I don't really understand the point of including c and non-c cores on the same package - either you want an efficient chip or a high performance one. To make a hybrid means you get neither. This is not like Intel's design, where the E cores are functionally different from P cores (for better or worse). For Intel's architecture, a hybrid design is necessary for anything that isn't a high-end workstation (all P-cores) or a large-scale web host server (all E-cores). I know people here will disagree with me on that, but considering Intel's situation, their hybrid design is the only thing they can lean on. The Zen c and non-c cores are basically the same thing but with the c cores having lower clocks and less cache (because you don't need as much cache if the clock speeds don't demand more). So that means unlike E cores, c cores can process the most advanced workloads. In a world with chips where CPUs dynamically overclock themselves, idle power draw is practically negligible, and (in the case of AM5) a package with plenty of room to spare for full-size cores, in what use case does it make sense to have some cores that can handle the exact same task slower than others? The Windows scheduler is notoriously stupid - it will have a problem with putting background processes on the non-c cores and it will also put heavy single-threaded tasks on the c cores. I've spent several minutes thinking about this and can't find a single good reason for AMD to make a hybrid design. Just give us an all-c and all non-c model. I firmly believe the c models will sell well - I would certainly consider it.

I think it`s a question of being cheaper to produce than a regular 8 core part with just "normal" cores because the C cores have less cache, something that occupies a lot of space on a CPU.

H83:

I think it`s a question of being cheaper to produce than a regular 8 core part with just "normal" cores because the C cores have less cache, something that occupies a lot of space on a CPU.

Then why not do an all-c-core model? That would still be a very worthwhile product.

schmidtbag:

Then why not do an all-c-core model? That would still be a very worthwhile product.

Benchmarks?...

H83:

Benchmarks?...

I don't really get the point of your question - are you implying nobody would be interested in an 8c/16t CPU with only c cores? If the goal is a lower power and/or lower-cost model, sacrificing the peak boost clock doesn't matter if the CPU is otherwise equally capable to the non-c part. Seems like a perfect option for laptops, embedded devices, and high-end server chips (high boost clocks are rough on server PSUs and cooling) as far as I'm concerned. To paraphrase something I said before: the cache needs to be big enough to supply the CPU's demand. If the clock speed is low enough, you don't need as much cache. Less cache means fewer transistors, which means further improved efficiency and a cheaper die to make. This means the smaller cache usually shouldn't be detrimental to performance. Anyway if you must have benchmarks, there's currently no apples to apples comparison because to my knowledge, there are no benchmarks of a Zen4 CPU with an equal core count to a Zen4c CPU. However, the 9754 certainly has proved to not be slow: https://www.phoronix.com/review/amd-zen1-zen4c https://www.phoronix.com/review/amd-epyc-9754-bergamo It has a max boost of 3.75GHz, which can apply to all cores. The 9654, meanwhile, has fewer cores at a max boost of 3.7GHz for just a single core (3.55 for all cores). The 9654 has regular Zen cores. When you look at performance-per-watt, the Zen4c cores win by a wide margin, despite operating at higher boost clocks. These are very good CPU cores for anyone who doesn't care about a self-overclocking chip.

I'm just joking because sometimes HW makers make some decisions in order to make their products look better when running benchmarks... Like factory overclocks that destroy the efficiency of Cpus and GPUs nowadays... And I mostly agree with your points and I dislike the big/little approach to desktop/laptop Cpus but Intel and AMD think otherwise.

schmidtbag:

The Windows scheduler is notoriously stupid - it will have a problem with putting background processes on the non-c cores and it will also put heavy single-threaded tasks on the c cores.

actually i think the Windows scheduler is the raison d'etre for the c cores. i believe the reason for the incorporation is power draw and the processes run time impact will be lesser than Intel big/little as all cores are competent at all tasks. although as you imply not as good as a "full fat" core count and better than an all c core count.

schmidtbag:

I've spent several minutes thinking about this and can't find a single good reason for AMD to make a hybrid design. Just give us an all-c and all non-c model. I firmly believe the c models will sell well - I would certainly consider it.

Keep the speedy singlecore 5 Ghz and when the CPU is slowly loaded up with work and hitting power and thermal limits, the fast main cores and C cores are going to be very close in speed anyway, so there is almost no A team and B team in the CPU anymore, when fully loaded. In my opinion delivering work to normal and C cores should be easier then a 7950X3D, because they are more equal then a big cache and small cache cores. If power and thermal are practically unlimited then there is still the size difference and cost difference of the C cores, a 16 core CPU could become a 20-24 core CPU depending on how many of each there is.

ZAMDus: Release the Kraken

TLD LARS:

Keep the speedy singlecore 5 Ghz and when the CPU is slowly loaded up with work and hitting power and thermal limits, the fast main cores and C cores are going to be very close in speed anyway, so there is almost no A team and B team in the CPU anymore, when fully loaded.

Right so why include both when the c cores are cheaper and can maintain boost clocks for longer? At that point, the only benefit of a non-c core is short busts of single-threaded tasks.

If power and thermal are practically unlimited then there is still the size difference and cost difference of the C cores, a 16 core CPU could become a 20-24 core CPU depending on how many of each there is.

Well, AMD has already managed to fit 16 cores on AM4 and AM5 packages. I agree though that c cores make much more sense for anything higher than that on such sockets, not just because of physical size but because motherboards and most heatsinks likely can't handle power delivery of a >16 non-c core CPU.

schmidtbag:

Right so why include both when the c cores are cheaper and can maintain boost clocks for longer? At that point, the only benefit of a non-c core is short busts of single-threaded tasks.

Full C-cores can hurt gaming and non multicore tasks like general office usage. A lot of time at the computer is spend in burst loads and relatively low core count situations where 2 normal cores at 5-5.5Ghz are preferred over 4-8 C cores at 3.7Ghz. A CPU can maybe run dualcore 5Ghz with 2 normal cores without hitting limits, that is still 1.3Ghz faster then the C cores might hit at full speed. A lot of people are still afraid to sacrifice singlecore speed for more allcore performance. Personally I have not seen evidence to suggest that a C-core is more power efficient other then the fact that 4 cores at 3.5 Ghz can be more efficient then 2 cores at 5Ghz, if the load permit.

TLD LARS:

Full C-cores can hurt gaming and non multicore tasks like general office usage. A lot of time at the computer is spend in burst loads and relatively low core count situations where 2 normal cores at 5-5.5Ghz are preferred over 4-8 C cores at 3.7Ghz. A CPU can maybe run dualcore 5Ghz with 2 normal cores without hitting limits, that is still 1.3Ghz faster then the C cores might hit at full speed. A lot of people are still afraid to sacrifice singlecore speed for more allcore performance. Personally I have not seen evidence to suggest that a C-core is more power efficient other then the fact that 4 cores at 3.5 Ghz can be more efficient then 2 cores at 5Ghz, if the load permit.

I agree, that's why I think it makes sense to have a CPU with all of one core but not the other. If you have workloads that need high burst speeds or good single-threaded performance, get the non-c. If you are an enthusiast who likes to push your system to the limits, get the non-c. If you are known to use something cache-intensive, you're probably better off with an X3D model. For everyone else, get the c. Perhaps if the Windows scheduler was a little less stupid, asymmetric cores might have a use case for a general-purpose desktop CPU, but until then, you're better off doing all of one core type.

schmidtbag:

I agree, that's why I think it makes sense to have a CPU with all of one core but not the other. If you have workloads that need high burst speeds or good single-threaded performance, get the non-c. If you are an enthusiast who likes to push your system to the limits, get the non-c. If you are known to use something cache-intensive, you're probably better off with an X3D model. For everyone else, get the c. Perhaps if the Windows scheduler was a little less stupid, asymmetric cores might have a use case for a general-purpose desktop CPU, but until then, you're better off doing all of one core type.

You might be forgetting the size difference of the cores, a 4 N (normal) + 4 C would be cheaper to make and sell compared to 8 N + 0 C. Something exotic like a 2 N + 4 C + spare room for extra cache functioning like 3D cache, but on the same level instead of 3D, could be a faster APU then a regular 6 Normal core APU. A single chip 4N + 6 C laptop could maybe be possible on the same size as a 8N CPU, giving a faster allcore pc without sacrificing the 1-4 core boost speed. The windows scheduler might not be so much of a problem, it is my understanding that AMD found a way to prefer the cores attached to RAM in the Threadripper 2000 series via software (working like process lasso), this is going to be the same thing: Use Normal cores first then spill over to C-cores when full, if the cores are equal at 3.5Ghz (except the cache) sending work to the "wrong" core at full multicore load might have very little consequence, compared to a Intel P and E core setup.

schmidtbag:

I don't really understand the point of including c and non-c cores on the same package - either you want an efficient chip or a high performance one. To make a hybrid means you get neither. This is not like Intel's design, where the E cores are functionally different from P cores (for better or worse). For Intel's architecture, a hybrid design is necessary for anything that isn't a high-end workstation (all P-cores) or a large-scale web host server (all E-cores). I know people here will disagree with me on that, but considering Intel's situation, their hybrid design is the only thing they can lean on. The Zen c and non-c cores are basically the same thing but with the c cores having lower clocks and less cache (because you don't need as much cache if the clock speeds don't demand more). So that means unlike E cores, c cores can process the most advanced workloads. In a world with chips where CPUs dynamically overclock themselves, idle power draw is practically negligible, and (in the case of AM5) a package with plenty of room to spare for full-size cores, in what use case does it make sense to have some cores that can handle the exact same task slower than others? The Windows scheduler is notoriously stupid - it will have a problem with putting background processes on the non-c cores and it will also put heavy single-threaded tasks on the c cores. I've spent several minutes thinking about this and can't find a single good reason for AMD to make a hybrid design. Just give us an all-c and all non-c model. I firmly believe the c models will sell well - I would certainly consider it.

it allows them to fit more cores on chip and allows them to exploit the efficiency curve of the silicon better, with zen4, the zen4c cores take up roughly half the space. chip yields are probably the main reason for this choice on apus, and also potentially lower power consumption for ultra low power skus. ( the main purpose of the hybrid design is to allow them to keep the high single threaded performance while saving space ect.)

TLD LARS:

You might be forgetting the size difference of the cores, a 4 N (normal) + 4 C would be cheaper to make and sell compared to 8 N + 0 C. Something exotic like a 2 N + 4 C + spare room for extra cache functioning like 3D cache, but on the same level instead of 3D, could be a faster APU then a regular 6 Normal core APU. A single chip 4N + 6 C laptop could maybe be possible on the same size as a 8N CPU, giving a faster allcore pc without sacrificing the 1-4 core boost speed.

user1:

it allows them to fit more cores on chip and allows them to exploit the efficiency curve of the silicon better, with zen4, the zen4c cores take up roughly half the space. chip yields are probably the main reason for this choice on apus, and also potentially lower power consumption for ultra low power skus. ( the main purpose of the hybrid design is to allow them to keep the high single threaded performance while saving space ect.)

We have already addressed that the c cores are smaller. As stated before, AMD already managed 16c/32t with a v-cache, so what's the extra room needed for on a 4N/4C configuration? Perhaps it's for a giant iGPU, but there's not much point in going bigger than what they've got now considering memory bandwidth is the current bottleneck. Considering TSMC seems to have rather good yields (hence AMD not bothering with hardly anything lower-end than 4c/8t in the past couple generations), I don't think yields have much to do with it.

The windows scheduler might not be so much of a problem, it is my understanding that AMD found a way to prefer the cores attached to RAM in the Threadripper 2000 series via software (working like process lasso), this is going to be the same thing:

AMD does seem to have preferred cores figured out but the Windows scheduler is still quite bad, compared to Linux for example.

Use Normal cores first then spill over to C-cores when full, if the cores are equal at 3.5Ghz (except the cache) sending work to the "wrong" core at full multicore load might have very little consequence, compared to a Intel P and E core setup.

In theory, this sounds very good. The problem is, it's more complicated than that, because the OS doesn't know which tasks are going to be more demanding or what deserves top priority. For example, let's say you have a game that only needs 8 threads. You have a 4N/4c configuration. Logic dictates that the Normal cores get top priority so the game runs only on those, but that's not necessarily what would happen: something like Windows Update could start hogging some of the cycles of the N cores, because that can still be a CPU-heavy task, even though it really doesn't need N cores. Even for something as simple as a notification could for a brief moment use one of the N cores, which in and of itself isn't a big deal but it may require the cache to be wiped from one of the cores, and stuff like that cause microstutter in a game.

schmidtbag:

We have already addressed that the c cores are smaller. As stated before, AMD already managed 16c/32t with a v-cache, so what's the extra room needed for on a 4N/4C configuration? Perhaps it's for a giant iGPU, but there's not much point in going bigger than what they've got now considering memory bandwidth is the current bottleneck. Considering TSMC seems to have rather good yields (hence AMD not bothering with hardly anything lower-end than 4c/8t in the past couple generations), I don't think yields have much to do with it. .

wafers are expensive on 4/5nm tsmc, so even if the defect rates are low, more chips per wafer are going to be significant cost savings.

user1:

wafers are expensive on 4/5nm tsmc, so even if the defect rates are low, more chips per wafer are going to be significant cost savings.

Right, which is why I was saying it makes sense to have an all-c-core chip as an option. Either c-cores are sufficient or they're not - why have them if your workload demands more processing power? Conversely, while it might make sense to have a couple normal cores (assuming the scheduler isn't dumb), if you bought c-cores, chances are, you're looking to keep the power draw to a minimum. If you're building a 16c/32t workstation, either you're trying to crunch a lot of numbers quickly (where c-cores are just slowing you down unnecessarily) or you're in an environment where the boost clocks can't/shouldn't be sustained (where normal cores would be a waste). If you're building a gaming rig, the last thing you want is an inconsistent frame rate, whether that be because the c-cores take priority (and perhaps give you a lower average frame rate) or because swapping between core clusters introduces microstutter. You're better off with just a single set of cores. If you're building a basic home/office PC, why even bother with normal cores? c cores will work just fine while sipping power.

schmidtbag:

Right, which is why I was saying it makes sense to have an all-c-core chip as an option. Either c-cores are sufficient or they're not - why have them if your workload demands more processing power? Conversely, while it might make sense to have a couple normal cores (assuming the scheduler isn't dumb), if you bought c-cores, chances are, you're looking to keep the power draw to a minimum. If you're building a 16c/32t workstation, either you're trying to crunch a lot of numbers quickly (where c-cores are just slowing you down unnecessarily) or you're in an environment where the boost clocks can't/shouldn't be sustained (where normal cores would be a waste). If you're building a gaming rig, the last thing you want is an inconsistent frame rate, whether that be because the c-cores take priority (and perhaps give you a lower average frame rate) or because swapping between core clusters introduces microstutter. You're better off with just a single set of cores. If you're building a basic home/office PC, why even bother with normal cores? c cores will work just fine while sipping power.

the high frequency cores do benefit office workloads , These kind of designs are a tradeoff, and are specific to their market segments, there will be probably be c core only designs, in the <10w sku range. I also found this slide. https://cdn.mos.cms.futurecdn.net/H6QsYXY7mY6TRhuzyijMtn.jpg this is interesting because they basically perform the same in the typical laptop power range, But as mentioned earlier the 2+4 design will be smaller than a 6 core zen4 design and will retain the single threaded performance for lightly threaded workload, so I would say that pretty much seals the deal on this one. these aren't really for gaming anyway also in the lower power range the hybrid design performs better, ( again retaining single threaded performance)

schmidtbag:

Right, which is why I was saying it makes sense to have an all-c-core chip as an option. Either c-cores are sufficient or they're not - why have them if your workload demands more processing power? Conversely, while it might make sense to have a couple normal cores (assuming the scheduler isn't dumb), if you bought c-cores, chances are, you're looking to keep the power draw to a minimum. If you're building a 16c/32t workstation, either you're trying to crunch a lot of numbers quickly (where c-cores are just slowing you down unnecessarily) or you're in an environment where the boost clocks can't/shouldn't be sustained (where normal cores would be a waste). If you're building a gaming rig, the last thing you want is an inconsistent frame rate, whether that be because the c-cores take priority (and perhaps give you a lower average frame rate) or because swapping between core clusters introduces microstutter. You're better off with just a single set of cores. If you're building a basic home/office PC, why even bother with normal cores? c cores will work just fine while sipping power.

A laptop can have the following frequency per active cores (just a guess for examples sake): 1 core 5000Mhz 2 core 4500Mhz 3 core 4100Mhz 4 core 3800Mhz All these are normal cores in a cluster of 4 that are used first, at higher multicore load with all the normal cores busy the jobs are directed to the C cores. 5 core 3600Mhz 6 core 3500Mhz 7 core 3400Mhz 8 core 3300Mhz The slower max speed of the C-cores does not matter because the normal cores are already down to that level because of power or thermal limits, so there is practically no difference between normal and C cores when the CPU is loaded up. I do not want to comment more at this time because I have not found enough info on the precise workings yet.