Question Zen 6 Speculation Thread

[Joe NYC]

i would assume many other supply chain issues before wafers , packaging , HBM etc

Just contemplating of what it would take run into N2 supply issues that could only arise if the actual demand far exceeds the number of wafers ordered.

A lot of people ignore the fact that when AMD placed order for N2 and TSMC confirmed it, those wafers are as good as money in the bank, independent of what else is taking place in the industry with other players.

 

[MS_AT]

They care about DIY very much; they own the place and will collect hearthy margins there.

You haven't commented directly about the delay. I mean you seem to give the impression it might still launch this year, but without taking explicit stance. Or maybe I am reading too much or too little into this

 

[Joe NYC]

wut?
2-4 years to what

When AMD handed Alder Lake undisputed performance leadership over Zen 3 (rather than a tie vs. Zen 3 V-Cache) that loss of Halo effect persisted until Zen 5 V-Cache.

That's because there were some cases where Zen 4 V-Cache did not build Raptor Lake, and Intel fans claimed that those CPUs were tied.

Only Zen 5 V-Cache reclaimed undisputed leadership, the same as what AMD gave away to Alder Lake.

 

[adroc_thurston]

Or maybe I am reading too much or too little into this

Yeah that's definitely this.

 

[adroc_thurston]

When AMD handed Alder Lake undisputed performance leadership over Zen 3 (rather than a tie vs. Zen 3 V-Cache) that loss of Halo effect persisted until Zen 5 V-Cache.

That's because there were some cases where Zen 4 V-Cache did not build Raptor Lake, and Intel fans claimed that those CPUs were tied.

Only Zen 5 V-Cache reclaimed undisputed leadership, the same as what AMD gave away to Alder Lake.

This has nothing to do with this babble and everything to do with RPL being good nuff in gaming.
Intel just got hit with a double whammy of RPL suicides and ARL-S being poo in gaming.

 

[eek2121]

You haven't commented directly about the delay. I mean you seem to give the impression it might still launch this year, but without taking explicit stance. Or maybe I am reading too much or too little into this

OOH, I WILL!!!!

THERE IS NO DELAY.

There, happy? Things are on track to launch just as they are shown on the roadmap.

 

[Joe NYC]

Both Zen 6 desktop and Nova Lake are 2027?

I am not buying either.

https://twitter.com/x/status/2025014474986717447

 

[poke01]

Both Zen 6 desktop and Nova Lake are 2027?

I am not buying either.

https://twitter.com/x/status/2025014474986717447

Nova Lake S was 2027. NVL-H is late 2026.

Golden Pig IS accurate

 

[OneEng2]

I suspect NVL bLLC will be competitive with 9800X3D so to give Intel a few months of being on top in 1T, MT, and good enough in gaming would be hilarious self-sabotage.

I think bLLC will be better; however, the latency will not be close to X3D in AMD's designs. I don't believe NVL with bLLC will even best 9800X3D rather on 10800X3D. Just my speculation.

Huh? Zen 6 does not seem to be a big bump - it is more like in Intel's territory of ~10% IPC.

I guess nobody expects Zen 7 "soon", as before 2028/2029.

I think IPC will be north of 10%. Perhaps 12-14%; however, clocks will be in the 6.2-6.5Ghiz region so when combined with the IPC increase, it will be a pretty significant bump. At least that is how I am reading the tea leaves.

 

[Fjodor2001]

I think bLLC will be better; however, the latency will not be close to X3D in AMD's designs. I don't believe NVL with bLLC will even best 9800X3D rather on 10800X3D. Just my speculation.

It’s a big and thus expensive slab of silicon, so to be worth it I guess it must be close.

But I also wonder how big latency difference can be expected. And what the net result difference will be in gaming performance.

 

[DrMrLordX]

I think bLLC will be better

Yeah but why compete with an old part like that? All they're doing is setting themselves up to get smashed by Zen6.

 

[adroc_thurston]

Yeah but why compete with an old part like that? All they're doing is setting themselves up to get smashed by Zen6.

Because right now they're 20-30% behind in gaming.

 

[coercitiv]

Yeah but why compete with an old part like that? All they're doing is setting themselves up to get smashed by Zen6.

Don't underestimate NVL bLLC when it comes to gaming, even Raptor Cove + big cache would likely get close enough to Zen5 3D. If Intel fixes their uncore and cache latency (big IF, but still) they should be competitive with Zen6 3D.

 

[adroc_thurston]

even Raptor Cove + big cache would likely get close enough

Intel no longer has IMC on the ring.

If Intel fixes their uncore and cache latency (big IF, but still) they should be competitive with Zen6 3D.

This only works under a bogus assumption of 0% fmax gain for z6.

 

[LightningZ71]

To me, the only wildcard for NVL is the paired P cores sharing a larger block of L2. If they can keep the latency of that L2 decent, get better Fmax on the P cores and also increase the average instruction throughput per clock of the core itself, they'll win a few benchmarks that spill the L2 on Zen5/6 heavily. Beyond that, I remain to be convinced that NVL is going to be significantly faster than the seppuku prone RPL K parts.

 

[OneEng2]

There are certainly lots of "ifs" about how competitive a NVL with bLLC might be with ZEN6 X3D or even 9800 X3D.

I tend to believe that AMD simply has a huge lead in technology like chiplets, and stacked L3, and that it is unlikely that the huge gap can be overcome by Intel by the next attempt.

I also believe that Intel is STILL being weighed down by its fab work and the old philosophy of win by brute force through fab superiority.

But mostly I believe that AMD simply has a very good strategy in its server first design. They are really racking up the margins year after year. They also have the momentum gained over the good decision making done a decade back.

I expect NVL to be much better than we believe, but ZEN 6 to be even more so.

 

[adroc_thurston]

I tend to believe that AMD simply has a huge lead in technology like chiplets, and stacked L3

They have much better cores and a much better L3.

I also believe that Intel is STILL being weighed down by its fab work and the old philosophy of win by brute force through fab superiority.

Their CPU IP flat out sucks. It's very simple.

I expect NVL to be much better than we believe, but ZEN 6 to be even more so.

NVL sucks.
Zen6 tbd.

 

[Fjodor2001]

I expect NVL to be much better than we believe, but ZEN 6 to be even more so.

Are you talking about gaming specifically, or also for other use cases including MT workloads?

 

[Doug S]

Just contemplating of what it would take run into N2 supply issues that could only arise if the actual demand far exceeds the number of wafers ordered.

A lot of people ignore the fact that when AMD placed order for N2 and TSMC confirmed it, those wafers are as good as money in the bank, independent of what else is taking place in the industry with other players.

AMD has a certain amount of confirmed N2 wafers, and no guarantees they will get any more than what they already have confirmed. So yes those are like money in the bank, but like money in the bank they'll want to invest it what makes them the most profit.

If they see enough demand from the datacenter market they could easily decide to allocate nearly all of those wafers there, leaving precious little for Zen 6 PC/laptop parts. They might trickle some out at the very high end / high profit range, but unless the AI bubble bursts this year I can't see them making a big launch with wide availability. If they can earn x dollars of profit per wafer making Zen 6 PC chips or 2x, 3x, 5x who knows as much profit per wafer making AI/Epyc chips it would be financially stupid to do the former.

They can keep selling Zen 5, or port Zen 6 to N3 family processes to avoid wasting that "money in the bank" making peanuts by comparison to what they could be making. If the AI bubble keeps inflating DRAM/NAND will continue to get more expensive, causing PC demand to crater so it might not matter that much anyway.

 

[Hulk]

Do you think this article right? 2027 for Zen 6/Nova Lake?

https://www.tomshardware.com/pc-components/cpus/amd-zen-6-and-intel-nova-lake-cpus-reportedly-arriving-late-delayed-to-ces-2027-next-gen-chips-rocked-by-industry-turmoil

 

[adroc_thurston]

or port Zen 6 to N3 family processes

N3 is a busier node.

 

[OneEng2]

Are you talking about gaming specifically, or also for other use cases including MT workloads?

Both I think. I believe that ARL was critically hamstrung by the high latency of off-chip interconnects. Zen 5 X3D was very very good with respect to latency.

It isn't hard to imagine that Intel (like AMD before it) had problems reconciling the architecture around off-chip resources. If that was able to be fixed (even a little bit) it could result in some big wins across a broad range of applications.

Zen 5 has long since (a couple of generations) found ways to deal with off chip latency ...... AND has perfected chip stacking which provides much lower latency access to L3 that Intel's bLLC is every likely to achieve (physics).

So where I am coming from is that ARL->NVL there is great opportunity for improvement. Zen 5-6 will likely find most of its gains from faster clock speeds made possible by the shrink from N4P to N2. Intel on the other hand will likely get much less of a boost moving from N3B to either N2 or 18A .... and I have great concerns about the fmax for 18A.

I think that NVL 52c will dominate Zen 6 24c in MT; however, it may be a moot point since at its price point it may well be up against thread ripper which, IMO, will beat it badly in highly MT loads that people are likely to shell out that kind of money to run.

Do you think this article right? 2027 for Zen 6/Nova Lake?

https://www.tomshardware.com/pc-components/cpus/amd-zen-6-and-intel-nova-lake-cpus-reportedly-arriving-late-delayed-to-ces-2027-next-gen-chips-rocked-by-industry-turmoil

I think that AMD has little reason to rush. Zen 5 already dominates ARL. I think it more likely that AMD will simply not release Zen 6 DT until Intel manages NVL..... and I suspect Intel is having yield issues with NVL pushing back their release date. I think AMD is just waiting ..... and raking in the profit from Zen 5 every day that Intel doesn't have an answer.

I think you and I both remember a time when Intel did the same exact thing ..... for like 10 years!

Lack of competition has shown itself to be VERY bad for consumers. Let's all hope Intel gets their act together or we aren't going to see Zen 6 for quite a while.

 

[adroc_thurston]

I believe that ARL was critically hamstrung by the high latency of off-chip interconnects

LNC just plain sucks.
L3 perf was also ass. Many layers to that onion.

So where I am coming from is that ARL->NVL there is great opportunity for improvement

0% fmax says hello.

 

[Fjodor2001]

I think that AMD has little reason to rush. Zen 5 already dominates ARL. I think it more likely that AMD will simply not release Zen 6 DT until Intel manages NVL..... and I suspect Intel is having yield issues with NVL pushing back their release date. I think AMD is just waiting ..... and raking in the profit from Zen 5 every day that Intel doesn't have an answer.

I think you and I both remember a time when Intel did the same exact thing ..... for like 10 years!

2 * Only the paranoid survive.

 

[coercitiv]

I believe that ARL was critically hamstrung by the high latency of off-chip interconnects.

The added tile latency combined with moving the IMC away from the compute tile probably did most of the damage, but ARL also has weird problems with cache latency. That's why I argued NVL might register a big jump in gaming, that big cache version can hide the IMC mess as long as Intel fixes their L3 (relative to size ofc).

Zen 5-6 will likely find most of its gains from faster clock speeds made possible by the shrink from N4P to N2.

AFAIK Zen 6 makes use of the advanced packaging seen for Strix Halo, so it should also see significant improvements in latency and power. A great recap of the packaging used in Halo can be found here.