AMD: Zen 4, Speculation, Rumours and Discussion

[Gubbi]

The extra silicon on top of their compute dies, increases thermal resistance. To maintain the same T_junction they have to run at lower power, which means lower frequency. There is nothing AMD can do, it's physics.

AMD opted for a massive increase in TDP this gen to compete with Intel, - and their even more ridiculous TDP. This invariably renders the 3D SKUs less attractive.

The 3D cache might still make sense in mobile, and for high performance, lower power applications, like big AIOs (Apple iMac clones).

cheers

 

[pjbliverpool]

Looks really impressive. Trading blows with the 13900K in apps and a little faster on average in games (although sizable swings each way). But doing it all at a massively reduced power draw - particularly in the 'favor cache' config which is almost certainly what I would run this at due to the lower power draw and faster game performance with only a small loss in apps performance.

This is just from the Tech Power Up review mind you but from that I'm impressed. I'd likely choose this over the 13900K at the same price for my workloads. Granted the 13900K is a lot cheaper.

 

[pcchen]

I was looking forward to 7950X3D but I have to say it doesn't meet my expectations.
However, I do think it's the right way to have a lower TDP. Having a small chip using more than 200W is not very healthy.
There are some reports that Intel is aiming for 50% better efficiency (performance per watt) in their next generation CPU (Meteor Lake) which if true is good.

Personally I'm still considering this because I don't think 7800X3D will be faster (it'll likely to be better in value but probably not in performance). Also cost-wise a new MB and new RAM is going to cost a lot and CPU is only part of the equation. To me it's either this or maybe waiting for Meteor Lake.

 

[homerdog]

Is there a reason they don't put v-cache on both CCDs other than cost? This configuration is clearly problematic. Look at some of the results Hardware Unboxed got with the stock 7950X3D vs the simulated 7850X3D with the cacheless half disabled. The 7850X3D 7800X3D absolutely smokes the 7950X3D in certain games when the schedular gets it wrong. And I don't know how Windows is supposed to know which CCD any specific program should run on.

 

[Kaotik]

Is there a reason they don't put v-cache on both CCDs other than cost? This configuration is clearly problematic. Look at some of the results Hardware Unboxed got with the stock 7950X3D vs the simulated 7850X3D with the cacheless half disabled. The 7850X3D absolutely smokes the 7950X3D in certain games when the schedular gets it wrong. And I don't know how Windows is supposed to know which CCD any specific program should run on.

Because it limits clocks and going over 8 cores doesn't help games, where the cache is actually useful.

Windows doesn't know, currently they use the chipswt driver + Xbox Gamebar to recognize games running and then assign it to the CCD with cache and park the other CCD. Cores on the parked CCD activate on per need basis if there's other loads to take care of.

 

[Albuquerque]

Yeah, the 7850X3D seems like the optimal implementation to me.

 

[homerdog]

Because it limits clocks and going over 8 cores doesn't help games, where the cache is actually useful.

Windows doesn't know, currently they use the chipswt driver + Xbox Gamebar to recognize games running and then assign it to the CCD with cache and park the other CCD. Cores on the parked CCD activate on per need basis if there's other loads to take care of.

I see. So the Xbox Gamebar doesn't recognize Factorio as a game? The simulated 7850X3D is 80% faster than the 7950X3D in that game

The 7850X3D 7800X3D is definitely the chip gamers want. Much cheaper and no risk of decimated performance due to a scheduling error. Also I've read that the 7850X3D 7800X3D may have higher clocks on the v-cache cores than the 7950X3D and 7900X3D.

 

[Kaotik]

I'm pretty sure you both mean 7800X3D, there's no 7850X3D.
Yes, 7800X3D is the CPU gamers want. 7950X3D is the CPU for gamers who also do content creation etc. 7900X3D is for no-one and makes no sense whatsoever (especially since it's 6+6 instead of 8+4)

Also I've read that the 7850X3D may have higher clocks on the v-cache cores than the 7950X3D and 7900X3D.

Nope, 5 GHz max boost, same as 7900X3D/7950X3D

 

[Albuquerque]

I'm pretty sure you both mean 7800X3D, there's no 7850X3D.
Yes, 7800X3D is the CPU gamers want. 7950X3D is the CPU for gamers who also do content creation etc. 7900X3D is for no-one and makes no sense whatsoever (especially since it's 6+6 instead of 8+4)

Nope, 5 GHz max boost, same as 7900X3D/7950X3D

I actually hadn't done any reading; I presumed there was a "hypothetical" 7850X3D which was exactly half of a 7950X3D but only using the vcache CCD.

Turns out, you're precisely right, that part is the 7800X3D A single, fully-enabled 8-core CCD with 96MB of L3 with a max boost clock of 5GHz and an MSRP of $449. That seems the right answer for gamers on AMD...

 

[homerdog]

I'm pretty sure you both mean 7800X3D, there's no 7850X3D.
Yes, 7800X3D is the CPU gamers want. 7950X3D is the CPU for gamers who also do content creation etc. 7900X3D is for no-one and makes no sense whatsoever (especially since it's 6+6 instead of 8+4)

Nope, 5 GHz max boost, same as 7900X3D/7950X3D

Yea that's my bad. Shame that AMD knows the 7800X3D is what people want and delayed it to push people to the much more expensive and potentially inferior (for gaming) Ryzen 9.

Hardware Unboxed saw 4.8GHz on the v-cache cores so I'm not sure what's going on there. Looking for more data on how the CCDs actually clock in real use.

 

[homerdog]

I actually hadn't done any reading; I presumed there was a "hypothetical" 7850X3D which was exactly half of a 7950X3D but only using the vcache CCD.

Turns out, you're precisely right, that part is the 7800X3D A single, fully-enabled 8-core CCD with 96MB of L3 with a max boost clock of 5GHz and an MSRP of $449. That seems the right answer for gamers on AMD...

Having 8 v-cache cores and a few more smaller "efficiency" cores would be ideal IMO. The small cores could handle background stuff so that games could get all 8 of the big ones. It's easier for the OS to schedule things if it knows that one set of cores is always gonna the fastest, like on Intel. Having a situation where certain cores are faster in certain applications sounds like a nightmare for the scheduler and will require a lot of user/developer intervention to ensure you don't end up in a Factorio situation. Unless there is some scheduling magic that I'm not aware of. Which doesn't seem to be the case given some of the results I've seen.

 

[Albuquerque]

I'm torn on the whole asymmetric core count thing. I understand how it should work and why it would be cool; at the same time there's no good way for an OS scheduler to make good decisions in a vacuum about which services should operate on which cores. Sure, the native OS threads should be understood, but what about all the 3rd party apps and their related background tasks / services / threads? This isn't just Windows either, the same applies to Linux too.

For this to really work in my head, I think we need a way for applications to assign some metadata tagging to their threads to ask the OS to schedule in a specific way. Otherwise, it's just some dudes writing kernel logic making guesses or building whitelists or regexes or whatever. That's a solution that is destined to fail.

 

[homerdog]

Ryzen 7950X3D with One CCD Disabled - The 7800X3D Preview

Ryzen 7 7800X3D will be the most affordable Zen 4 X3D processor, but it doesn't release until April. We reconfigured the Ryzen 9 7950X3D to match the specs of the 7800X3D and ran it through our test suite, which includes 14 games at 720p, 1080p, 1440p and 4K. Results are impressive.

www.techpowerup.com

Simulated 7800X3D smashing the stock 7950X3D in Spiderman Remastered and Mount and Blade 2. Even setting the scheduler to "prefer cache" doesn't help in Mount and Blade 2. Turning off CCD1 improves FPS from 195 to 260. Oddly even disabling CCD0 improves things slightly over stock . The game is bouncing between CCDs?

 

[Kaotik]

Hardware Unboxed saw 4.8GHz on the v-cache cores so I'm not sure what's going on there. Looking for more data on how the CCDs actually clock in real use.

Nothing strange, 5 GHz is just stock max boost clock, so practically 1t max clock. 4.8 GHz for all is pretty good

The game is bouncing between CCDs?

That's what it sounds like

 

[homerdog]

Nothing strange, 5 GHz is just stock max boost clock, so practically 1t max clock. 4.8 GHz for all is pretty good

Right. Curious to see if the 7800X3D will have an easier time maintaining 5GHz.

If I had a 7950X3D I would leave it in the "prefer cache" mode. It's not optimal for non-gaming workloads but you don't lose that much in terms of frequency, and there should be way less chance of games running on the cacheless CCD which can be devastating. Still not a zero chance of that happening as evidenced by Mount and Blade.

Also I noticed that Doom Eternal takes a hit with both CCDs active no matter how you set the scheduler. It gains 60fps from disabling CCD1 and also slightly benefits from disabling CCD0 (the vcache one) so it must also be bouncing between CCDs. I wonder why this happens and if they plan to fix it.

 

[Albuquerque]

Also I noticed that Doom Eternal takes a hit with both CCDs active no matter how you set the scheduler. It gains 60fps from disabling CCD1 and also slightly benefits from disabling CCD0 (the vcache one) so it must also be bouncing between CCDs. I wonder why this happens and if they plan to fix it.

It happens because how would the Windows scheduler know to do any different? The only good answer is some sort of binary whitelist where Microsoft somehow keeps track of which apps shouldn't behave that way. They'll eventually fix Doom, which app will be next?

The current paradigm for all these esoteric, non-homogenous cores results in exactly this behavior. We need a better method, and the answer isn't "make it the kernel scheduler's problem."

 

[Bondrewd]

and the answer isn't "make it the kernel scheduler's problem."

oh it really is, nothing else works, period.

 

[hoom]

I had thought the IO die would be the best place for it, how much SRAM could they have fit in there instead of that GPU & displayport IO etc?
Or stick an SBM stack or two in place of one of the CCDs.
Though either would have to be L4 and would mean a bunch of extra CCD -> IO die traffic -> extra power and latency.

 

[homerdog]

It happens because how would the Windows scheduler know to do any different? The only good answer is some sort of binary whitelist where Microsoft somehow keeps track of which apps shouldn't behave that way. They'll eventually fix Doom, which app will be next?

The current paradigm for all these esoteric, non-homogenous cores results in exactly this behavior. We need a better method, and the answer isn't "make it the kernel scheduler's problem."

What will they do to fix Doom? And who will fix it (Microsoft or id)?

 

[arandomguy]

I feel we're in situation in which there isn't a clear "best" CPU even just isolating for gaming. So if that was the expectation it's a bit disappointing.

I originally felt Zen 4 3D parts would open up a clear lead but later on Raptor Lake's higher scaling DDR5 and availability of said DDR5 started to give me some pause which turns out to be the case.

The cache structure though still follows a pattern in that certain games highly benefit from it. 7800X3D will the one to wait for (no wonder they've delayed it) for those that really value those games.

I was looking forward to 7950X3D but I have to say it doesn't meet my expectations.
However, I do think it's the right way to have a lower TDP. Having a small chip using more than 200W is not very healthy.
There are some reports that Intel is aiming for 50% better efficiency (performance per watt) in their next generation CPU (Meteor Lake) which if true is good.

Personally I'm still considering this because I don't think 7800X3D will be faster (it'll likely to be better in value but probably not in performance). Also cost-wise a new MB and new RAM is going to cost a lot and CPU is only part of the equation. To me it's either this or maybe waiting for Meteor Lake.

It's not going to be relevant, all these new uarchs have been offering significant perf/W improvements but actual product power consumption trends the other way. DIY enthusiast targeted products are largely in demand, and therefore the price they can carry, based on their performance. As such the stock product configurations will continue to be tuned for peak performance over efficiency until that changes. Actual lower TDP drops, such as with this, only occur if the chip cannot scale that high with more power (which was a contributing factor to Intel vs AMD back with Zen 3 and earlier in terms of peak TDP differential). So unless Meteor Lake (or whatever is next) and Zen 5 simply don't scale there isn't going to be significant drops in stock power consumption.