Question Zen 6 Speculation Thread

[adroc_thurston]

Any poster on this forum could tell you that there will be no new socket unless there is DDR6 for client, and DDR6 for clients is years away.

They can go for LPDDR6 just fine.

 

[DAPUNISHER]

His off-script statements reflect the sentiment at the time, which was over a year ago - that Zen 6 (Edit: *on AM5) is still (at the time of the video) unconfirmed from AMD. That LGA1851 longevity was a complete unknown (at the time of the video), that Zen 7 was completely unknown, and so that he was not going to use Platform Longevity in his upcoming Zen 5 vs ARL video (which, btw, he hadn't even tested ARL at time of recording).

Anybody who watches HUB regularly knows that he constantly brings up the value of platform longevity in his videos, despite you finding an out of context sentence from a year and a half old video, where he was speaking from a place of unknowns.

That's an excellent summary. Steve was not wrong that longevity becomes hard to promote if the max performance gains over the course of the platform's lifespan are pedestrian.

He did end up on the wrong side of AM4 however. This is how it went between him and budget gamers -

 

[basix]

The usage model for your average DIY PC is gaming.
And sometimes emails.

Office and gaming tasks and maybe some light multimedia stuff, indeed.

But anyways, it is likely that we see now +50% cores with Zen 6 and again +33% cores with Zen 7.
Sure, it was a little bit special to see core count stagnation since Zen 2. But: AMD brought us from Intel's 4/6C to 8C wit Zen 1 and then 16C with Zen 2.
And you can buy a very fast 9950X today and Intel does not really have faster MT parts.

So what can we expect from Zen 6 & Zen 7 as best case compared to a 9950X? 2x and 3x MT performance. So if you need MT performance, AMD will deliver.
Will NVL-S with 52C be faster? Maybe. But we enter a very special and narrow use cases even for enthusiasts. 99% of users will be happy with a 16c 7.0 GHz X3D Zen 7 CPU 😉

I hope that AMD builds high clocking X3D Threadrippers to shut down any further MT arguments 😉
What about 48C Zen 6 and 64C Zen 7 X3Ds? 🙂

 

[Joe NYC]

They can go for LPDDR6 just fine.

Which is actually something I predicted here. Not as a replacement (LPDDR6 replacing DDR5 AM5) but as an additional platform, geared to higher end APUs MiniPCs.

That market is already there, and if memory was in form of LPCAM, it could grow faster.

 

[DAPUNISHER]

I like Steve, I watch his videos, but that one particular appearance and what Steve said was a complete head scratcher.

He was bent with AMD for jerking him around for a week. Forcing him to put in long hours day after day. The pedestrian gains threw into question what happens if Z6 is mid too. Why would you talk about longevity if there is a 20% difference between the start and finish of the socket?

 

[adroc_thurston]

But anyways, it is likely that we see now +50% cores with Zen 6 and again +33% cores with Zen 7.

Stuff you have to pay for, which is why the bulk of Z6 will be sold as 8/10c.
And Zen7 gives you an 8c CCD for mainstream usecases.

Sure, it was a little bit special to see core count stagnation since Zen 2. But: AMD brought us from Intel's 4/6C to 8C wit Zen 1 and then 16C with Zen 2.

It's not special at all. AMD core counts are a byproduct of their server roadmap.

99% of users will be happy with a 16c 7.0 GHz X3D Zen 7 CPU 😉

12c maybe and that's a stretch. 8c really.

Which is actually something I predicted here. Not as a replacement (LPDDR6 replacing DDR5 AM5) but as an additional platform, geared to higher end APUs MiniPCs.

hahahahah no I mean replacing AM5 with an LPDDR6 socket.
NUCs don't need a socket, they can do with standard laptop BGAs just fine.

 

[Joe NYC]

hahahahah no I mean replacing AM5 with an LPDDR6 socket.

AMD already tried the stunt of cutting itself off from mainstream memory. With Rembrandt and Raphael.

It worked out very poorly.

 

[adroc_thurston]

AMD already tried the stunt of cutting itself off from mainstream memory. With Rembrandt and Raphael.

It worked out very poorly.

You do understand that LPDDR is the most mainstream memory there is?

 

[MadRat]

Time to sell on package RAM until prices come down. Expansion RAM on the motherboard.

4 GB for entry level.
16 GB for mainstream.
32 GB for gaming edition.
64 GB for professional.

 

[Joe NYC]

You do understand that LPDDR is the most mainstream memory there is?

NVidia's attempt to use it as a socket LPDDR5 (SOCAMM1) ended with a disaster.

If AMD tried the same, for, say Zen 6, it would have been a disaster 2x the scale of Rembrandt and Raphael disaster. At least DDR5 was available, it only made systems using it uncompetitive because of cost. With SOCAMM unavailable, it would have been worse than the platform being too expensive.

 

[basix]

Stuff you have to pay for, which is why the bulk of Z6 will be sold as 8/10c.
And Zen7 gives you an 8c CCD for mainstream usecases.

12c maybe and that's a stretch. 8c really.

It is clear that you have to pay more for 12/16c. But it is still single CCD. And since Zen 2 this has been the standard go-to option for most people.

That the majority of users will also be happy with 8...10c or 8...12c is also clear. But you have the option to add a few more cores (with spending a little bit more $$) if you want a little more MT oomph.
Currently you are stuck with 8c unless you go dual CCD (which has some drawbacks in games and also regarding power draw).

 

[Thibsie]

Wonder if they’ll do dual CCDs 8-10-12 too. Probably not (hope not).

 

[Joe NYC]

Also remember AM4 got that massive boost because of the 5800X3D. AM5 started with that so it is unlikely to see gains that match AM4, at least in gaming.

AM5 started with 7700x and 9850x3d is already ~35% performance boost.

 

[Joe NYC]

That's an excellent summary. Steve was not wrong that longevity becomes hard to promote if the max performance gains over the course of the platform's lifespan are pedestrian.

Steve made a bad recommendation period.

Excuse that his recommendation was sound, on basis of his prediction for the future - prediction which also turned out to be flawed - is not exactly the best excuse.

He did end up on the wrong side of AM4 however. This is how it went between him and budget gamers -

View attachment 138714

The budget gamers, who got 7600x when AM5 came out, will again come out big winners

 

[Tuna-Fish]

NVidia's attempt to use it as a socket LPDDR5 (SOCAMM1) ended with a disaster.

That's because they built their own shitty module standard. LPDDR6 will come in a JEDEC-specified CAMM2 modules.

 

[inquiss]

Any poster on this forum could tell you that there will be no new socket unless there is DDR6 for client, and DDR6 for clients is years away.

How is that an excuse for Steve giving a bad advice? Because he was not spoon-fed?



Tom told him, in the exchanges that Zen 6 on AM5 was nearly certain, and ARL platform was likely on 1 gen + maybe a refresh.

Again, how much spoon feeding is needed for Steve to make a proper analysis?



I like Steve, I watch his videos, but that one particular appearance and what Steve said was a complete head scratcher. Maybe he was smoking crack or was doing it in half asleep being in the opposite time zone.

That's why that appearance stuck in my memory, because it was so out of norm, and then, when this new video came out today, on platform longevity, it is a complete 180 from that appearance on MLID.

Uh, huh, tell me more, it was STEVE smoking crack you say, uh huh, *writes notes*

 

[Joe NYC]

Uh, huh, tell me more, it was STEVE smoking crack you say, uh huh, *writes notes*

I said smoking crack or half asleep. I have said and done dumb things while sleep deprived, I can relate.

He was not exactly himself, on top of his game, like he usually is.

 

[adroc_thurston]

NVidia's attempt to use it as a socket LPDDR5 (SOCAMM1) ended with a disaster.

Nvidia sucks.

If AMD tried the same, for, say Zen 6, it would have been a disaster 2x the scale of Rembrandt and Raphael disaster. At least DDR5 was available, it only made systems using it uncompetitive because of cost. With SOCAMM unavailable, it would have been worse than the platform being too expensive.

You know LPCAMM exists and shipped a while ago.

 

[Joe NYC]

Nvidia sucks.

You know LPCAMM exists and shipped a while ago.

First, these are LPDDR5.
Second, When Rembrandt and Raphael shipped, DDR5 existed and was shipping.

Unlikely that AMD is going to launch a major new platform exclusively on a Hail Mary assumption about a new memory standard.

But I can see it as a secondary platform geared to APUs, or maybe local AI platform.

 

[adroc_thurston]

Unlikely that AMD is going to launch a major new platform exclusively on a Hail Mary assumption about a new memory standard.

Hail Mary assumption about a phone memory? Are you serious?

 

[MadRat]

If they add memory on the package they can maintain socket compatibility with a second memory controller for mainboard memory. Maybe you add a new bracket for the bigger package, but want to retain compatibility with the socket and previous coolers. Older CPUs still only use maiboard memory on the next generation. Older CPUs might benefit from the newer, optionally wider bracket when it comes to cooling the higher end Zen 5 CPUs. Or they just plunk on the old cooler because the bracket is made for both.

Maybe it makes the performance of integrated APUs possible that turn old competitor low end options obsolete and raises the bar to make Intel ARC/Iris Xe look feeble. Doesn't hurt AMD/ATI stockpiles as hard as the competition because they don't have it to spoil. Done with modest modification of an existing socket, not something requiring costly retooling by vendors.

How much more heat do you generate with 32 GB of LPDDR6 on the package? Seems like we could handle it with an additional 25 mm (approx. an inch) to 40 mm of packaging.

 

[regen1]

Meta and AMD Partner for Longterm AI Infrastructure Agreement

We’re announcing a long-term agreement with AMD to power our AI infrastructure with up to 6GW of AMD Instinct GPUs, helping us build a flexible, resilient tech stack for our AI workloads.

about.fb.com

AMD and Meta Announce Expanded Strategic Partnership to Deploy 6 Gigawatts of AMD GPUs

News Highlights Meta is partnering with AMD to rapidly scale AI infrastructure and accelerate the development and deployment of cutting-edge AI…...

ir.amd.com

Shipments supporting the first gigawatt deployment are scheduled to begin in the second half of 2026 powered by the custom AMD Instinct MI450-based GPU and 6th Gen AMD EPYC™ CPUs, codenamed “Venice,” running ROCm™ software and built on the AMD Helios rack-scale architecture.

Building on deep roadmap alignment, Meta will be a lead customer for 6th Gen AMD EPYC CPUs, codenamed “Venice,” and “Verano,” a next-generation EPYC processor designed with workload-specific optimizations to deliver leadership performance-per-dollar-per-watt.

As part of the agreement, to further align strategic interests, AMD has issued Meta a performance-based warrant for up to 160 million shares of AMD common stock, structured to vest as specific milestones associated with Instinct GPU shipments are achieved. The first tranche vests with the initial 1-gigawatt of shipments, with additional tranches vesting as Meta’s purchases scale to 6 gigawatts. Vesting is further tied to AMD achieving certain stock price thresholds and exercise is tied to Meta achieving key technical and commercial milestones.

 

[adroc_thurston]

If they add memory on the package they can maintain socket compatibility with a second memory controller for mainboard memory. Maybe you add a new bracket for the bigger package, but retain want to retain compatibility with the socket and previous coolers. Older CPUs still only use maiboard memory on the next generation. Older CPUs might benefit from the newer, optionally wider bracket when it comes to cooling the higher end Zen 5 CPUs. Or they just plunk on the old cooler because the bracket is made for both.

How much more heat do you generate with 32 GB of LPDDR6 on the package? Seems like we could handle it with an additional 25 mm (approx. an inch) to 40 mm of packaging.

MoP is nightmarishly margin dilutive and is never ever at AMD.
Forget about iti.

 

[adroc_thurston]

Meta and AMD Partner for Longterm AI Infrastructure Agreement

We’re announcing a long-term agreement with AMD to power our AI infrastructure with up to 6GW of AMD Instinct GPUs, helping us build a flexible, resilient tech stack for our AI workloads.

about.fb.com

AMD and Meta Announce Expanded Strategic Partnership to Deploy 6 Gigawatts of AMD GPUs

News Highlights Meta is partnering with AMD to rapidly scale AI infrastructure and accelerate the development and deployment of cutting-edge AI…...

ir.amd.com

i love venice so much bros...

 

[Joe NYC]

Hail Mary assumption about a phone memory? Are you serious?

Phone memory does not come in removable modules

LPDDR5 came out in CAMM package recently. It is for laptops and its price premium is high over any other memory option.

LPDDR6 is not available even in phones yet.

But you know that...