache pool.
And even Ryzen 4000 series APUs that had 2, 4 core Zen-2 CCX Units had no need for any IFOP(Infinity Fabric Off Package/Off Die) connectivity as the 2 CCX Units where on the same Monolithic slab of APU silicon so the latency was different and so was the memory access latency different compared to Chiplet based processors that had to go off die to access the memory controller on Zen-2/Later chiplet based processors where the I/O die was separate from the CCD DIE/DIEs.
I think it's good this is only going to be done on the R3 and R5. Those two get to see some of the worst and lowest cost cooling solutions anyone could come up with. Literal penny pinching stuff.
So AMD will keep one Zen4 core at least in each chip to skirt the line of dishonest advertising when it comes to maximum clocks (See ? See ? That core CAN hit the clockspeed so all is fine) while not specifying the maximum clock speed for most of the chip ... I expected better ....
Are there any laptops announced with this chip? Based on previous releases, it will take ages before we see it available in retail. I still have not seen any Zen 4 laptops at Costco or Best buy though online availability and deals have been good in past month or so. Big Achilles heel for AMD is laptop is the least important part of their portfolio. That is surprising considering laptops sell way more than desktop.
Seems great for Steam Deck and competitors, they're cooling and power bound anyways, so they're already operating where this would be most efficient and doesn't need to clock higher. You'd save power and die size both to give more to the IGP.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
43 Comments
Back to Article
TristanSDX - Thursday, November 2, 2023 - link
as for Zen4c clocks, provided CB graph shows that difference is pretty small, maybe 7-9% slower than HP corestwotwotwo - Sunday, November 5, 2023 - link
It's trickier than that: running a many-threaded workload, you can't run all the cores at their top clocks anyway, because of power and cooling. The Zen 4C cores can (at least on Cinebench) add almost as much oomph as Zen 4 cores would, but may not be able to hit the higher frequencies used in single-thread workloads at all. It can all work out well on your workload (which is the point the slide is trying to make) but it doesn't mean that the Zen 4C cores are able to hit the same speeds as Zen 4 cores.A5 - Thursday, November 2, 2023 - link
As long as OEMs insist on putting terrible cooling in their laptops, this is a smart move by AMD to help their margins.Samus - Friday, November 3, 2023 - link
If anything, this will further cement AMD is the value proposition for mobile, as U-series chips are seldom used in budget Intel laptops over the less expensive and power hungry P-series chips, both which scale down terribly to 15-25w from their TDPup of 35-45w compared to Zen4.So while no cheap laptop (and even many mid-range models upward of $1000) seem to allow modern mobile CPU's to consistently maintain a 25+ watt power profile, at that ceiling, AMD performs better. Intel's current architecture seems to be a spiritual reincarnation of Netburst where the only way to edge out an extra 10-20% performance and stay competitive is to double (or triple) power consumption. It's ridiculous.
lemurbutton - Friday, November 3, 2023 - link
Apple Silicon doesn't even require cooling. When will AMD actually make efficient mobile chips that can match Apple Silicon?TheinsanegamerN - Friday, November 3, 2023 - link
Why do macbook pros, mac minis, mac studios, and mac pros have cooling then?ET - Saturday, November 4, 2023 - link
The M2 uses 20W+ when running Cinebench 23 multithreaded, according to the Notebookcheck test. The M2 gets a score of 8500+, while a 7640U gets a score of 10500+. Although I haven't seen a test of the 7640U at 20W, it still seems to me like it will beat the M2. Therefore, AMD is already more efficient.And sure, Apple is now at 3nm and so might have caught up with AMD's 4nm CPUs, but I wouldn't bet on it.
Qasar - Sunday, November 5, 2023 - link
" Apple Silicon doesn't even require cooling. When will AMD actually make efficient mobile chips that can match Apple Silicon?"right after intel does the same ? as hard as you try lemur, apple isnt the best out there... for anything.. regardless of the BS and misinformation you keep posting... either post links to what you claim, or stop posting...
PeachNCream - Tuesday, November 7, 2023 - link
Please don't try to act as a gatekeeper for everyone else's ability to make comments you may disagree with. Instead of trying to control the actions of others that you obviously will not be able to do, consider controlling your own reaction to disagreeable information. You will likely feel a lot happier if you take that approach to life.Qasar - Thursday, November 9, 2023 - link
sure Peachandcream, right after you do. which i have seen you post very similar posts...practice what you preach
pogsnet - Wednesday, November 22, 2023 - link
All ARM processors does not require cooling. AMD is x86 which is far more complex than ARM. You can even put Apple ARM cpu to a phone but never an x86. Comparison is far off.pogsnet - Wednesday, November 22, 2023 - link
Add:x86 CPUs are generally more powerful than ARM CPUs, especially for demanding tasks like gaming, video editing, and scientific computing. Let's try your Apple Silicon match with AMD Ryzen on COD if you can win or can even play at all.
tipoo - Monday, November 6, 2023 - link
Everything above the Macbook Air with the base M2 requires cooling. Even the base M3 Macbook Pro has a heatsink and fan to sustain the performance, where the Macbook Air loses 30% over time when it's warm.I think someone could probably take these "C" cores and make a take on the fanless Macbook Air though, I'd like to see it
NextGen_Gamer - Thursday, November 2, 2023 - link
Lot of speculation here and slamming AMD, without knowing full details. Zen 4 is designed for 5+ GHz clockspeeds, as shown by the up to 5.70GHz it reaches. How do we really know Zen 4c can't clock up much past 3GHz? Because of the server clockspeeds? That makes no sense - Zen 4, in its server versions, also barely gets above 4GHz in most parts - a full 1.5GHz below desktop chips. My point is: Zen 4c could very well easily reach 4+ GHz ranges, making the all-turbo clockspeeds on Phoenix 2 no different than Phoenix 1 processors. Someone should be able to verify this, somewhat easily, on the two different ROG Ally configurations.Xajel - Thursday, November 2, 2023 - link
Actually, AMD said in some interviews that the optimizations they made to the process technology to make this dense & efficient came at the expense of clock speeds. So we know for sure that clocks are limited for the "c" cores. Clocking these cores higher will put them way past their efficiency goals.But, I personally don't mind that, as long as it brings performance and efficiency to the chip, in fact, I don't mind having a desktop Ryzen with 8xZen4 and 16xZen4c chiplets, but I guess AMD doesn't see this happening now, maybe in Zen5 or Zen6 with even more cores.
mczak - Thursday, November 2, 2023 - link
This was already done quite a while ago:https://www.hardwaretimes.com/amds-phoenix-2-hybri...
At least there it seems like the zen4c cores were capped at 3.5 Ghz. It might well be possible they can go a bit higher (the voltage is already quite a bit higher there than zen 4 for the same clock, but not sky high and certainly lower than the maximum voltage of zen 4). (Also keep in mind zen4c will use less power if it's running at the same voltage and clock as zen 4, so not sure at which clock exactly it will be less efficient.) Based on the voltage / frequency graph though I'd guess there's no way these cores can go past 4 Ghz.
NextGen_Gamer - Thursday, November 2, 2023 - link
@ Xajel: Oh no, I definitely agree that these are designed to clock lower - I was just wondering how MUCH lower. Keeping in mind that both the Ryzen 5 and 3 have lower top turbos anyway, I was just curious if Zen 4c could keep up in that environment. Zen 4c should definitely not be in a the Ryzen 7 or 9 series, as people expect all-core turbos there well above 4GHz or 5GHz.@mczak - Thank you for that link! So yeah, it does appear Zen 4c tops out at 3.5GHz then. And it does confirm that its IPC is identical to Zen 4. However, what I would really like to see is a ROG Ally chart, one with Z1 Extreme and another with Z1, running a benchmark capped at 6 cores (12 threads), running the same TDP levels. That would show how a typical multi-core run is between Zen 4/4c. If a lot of the Zen 4 cores on Z1 Extreme clock down to 3.5GHz or less anyway during a benchmark, then Zen 4c isn't holding back anything.
mczak - Thursday, November 2, 2023 - link
The die size comparison fails to mention another cut (apart from cpu and gpu changes, and missing AI accelerator) - Phoenix 2 has only 14 instead of 20 pcie lanes. Shouldn't really matter in the mobile devices this will appear in, although if it ever comes as AM5 version, it very well might.Valantar - Thursday, November 2, 2023 - link
Judging by the original specs of the 7540U and 7440U, seeing how they both fit within the capabilities of this new chip, I'm guessing that this was a calculated play by AMD to wait and see how the specifics played out - if there was early demand, they'd launch full Zen4 versions, but if there wasn't, they'd hold off until Phoenix 2 was ready. And clearly there was early demand for the 7540U (as demonstrated by laptops with this chip being on the market as we speak), while as far as I can tell no product has ever reached the market with the 7430U, allowing AMD to keep that name for the Phoenix 2 chip only.As for these chips, I agree that it's pretty bad to not disclose 4c core clockspeeds, but as long as these chips are kept in low-power designs it won't matter much anyhow - all-core clock speeds in 15W laptops tend to be very low, not to mention if the iGPU is active (easily bringing many U-series chips into the 1GHz range for the CPU). So, if Zen 4c is actually more efficient per clock, this would actually result in a net performance increase in these scenarios. This also somewhat explains the decision to put this chip in the cheaper ROG Ally - while at first I thought it was rather silly due to the 4c iGPU, it makes a lot of sense for a low power casual/emulation handheld. It's just kind of a shame that the price differential to the Z1 Extreme version is relatively small.
dhruvdh - Thursday, November 2, 2023 - link
Anandtech's position on this makes no sense.All cores by default hit differing clockspeeds, only final performance matters. If the IPC is identical, cache is identical, definition of base and turbo clocks hasn't changed, and MOST importantly performance is BETTER - what exactly is the issue?
How is AMD doing a disservice to consumers by making a better product?
Valantar - Thursday, November 2, 2023 - link
It's not about better or worse, it's about predictability and transparency. This chip is indeed most likely a tad better in low power MT scenarios, but it might be worse in others. Without transparency on the behaviour of various cores, and especially without clearly disclosing how many of each core the SKUs have, it becomes difficult to predict as their relative scaling across tasks will be non-linear. Is this a huge issue? No, especially not for low end parts. But it is nonetheless valid criticism and something AMD should do better on.dhruvdh - Thursday, November 2, 2023 - link
Why predict when you can just check reviews?If you put 2x Zen 4 and 4x Zen 4c on marketing material doesn't that end up making things more unclear and opaque? How do you teach consumers that Zen 4c is not the same as other efficiency cores?
Isn't just saying 6 cores better for the consumer? I would be surprised if there is much difference between Zen 4 and Zen 4c clocks for the non-prefered cores on a laptop anyway.
The specifications on AMD's product page tells you the configuration if you are interested.
Valantar - Friday, November 3, 2023 - link
Because checking reviews isn't always possible, especially for low end products that rarely get reviewed at all. Reviews of other products then provide data for rough estimates of performance, assuming accurate specs are available - but without accurate specs, those guesses become extremely unreliable. And yes, these cores will outperform Intel's e cores by quite a bit - but they'll also behave differently from regular Zen 4 cores. And because there are both pure Zen 4 and Zen 4+4c designs in the same product stack, internal comparisons are the most important, hence a distinction is necessary. Saying '6c (2+4, 12t)' really isn't very hard, nor is listing peak boost clocks for both types of cores. Intel manages, and they can barely make a cpu that doesn't cook itself these days, so why should amd be unable to do this? Have they run out of rows in their table budget for the year?Dolda2000 - Thursday, November 2, 2023 - link
>and MOST importantly performance is BETTERAssuming that's true, that also means that the performance of the Phoenix 1 variants is worse, so that's enough reason to be transparent just there.
Valantar - Friday, November 3, 2023 - link
Exactly. Different behaviour needs to be highlighted. Doesn't matter if it's faster or slower or a mix of both, the point is transparency and accurate information.brucethemoose - Thursday, November 2, 2023 - link
> and tossing out the Ryzen AI NPUSay what you will about Nvidia (and I have many negative things to say), but CUDA is performant and consistent across the entire lineup, especially post RTX 2060.
Meanwhile, AMD has split the server/client GPU arch, is asking if the NPUs are even worth supporting on linux, and now is excluding the NPUs from some new SoCs... This is not how to drive software adoption.
Intel SoC/dGPU support is sneaking into all sorts of "local AI" projects, even with Intel's massive current hardware disadvantage. ROCm is here and there, but I haven't seen a single implementation of Ryzen AI outside the quick AMD demo.
kpb321 - Thursday, November 2, 2023 - link
What kind of speeds can an existing 7540u hit on 15w for the CineBench MT test? According to their testing the new chip is faster at that point. That would have to mean either the existing chip is stuck pretty close to it's base 3.2ghz speeds or the 4c cores can still turbo up a fair bit beyond what we've seen in server parts. I do agree that it would be nice if AMD would come out and provide a more details boost speed than just the one fastest single core.ToTTenTranz - Thursday, November 2, 2023 - link
I wish AMD had made a version of Phoenix for PC gaming handhelds with 6-8x Zen4c cores and a wider 8-12CU GPU, with perhaps some 8-16MB Infinity Cache.Size might have been close to the original Phoenix but performance at 15-20W could be so much better.
Valantar - Friday, November 3, 2023 - link
I think you're... eh, let's say optimistic about AMD's budget for taping out new chip designs. Yes, they did make the Steam Deck chip, but only after Valve pre-ordered likely hundreds of thousands of them. Taping out a new design costs millions of dollars at the minimum. Handheld gaming is still a niche market, so it would take someone willing to take on that risk for AMD to do something like this, and the only real candidate there would be an SD sequel, which Valve has confirmed isn't coming for a while yet. I wouldn't be surprised to see a 6-8c Zen *c +large-ish GPU design for that when it arrives though, but I wouldn't expect that until... late next year, or 2025?andrewaggb - Thursday, November 2, 2023 - link
Questions:Can a 7840U/HS run accelerated torch, tensorflow, onnx etc via rocm? On linux only? or windows also? IS the performance any good? (I don't think these have decidated vram)
Is the AI a separate NPU in addition to the GPU or are they just using the GPU?
Ryan Smith - Thursday, November 2, 2023 - link
The NPU in Phoenix is a separate functional block. It is not reusing the GPU.As for what libraries are supported under what OSes, I'm afraid I don't have that information handy. I know that Ryzen AI is accessible via ONNX, but I don't have any further details than that.
https://github.com/amd/RyzenAI-SW
systemBuilder33 - Thursday, November 2, 2023 - link
ERROR : "The Chips: Ryzen 5 7545U and Ryzen 3 7440U" section, first table "AMD Ryzen 7040U series"The 7545U with 2+4c Ryzen chips must have a 12MB L3 cache, not 16. That makes sense. The 7640U has 16MB cache. The 7545 changes 4 cores to 'c' cores so it must save at least 4MB of cache. There is no way that the 7545U and 7640U have the same cache size. The whole point of 'c' cores is to save on cache, and if both cores are 16MB cache, that would be stupid.
Ryan Smith - Thursday, November 2, 2023 - link
Hi. The specs in that table are correct. The L3 cache is shared over the entire CCX; it's not a per-core arrangement. Even though AMD has synthesized a CCX with just 6 CPU cores, they put a full 16MB of L3 cache in there.More broadly speaking, Zen 4c in servers exists in part to save on L2 cache. But this isn't true for the mobile parts. The server parts had 2MB of L2 per core with Zen 4, and 1MB of L2 with Zen 4c. But the mobile parts always had 1MB of L2 per core.
garloff - Saturday, November 4, 2023 - link
Ryan: You meant to write L3 cache in the 2nd paragraph rather than L2 cache, right?Ryan Smith - Saturday, November 4, 2023 - link
As odd as it may seem, in this case the answer is no. I do mean L2 cache.Server Zen4c parts have less private L2 cache per core; it's part of how they managed to shave off 35% of the area of a CPU core.
As for L3 cache, AMD's chiplet CCDs have the same 64MB of shared L3 cache each, regardless of the core type used. It just happens to be that with 16 cores per Zen 4c CCD, you have fewer CCDs altogether than with Genoa (8x16 instead of 12x8). Which is moot for mobile, because the mobile CCX designs all are built with 16MB of L3.
FWhitTrampoline - Friday, November 3, 2023 - link
Monolithic APU Tape-outs are Bespoke affairs so can not really be compared to CCD/Chiplet based processors from AMD. And the Die Shots of some Zen-4/Zen-4C based APU designs all show the cores arranged around the same Ring Bus like topology on that monolithic APU die so the L3 is maybe just another stop along that Ring Bus(?) there and all the cores since Zen-3 have a unified L3 cache pool.And even Ryzen 4000 series APUs that had 2, 4 core Zen-2 CCX Units had no need for any IFOP(Infinity Fabric Off Package/Off Die) connectivity as the 2 CCX Units where on the same Monolithic slab of APU silicon so the latency was different and so was the memory access latency different compared to Chiplet based processors that had to go off die to access the memory controller on Zen-2/Later chiplet based processors where the I/O die was separate from the CCD DIE/DIEs.
meacupla - Thursday, November 2, 2023 - link
I think it's good this is only going to be done on the R3 and R5. Those two get to see some of the worst and lowest cost cooling solutions anyone could come up with. Literal penny pinching stuff.haplo602 - Friday, November 3, 2023 - link
So AMD will keep one Zen4 core at least in each chip to skirt the line of dishonest advertising when it comes to maximum clocks (See ? See ? That core CAN hit the clockspeed so all is fine) while not specifying the maximum clock speed for most of the chip ... I expected better ....nandnandnand - Friday, November 3, 2023 - link
"U" parts aren't going to be running all cores at 4.9 GHz in the first place.ballsystemlord - Friday, November 3, 2023 - link
@Ryan , "confirmed" was accidentally typed twice."Diving a bit deeper into the technical specifications for Phoenix 2, AMD has confirmed confirmed ..."
Ryan Smith - Saturday, November 4, 2023 - link
Thanks!trivik12 - Monday, November 6, 2023 - link
Are there any laptops announced with this chip? Based on previous releases, it will take ages before we see it available in retail. I still have not seen any Zen 4 laptops at Costco or Best buy though online availability and deals have been good in past month or so. Big Achilles heel for AMD is laptop is the least important part of their portfolio. That is surprising considering laptops sell way more than desktop.tipoo - Monday, November 6, 2023 - link
Seems great for Steam Deck and competitors, they're cooling and power bound anyways, so they're already operating where this would be most efficient and doesn't need to clock higher. You'd save power and die size both to give more to the IGP.