So it is more efficient at lower clocks but less at higher clocks. I would love a desktop version with 8 large cores, 8 efficient cores and all of them with a shared 3Dvcache, i can hope.
So Zen4C is identical to Zen4 in function with no other downside then they scale worse at higher power?
Sounds like a 4 core Zen4C only, would be fine and plenty for a browser and office pc replacement. Would maybe be better then a Intel 2p 8e because of the full instruction set.
I could see an Epyc chip having nothing but 4C cores in it. The better efficiency and smaller size means I'm sure they could increase the core count by 33% without there being a power draw issue. Crazy to think about having 256 cores, 512 threads on a single CPU package. And then doubling that up in a dual socket configuration.
Not all applications need the extra cache; the extra cores or lower power draw would be appealing to some.
TLD LARS:
Would maybe be better then a Intel 2p 8e because of the full instruction set.
Depends how you look at it. Intel's E-cores are roughly 1/4 the size of P-cores but they don't have SMT, they operate at a lower clock speed, they don't have the full instruction set, and I don't think they have an L3 cache (or at least it's shared with P-cores). The only thing we don't know for sure is the power efficiency, but considering AMD's "vanilla" Zen4 cores aren't that much more power hungry than Intel's E-cores, I assume the 4C cores must be a lot better to warrant their R&D.
I think the real porpose of the Zen4c cores is to save money because they are much smaller.
Being more efficient at lower speeds is nothing special, normal Zen4 cores can probably be as efficient as 4c if they also lower the clocks and the voltage.
And i also agree that a CPU with "just" 8 Zen4c cores would be more than enough for most people.
Being more efficient at lower speeds is nothing special, normal Zen4 cores can probably be as efficient as 4c if they also lower the clocks and the voltage.
I don't see how being as-efficient is possible when there are physically less transistors. There are situations where a cache-intensive process would run more efficiently on the full-size cores, but, at lower clock speeds, there usually isn't as much of a need for a big cache.
Depends how you look at it. Intel's E-cores are roughly 1/4 the size of P-cores but they don't have SMT, they operate at a lower clock speed, they don't have the full instruction set, and I don't think they have an L3 cache (or at least it's shared with P-cores). The only thing we don't know for sure is the power efficiency, but considering AMD's "vanilla" Zen4 cores aren't that much more power hungry than Intel's E-cores, I assume the 4C cores must be a lot better to warrant their R&D.
The E-cores are still much smaller then Zen4C cores yes.
My hope is that the not perfect working thread director can be vastly simplified with AMD and all work can be shared fully between normal and C cores, that should make for much more stable performance at low TDP.
My Intel 2P and 8E core work pc is fast enough for most office tasks, but if I hit avx load or some other missing E core instruction the CPU is still only a dual core with 8 helpers that may not be smart enough to help doing work.
From the slides, AMD could be the other way around, the C cores can be main workers for efficient work and the normal cores are there for peak single core speeds when needed.
Full Zen4C chips could be interesting for 10 watts chips.
Zen4C cores are not meant to be E-cores.
They are the same architecture and have the same IPC as Zen4. But they use higher density cells, and this is why they get a 35% reduction in size for the cores+L2.
The problem is that higher density cells mean lower signal integrity and higher temperatures per area. So they have to be clocked lower.
Another thing to consider is that the design has changed. We can see on the die shots that the layout of units in different and it's packed much closer.
Usually, AMD design their CPUs by designing each part individually and then putting it back together. This has the advantage of being faster to develop and if there is a problem in one individual part, AMD only has to fix one.
The downside is that the layout of the chip is not very optimized, resulting in some wasted space.
With Zen4C, the layout was optimized not to waste any die space.
The advantage of a Zen4C core is that they are fully identical in instruction support, as the big cores. They just lack L3 cache and have lower clocks.
But they will have the same IPC at a much lower power and die space budget.
My hope is that the not perfect working thread director can be vastly simplified with AMD and all work can be shared fully between normal and C cores, that should make for much more stable performance at low TDP.
I'm not saying you're wrong, but what suggests the thread director isn't sufficient? I got the impression the OS scheduler is more often than not the problem with CPU inefficiencies. That's why Alder Lake was poorly received - I didn't see anything particularly bad about the design but schedulers just have a hard time figuring out what a process needs. Kinda a shame since Alder Lake was kinda doomed to fail in initial reviews.
I'm not saying you're wrong, but what suggests the thread director isn't sufficient? I got the impression the OS scheduler is more often than not the problem with CPU inefficiencies. That's why Alder Lake was poorly received - I didn't see anything particularly bad about the design but schedulers just have a hard time figuring out what a process needs. Kinda a shame since Alder Lake was kinda doomed to fail in initial reviews.
There have been multiple examples with games where the thread director has not been working correctly, leading to bad performance with E cores on.
When turning the E cores off many have seen an improvement in performance just because the CPU is no longer in doubt where to send the work.
The increase in performance could not be explained by limited in TDP, busy P cores or other limits being hit.
If the thread director was perfect, Intel could in theory have AVX512 working on the P cores and AVX threads could be directed to the core able to do the work.
Just like with 7950X3D, work is not send to the correct cores perfectly leading to the 7800X3D often being faster at game loads.
With the C cores things are a lot more simple because they are much closer to "normal" cores, so the AMD version of the thread director can just send work to C cores when thermal, power, battery or other limits are in place and to main cores when those limits are far away.
I am guessing a Intel E core is 15% speed of a P core and AMD C cores look more like 70-80% of the normal cores (even closer if normal cores are throttled by limits) the consequence of sending the work to the wrong core is much lower and the user should be much less effected.
There have been multiple examples with games where the thread director has not been working correctly, leading to bad performance with E cores on.
When turning the E cores off many have seen an improvement in performance just because the CPU is no longer in doubt where to send the work.
But that's what I'm trying to say: how do we know it's the thread director and not the OS scheduler? Windows' scheduler is notoriously terrible. It was competent during the Core2 days and it got a little better once Intel re-introduced HT, but it took years for AMD's SMT to perform decently, I don't think it ever behaved as expected with ARM's big.LITTLE configurations, and it's been notoriously terrible with multiple NUMA nodes. So considering Microsoft's track record with developing a competent scheduler, I have a hard time believing the CPU's thread director is the problem. There is one little inkling that it may be the problem: Linux performance. Alder Laker had director/schedule issues on Linux too, and unlike Windows, Linux has a lot of good schedulers and you can swap between them for a more fine-tuned experience. In fact, for the 5.16 kernel, it actually got slower. When Windows 11 was released, there was basically no real improvement for the E-core performance but Linux I know has had multiple kernel patches to alleviate performance issues. I never saw benchmarks to show whether this worked or not.
Just like with 7950X3D, work is not send to the correct cores perfectly leading to the 7800X3D often being faster at game loads.
Right but that is definitely a case of crappy schedulers.
I am guessing a Intel E core is 15% speed of a P core and AMD C cores look more like 70-80% of the normal cores (even closer if normal cores are throttled by limits) the consequence of sending the work to the wrong core is much lower and the user should be much less effected.
Depends on the workload and how you're measuring it. In terms of performance-per-clock, I get the impression the E cores have a nearly 1:1 ratio with P cores so long as you're not using any advanced instructions. With AMD's C cores, I'm sure it's actually faster so long as you aren't doing anything cache-intensive.
The advantage of a Zen4C core is that they are fully identical in instruction support, as the big cores. They just lack L3 cache and have lower clocks. But they will have the same IPC at a much lower power and die space budget.