Question Zen 6 Speculation Thread

[Fjodor2001]

My take, Zen6 will win everything thats not embarrassing parallel and Nova Lake will comsume 350-400w when tuned to actually have decent cinebench performance

Regarding how much NVL-S will consume, it depends on how far you push the frequency on the cores.

Could be that they are not intending to push the cores to max frequency when running all cores in parallell. And when not pushing the cores to max frequency the Intel cores in 285K perform very well in perf/watt, and in fact better than Zen.

As posted a while ago from this source:

 

[Io Magnesso]

It's true that a chiplet moves multiple chips as a single chip, but The goals that you achieve with a chiplet are different.
It's fun to see different design ideas

 

[Fjodor2001]

if the rumored 8P16E + 8P16E + 4LPE(SOC) on two different compute dies are true, making the windows scheduling doomed to fail

Why would it be doomed to fail? Sure having more compute dies has a (usually quite small) scheduling penalty, but both AMD and Intel will have that since both Zen6 and NVL-S will have two compute dies on DT.

Also, scheduling is working fine in current CPU generations, despite having multiple compute dies too, so I don’t see why it should be doomed now all of a sudden in the next CPU generations.

 

[adroc_thurston]

And when not pushing the cores to max frequency the Intel cores in 285K perform very well in perf/watt, and in fact better than Zen.

Put a Strix halo onto datboi chart and see wonderful miracles happen.

 

[511]

Put a Strix halo onto datboi chart and see wonderful miracles happen.

It doesn't tbh vs 275HX can't find one with higher wattage

 

[DrMrLordX]

My take, Zen6 will win everything thats not embarrassing parallel and Nova Lake will comsume 350-400w when tuned to actually have decent cinebench performance

With AMD launching Threadripper 9xxx parts next month, how many people are seriously going to wait for a 400w Nova Lake-S? That part will probably be vaporware and will lose to 64c Threadripper, not to speak of the 96c PRO part.

 

[basix]

It doesn't tbh vs 275HX can't find one with higher wattage

Well you can see a few things there:
- Strix Halo closes the gap at 40-50W (where desktop Zen was far away due to the IFOP -> IOD)
- If you would extend the measurements to lower power draw, Strix Halo would beat Arrow Lake
- Strix halo would probably beat Arrow Lake at 100W and above

That Arrow Lake scales higher between 50-90W is primarily a core count thing. Not much more.

I expect, that Zen 6 will look quiet differently in this graph. Nova Lake will be interesting though, because of the huge number of cores.

 

[Abwx]

Regarding how much NVL-S will consume, it depends on how far you push the frequency on the cores.

Could be that they are not intending to push the cores to max frequency when running all cores in parallell. And when not pushing the cores to max frequency the Intel cores in 285K perform very well in perf/watt, and in fact better than Zen.

As posted a while ago from this source:

Only in CB 2024, in Handbrake it consume 30% more for lower perf, 263W vs 196W.

AMD Ryzen 7 9800X3D vs. 7800X3D, 5800X3D & Core Ultra im Test: Leistungsaufnahme und Temperaturen

AMD Ryzen 7 9800X3D im Test: Leistungsaufnahme und Temperaturen / Leistungsaufnahme auf dem Windows-Desktop

www.computerbase.de

 

[Fjodor2001]

With AMD launching Threadripper 9xxx parts next month, how many people are seriously going to wait for a 400w Nova Lake-S? That part will probably be vaporware and will lose to 64c Threadripper, not to speak of the 96c PRO part.

It’s 150W TDP and much, much, cheaper than the expensive TR platform for corresponding level of MT perf. So yes, likely a lot of people will wait for NVL-S.

 

[Fjodor2001]

Only in CB 2024, in Handbrake it consume 30% more for lower perf, 263W vs 196W.

AMD Ryzen 7 9800X3D vs. 7800X3D, 5800X3D & Core Ultra im Test: Leistungsaufnahme und Temperaturen

AMD Ryzen 7 9800X3D im Test: Leistungsaufnahme und Temperaturen / Leistungsaufnahme auf dem Windows-Desktop

www.computerbase.de

We’re talking about different things.

I was talking about perf/watt when not pushing the cores to max perf (or max TDP).

 

[Det0x]

Why would it be doomed to fail? Sure having more compute dies has a (usually quite small) scheduling penalty, but both AMD and Intel will have that since both Zen6 and NVL-S will have two compute dies on DT.

Also, scheduling is working fine in current CPU generations, despite having multiple compute dies too, so I don’t see why it should be doomed now all of a sudden in the next CPU generations.

Windows already can have scheduling problems with current gen, on both AMD and Intel, and this is a much easier sheme since its either all P-core on two different dies (Zen5) or a mix og P and E cores on a single die (Arrow Lake)

Nova Lake will take the issue to a other level with a mix of 3 different cores, intertwined on 3 different dies..

How do you suggest Windows should schedul a 12 thread gaming-workload with lots of crosstalk between the cores ?

Put all the work on a single compute-die and use 8P + 4 Ecores ? (Since Ecores have lower performance than Pcores you get lower performance)
Or split the work up between two compute-dies to only run on the Pcores ? (will get same crosstalk/latency issues AMD have been fighting for years and have pretty much solved at this point)

My main point is that Nova Lake will now pay the dual chiplet/compute-die tax which they avoided all the way up until now and "fanboys" have been criticizing AMD for.
While the Zen team almost have 10 years experience dealing with this now..

So i kinda fear we will see many problems with the 8P16E + 8P16E + 4LPE splitup since this is time they have to face scheduling workloads to different compute-dies on a highend desktop K-SKU, in recent times, and we already know how it went with Arrow Lake with its much easier chiplet architecture compared to Nova Lake (latency and gaming performance is very bad)

 

[Fjodor2001]

Well you can see a few things there:
- Strix Halo closes the gap at 40-50W (where desktop Zen was far away due to the IFOP -> IOD)
- If you would extend the measurements to lower power draw, Strix Halo would beat Arrow Lake
- Strix halo would probably beat Arrow Lake at 100W and above

That Arrow Lake scales higher between 50-90W is primarily a core count thing. Not much more.

I expect, that Zen 6 will look quiet differently in this graph. Nova Lake will be interesting though, because of the huge number of cores.

Yes, what's interesting is that Arrow Lake performs better than Zen5 w.r.t. perf/watt in the lower power consumption range, and vice versa in the highest range.

Now we have this:
ARL-S: 125W TDP and 8P+16E.
NVL-S: 150W TDP and 16P+32E.

So much lower TDP per core for NVL-S when all cores are used.

This means NVL-S must stay in the lower power consumption range per core, which is where it perform at it's best w.r.t. perf/watt.
(And going beyond that gains very little extra perf anyway, for a lot of additional power consumption.)

In short: Assuming the above will be the case, then perf/watt in MT loads should be very good for NVL-S.

 

[Io Magnesso]

It's just because Windows scheduling is poor…
There is a high possibility that a processor with heterogeneous cores will become more mainstream in the future…

 

[Io Magnesso]

It's just because Windows scheduling is poor…
There is a high possibility that a processor with heterogeneous cores will become more mainstream in the future…

Even AMD has a story that the next Zen6 laptop CPU will have a smaller LP core than the Zen "C" core.
Every AMD is installed in many models from the Zen5 laptop CPU to the Zen C. The weak Windows scheduling is a bad thing for AMD as well...

 

[StefanR5R]

Nova Lake will now pay the dual chiplet/compute-die tax which they avoided all the way up until now and "fanboys" have been criticizing AMD for.

Will NVL have split last level caches? Or will it have a unified LLC like Xeons do?

Windows already can have scheduling problems with current gen, on both AMD and Intel,

For forum members to whom the world of multitasking/multithreading consists of Cinebench and nothing else, scheduling problems do not exist.

It's just because Windows scheduling is poor…

Linux' CPU scheduler is pretty good but not omniscient either. It just can't be.

 

[CouncilorIrissa]

Yes, what's interesting is that Arrow Lake performs better than Zen5 w.r.t. perf/watt in the lower power consumption range, and vice versa in the highest range.

Now we have this:
ARL-S: 125W TDP and 8P+16E.
NVL-S: 150W TDP and 16P+32E.

So much lower TDP per core for NVL-S when all cores are used.

This means NVL-S must stay in the lower power consumption range per core, which is where it perform at it's best w.r.t. perf/watt.
(And going beyond that gains very little extra perf anyway, for a lot of additional power consumption.)

This has little to do with cores lol and everything to do with AMD paying a constant 20-25W penalty due to an antiquated IOD (which they will no longer be paying with Zen 6: looking at you, Strix Halo). The lower the total power, the bigger the relative penalty from the IOD (because it's pulling 20W of power no matter what. So when limited to 100W, only 80% of power are going to the cores, whereas at 200W, 90% are used by the cores)

 

[gdansk]

I just can't imagine anyone but the cinebench graphers buying a 48 core CPU on 2 channel memory in 2026/7.

Unless Intel does some absurd segmentation causing it to also have the best 1T performance there's no reason to buy it for almost all their customers.

Buy a big chip that will trip over itself in bus contention in any real workload. Or buy half the chip and get a part that performs just as well in every workload but cinememe? I do not see the value.

I also think AMD will have to play stupid segmentation games if they want people to buy more than a 1 x 12 core part.

 

[poke01]

Will mobile zen 6 be 6GHz too? Or more like 5.5GHZ?

 

[Kepler_L2]

Will mobile zen 6 be 6GHz too? Or more like 5.5GHZ?

All the N2P stuff should be over 6GHz, dunno about N3P versions of Zen6.

 

[DavidC1]

Yes, what's interesting is that Arrow Lake performs better than Zen5 w.r.t. perf/watt in the lower power consumption range, and vice versa in the highest range.

That says nothing about Novalake being better than Zen 6.

Because you could have a 50W 16+32 chip just by clocking it low enough. But your performance will suck.

 

[CouncilorIrissa]

I just can't imagine anyone but the cinebench graphers buying a 48 core CPU on 2 channel memory in 2026/7.

Unless Intel does some absurd segmentation causing it to also have the best 1T performance there's no reason to buy it for almost all their customers.

Buy a big chip that will trip over itself in bus contention in any real workload. Or buy half the chip and get a part that performs just as well in every workload but cinememe? I do not see the value.

I also think AMD will have to play stupid segmentation games if they want people to buy more than a 1 x 12 core part.

day 9592 of anandtech posters pretending there's a mythical workload that scales infinitely with cores, while not giving a crap about membw AND not needing to be correct (aka requiring ECC from workstation/HEDT boards).

 

[Timorous]

I just can't imagine anyone but the cinebench graphers buying a 48 core CPU on 2 channel memory in 2026/7.

Unless Intel does some absurd segmentation causing it to also have the best 1T performance there's no reason to buy it for almost all their customers.

Buy a big chip that will trip over itself in bus contention in any real workload. Or buy half the chip and get a part that performs just as well in every workload but cinememe? I do not see the value.

I also think AMD will have to play stupid segmentation games if they want people to buy more than a 1 x 12 core part.

I think Intel are doing it because they want a win in something at least.

AMD will do it because people are accustomed to them offering dual CCD parts and AMD will make a small amount for that niche.

I think in all likelihood the 12core X3D part will be the best seller in DIY, especially as it should be a drop in upgrade for Zen 4 or 5 users.

 

[Io Magnesso]

All the N2P stuff should be over 6GHz, dunno about N3P versions of Zen6.

You can tell if it's for desktop…
Isn't it difficult to get 6GHz on a laptop no matter what?

 

[Thunder 57]

I think Intel are doing it because they want a win in something at least.

AMD will do it because people are accustomed to them offering dual CCD parts and AMD will make a small amount for that niche.

I think in all likelihood the 12core X3D part will be the best seller in DIY, especially as it should be a drop in upgrade for Zen 4 or 5 users.

Probably too expensive. I'd think an 8 core X3D would sell better.

 

[Io Magnesso]

Probably too expensive. I'd think an 8 core X3D would sell better.

Compared to the 12-core X3D?