AMD: Zen 4, Speculation, Rumours and Discussion

Attachments

  • Spec 2017 total est score graphs anandtech.png
    Spec 2017 total est score graphs anandtech.png
    73.1 KB · Views: 47

yes this has V$ tiles only on one CCD for 2 CCD parts, very-very fun/cursed in that.
 

yes this has V$ tiles only on one CCD for 2 CCD parts, very-very fun/cursed in that.

I was looking forward to the new X3D series CPU, and was considering going for 7950X3D directly, but with the 'same size additional cache for all' and 120W TDP limitation, now I'll probably go for 7800X3D instead if the reviews are good enough.
 
Yes and coincidental that the 7800X3D is the only one with a lower boost clock than its non-X3D counterpart. Makes the choice harder.

I was really wanting to move up from 8 to 12 cores while still favouring gaming performance over productivity, but now the choice won't be so easy as 12 and 16 core variants have the potential of issues with thread scheduling, while they offer higher boost clock speeds in return... A tough choice, as if the price difference wasn't already enough.
 
I'm not sure to decode correctly the terrible naming convention, but it's official that there's no zen4 mobile at 15W?
 
Yes and coincidental that the 7800X3D is the only one with a lower boost clock than its non-X3D counterpart. Makes the choice harder.

I was really wanting to move up from 8 to 12 cores while still favouring gaming performance over productivity, but now the choice won't be so easy as 12 and 16 core variants have the potential of issues with thread scheduling, while they offer higher boost clock speeds in return... A tough choice, as if the price difference wasn't already enough.
+ afaik the v-cache / 3d cache is on one ccd only, so they will have to find a way to lock the thread needing it on the one ccd if they want max performances ?
 
+ afaik the v-cache / 3d cache is on one ccd only, so they will have to find a way to lock the thread needing it on the one ccd if they want max performances ?

Yes exactly. The 12 and 16 core variants will have to make use of the one CCD for gaming with the only difference between SKU's being default boost clocks.

I wonder if that would mean that the 7900X3D would have to have one fully enabled 8 core CCD with the 3DV-cache and one with only 4 enabled cores without? Possibly that's always what they've done for x900 SKU's, I just assumed it was 2 x 6 core CCD's.
 
The lower TDP is probably because they can't dissipate that much heat up through the cache die anyway (they already got big problems cooling those small chiplets under the ihs), so that would mostly affect all-core clocks (and probably not by a lot - the last 1-200mhz are really costly now).
But the 5 ghz max boost for 7800X3D would be for 1T anyway, where heat shouldn't be a problem, and the L3 speed shouldn't be coupled to core speed anyway, so why that limitation?

And, if it's not pure product segmentation (to let the expensive ryzen 9s be the ultimate gaming cpus), would that mean that the high boost clock of the dual CCDs only applies to the cache-less one? That would be really weird for the scheduler - 5.0 GHz and big L3 or 5.7 GHz and less L3.

I just assumed it was 2 x 6 core CCD's.

Mine is 2x6, but they are not equal. All core clocks:
1672955908322.png
(and it's not just worse silicon, at these clocks ccd0 is 12-13w/core, while ccd1 is 10w/core and lower temperatures)
 
Last edited:
+ afaik the v-cache / 3d cache is on one ccd only, so they will have to find a way to lock the thread needing it on the one ccd if they want max performances ?
It's going to be a scheduling nightmare, as up until now, pretty much every heterogeneous core architecture (Big.Little, etc) has had a very distinct hierarchy of cores, Intel's P-cores are always faster than the E-Cores, so your scheduling task is limited to figuring out which things you want to run the fastest and keeping that work on the P-Cores.

In this case though, the CCD with the v-cache is going to be at a ~500-700mhz clock deficit, so depending on the workload, either CCD could be faster. Is the Windows scheduler going to start real-time-profiling cache misses and instruction mix now? As much trouble as they've had with the scheduler thus far, this is going to be much harder to get right.
 
I wonder if that would mean that the 7900X3D would have to have one fully enabled 8 core CCD with the 3DV-cache and one with only 4 enabled cores without? Possibly that's always what they've done for x900 SKU's, I just assumed it was 2 x 6 core CCD's.
No. The AMD-provided code in Linux that figures out how many cores a CPU has provides two numbers: How many CCX and how many cores per CCX. Then amount of cores is A*B, and you address cores by CCX number and core number in CCX.

Or, to put it more plainly, there is no OS support for any configurations where there are enabled CCX with different core counts. If AMD intended to release such a product, they would change this code probably years in advance.

Yes and coincidental that the 7800X3D is the only one with a lower boost clock than its non-X3D counterpart. Makes the choice harder.

To be clear, not coincidental at all. All the CCDs with cache run at lower clocks. The dual-CCD products contain a high-clockspeed CCD and a high-cache one.
 
No. The AMD-provided code in Linux that figures out how many cores a CPU has provides two numbers: How many CCX and how many cores per CCX. Then amount of cores is A*B, and you address cores by CCX number and core number in CCX.

Or, to put it more plainly, there is no OS support for any configurations where there are enabled CCX with different core counts. If AMD intended to release such a product, they would change this code probably years in advance.

Good to know. For me that does make the 7900X3D an odd product. You're basically getting a 6 core CPU for gaming for more money than the 8 core. It's been proven again and again that in current games 6 fast cores is plenty, but I still can't help but wonder if this could change at some point in the near future. Basically you really need to value gaming and highly threaded productivity in equal amounts for it to be the best product for you.

To be clear, not coincidental at all. All the CCDs with cache run at lower clocks. The dual-CCD products contain a high-clockspeed CCD and a high-cache one.

I realised this must be what's going on a little while after posting. While somewhat misleading it does make perfect sense. It's not a case of purposefully hobbling the lower end SKU to try and introduce ladder creep, so that's nice. But at the same time all core performance for non-cache sensitive productivity tasks will be slightly lower than the non-3D counterparts and gaming performance likely very much the same as the 7800X3D.
 
All the CCDs with cache run at lower clocks. The dual-CCD products contain a high-clockspeed CCD and a high-cache one.
Why is this a given? It's not confirmed at least.

While the vcache die is obviously more power limited and thus more limited on all-core clocks, that should not affect the max boost clock, and the cores should be pretty decoupled from the L3 anyway.
 
The extra die on top of the CPU die in the X3D SKUs increases thermal resistance. The power hike this gen is really detrimental to the X3D SKUs because they *have* to be operated at lower power levels to keep T_junction the same. The benefit from massive caches might also be more modest given the increase in DDR5 bandwidth.

AIOs and mobile solutions are the best target markets for X3D SKUs now, IMO.

Cheers
 
Really interesting. I have a 7700x the 8/16 core. I think my next chip will be zen 5 with the ai cores and 3d cache and 16 cores lol. Gotta wait I think for that. Hopefully its socket compatible .
 
Back
Top