AMD Zen 6 Discussion

D

David_Schroth

Administrator
Staff member
Joined
Apr 23, 2019
Messages
1,345
Points
113
Sorry, it's an old Anandtech in joke 🙂 I absolutely don't expect SMT4, because at that point you start adding more resources prioritising SMT to the detriment of per-core performance. There's some utility as a licensing hack for software that is charged per core, rather than per thread, but that's a serious niche.

I do wonder what level of single thread performance we'll get at reasonable cache sizes and clock speeds. Just from the timing, I expect to see Zen 6 in the next generation of consoles, so it's probably going to set the gaming baseline performance level for the next 10 years.
 
Maybe don't expect 7 GHz. It'll probably be something more conservative than that. There's still a lot we don't know. Perhaps if some of the acerbic leakers follow us over here we can get a repeat of stuff leaked/teased elsewhere. Or people can just mine the Zen6 speculation thread from AT for fodder.

That being said, it would be very sad if Medusa Halo doesn't become a thing. An upgraded Strix Halo sounds kinda neat actually.
 
DrMrLordX said:
Maybe don't expect 7 GHz. It'll probably be something more conservative than that. There's still a lot we don't know. Perhaps if some of the acerbic leakers follow us over here we can get a repeat of stuff leaked/teased elsewhere. Or people can just mine the Zen6 speculation thread from AT for fodder.

That being said, it would be very sad if Medusa Halo doesn't become a thing. An upgraded Strix Halo sounds kinda neat actually.
Click to expand...
The ceiling for single-core (maybe even two CPU cores boost)will probably be around 6.5-6.7ghz. This refers to the strongest or best Zen 6 CPU-s.
 
Overclockers are currently hitting ~5.8GHz using bus clock and BIOS trickery on the best Zen 5 samples they can find (including X3D variants). AMD could easily stabilize higher clockspeeds by design IMO. 6.0GHz as a stock all-core might be reasonable for top-end samples.

Much more interested in a new memory controller though, personally. With 12 cores per CCD, they're going to need to get >DDR5-8000 stable across the board, i.e. out of the box with EXPO. I assume CUDIMM support will be included and help with this.
 
I just want 12 cores per CCD with more L3 on the CCD. Hopefully its not a 1:1 ratio on the increase and we get a little bump in L3 too.
 
LazyGamer said:
They'd need to abandon their multi-CCD configuration that adds another interconnect stage if they want to drop latency any significant amount really.
Click to expand...
I'm not going to armchair engineer the CPU's... not like back in the 90's when PC magazine deep dove into CPU design. Sigh.. I miss that mag.

Where was I... this isn't my house... WHO ARE YOU! STRANGER DANGER!!! STOP DROP AND ROLL..


Ok anyway I just think they need to find a better way of utilizing the memory and doing the interconnects better between CCD's. I'm of a mind that for those of us that WANT the big core counts for desktop use we will require the CCD's to be there. They save on cost in making them because you don't run the risk of bad silicone when you pack less into each CCD. (Meaning your failures are not as impactful) As it is going to 12 cores per CCD is a 50% pump in the top line, I expect to see at least a 30% cost increase for those top line CPU's.

And the interconnect issue should have a better solve as well, such as a fiber optic or light path traces between the CCD's. And a higher speed better designed interconnect. Don't forget that historically for these Ryzen CPU's at least, the interconnect and I/O controllers are a larger technology set altogether. It's distinctly possible with going to a smaller node with a bit better technology they can decrease trace sizes and increase throughput of the controllers and interconnects while still saving cost per CPU.

But I'm not that type of Engineer. IT's outside of my purview just making some logical statements based on observations and information that has stuck in my mind... for now.
 
Asterox said:
The ceiling for single-core (maybe even two CPU cores boost)will probably be around 6.5-6.7ghz. This refers to the strongest or best Zen 6 CPU-s.
Click to expand...

6.5 GHz seems likely. But what do I know, maybe they will aim for more? Might not be worth it, especially once they get to the X3D model.

LazyGamer said:
Much more interested in a new memory controller though, personally. With 12 cores per CCD, they're going to need to get >DDR5-8000 stable across the board, i.e. out of the box with EXPO. I assume CUDIMM support will be included and help with this.
Click to expand...

DDR5-8000 is being teased, though we don't know if that'll be 1:1 or not. Also I don't know if we have Det0x over here yet but I believe he's done DDR5-8000 on Zen5, albeit not 1:1.
 
DrMrLordX said:
DDR5-8000 is being teased, though we don't know if that'll be 1:1 or not. Also I don't know if we have Det0x over here yet but I believe he's done DDR5-8000 on Zen5, albeit not 1:1.
Click to expand...
Folks have pushed to DDR5-8500 - but obviously with excessive binning of all parts involved and less pedestrian cooling, on enthusiast platforms.

Main thing we know is that the memory can handle it. It's up to AMD to design a memory controller to do 4GHz for 1:1 at those speeds.

Will also note that folks are reporting better performance all around with tuned 8000+, despite running at 1:2 on Zen 5. And that Intel didn't even bother with 1:1 for their DDR5 memory controllers.
 
Chiplets are only a small part of the cross-CCX latency. Chips and Cheese tested the HX 370 in depth, and the 3950X has half the cross-CCX latency of the HX 370. Strix Point's cross-CCX latency is so bad, it's higher than the DRAM latency.

In comparison, the 9950X's cross-CCX latency is only ~20ns higher than Strix Point. Zen5 as a whole seems to be buggy in this regard, on client at least.

I'd be surprised if Zen6's cross-CCX latency wasn't a huge improvement.
 
You must log in or register to reply here.
Become a Patron!
Back
Top