Question Zen 6 Speculation Thread

[Timorous]

So the problem of ”loading threads fully” apply to Zen6 as well.

You miss the point, possibly intentionally.

If a workload scales well to 20-32 threads then shows diminishing returns Zen 6 will be faster because upto 24 or those threads will be on a full core upto 8 will be on a HT virtual core.

With NVL 16 will be on a p core and upto 16 will be on an e core.

Plenty of workloads don't even scale much beyond 8 threads but more cores means you can potentially run multiple instances of such workloads at the same time. With NVL you can run 2 before you start using E cores. With Zen 6 you can run 3 before you are onto HT threads.

There will be plenty of cases where Zen 6 is just better because it's performance is more consistent as you load up more threads.

 

[511]

You miss the point, possibly intentionally.

If a workload scales well to 20-32 threads then shows diminishing returns Zen 6 will be faster because upto 24 or those threads will be on a full core upto 8 will be on a HT virtual core.

With NVL 16 will be on a p core and upto 16 will be on an e core.

On NVL there are no Virtual core and the E cores are very good since ARL they are the only thing carrying ARL 🤣.

Plenty of workloads don't even scale much beyond 8 threads but more cores means you can potentially run multiple instances of such workloads at the same time. With NVL you can run 2 before you start using E cores. With Zen 6 you can run 3 before you are onto HT threads.

There will be plenty of cases where Zen 6 is just better because it's performance is more consistent as you load up more threads.

Yeah we will see in the benchmarks in H2 26.

 

[MS_AT]

Going from 5600 to 8000 I am figuring at 43%. Did I get that right?

I was using 6000MT/s as baseline but otherwise I think your math is correct.

Do you disagree?

I have already answered that. My issue was with "just fine" you used previously. That was giving the impression that it would fix the issue for people who are unaware of the memory gap and that the cores in principle could still be considered starved but just less so.

Now to rephrase, I think first thing they should do is increase the BW between CCD and IOD, that is actually more important for single CCD skus, and would move the bottleneck back to memory for both types of skus. And whatever mem speed boost they give is a nice bonus on top. You mention IOD improvements but you never say what kind of improvements you do expect.

It works perfectly fine on other big/little core CPUs, and it’ll work fine on AMD Zen6 too, unless someone screws up the OS scheduler code.

I just gave a real world example when it does not work perfectly on Intel CPUs, in the past I was linking to years old bugs that are still not resolved where the scheduler gets it wrong.

Not to mention your statement is contradictory as you are unlikely to run that CPU without an OS, the scheduler is the essential part of the experience and frankly nobody cares who screwed it up, just that without hybrid cores the issues is not there.

 

[511]

I just gave a real world example when it does not work perfectly on Intel CPUs, in the past I was linking to years old bugs that are still not resolved where the scheduler gets it wrong.

Not to mention your statement is contradictory as you are unlikely to run that CPU without an OS, the scheduler is the essential part of the experience and frankly nobody cares who screwed it up, just that without hybrid cores the issues is not there.

Just on this topic windows just scheduled my video game on LP-E core it started to feel a bit laggy and than i changed the affinity and it worked fine yesterday.

I want to Thank Microsoft for their awesome scheduler.

 

[eek2121]

Just on this topic windows just scheduled my video game on LP-E core it started to feel a bit laggy and than i changed the affinity and it worked fine yesterday.

I want to Thank Microsoft for their awesome scheduler.

I have issues playing Diablo 2: Resurrected thanks to Microsoft’s current incompetence (related to the “game bar” in this case, which they have basically made mandatory). Don’t even get me started on their scheduler.

 

[gdansk]

I have issues playing Diablo 2: Resurrected thanks to Microsoft’s current incompetence (related to the “game bar” in this case, which they have basically made mandatory). Don’t even get me started on their scheduler.

I'm curious, what's the problem? Stutters? I've been playing a lot of D2:R lately on an 9950X3D W10 system. That can also have automatic scheduling problems, but in either case it's playable.

 

[Joe NYC]

Wow. I guess I have to consider the source here .

Datacenter is higher profit and is growing faster than any other segment. AMD is also growing faster in DC than any other segment.

I believe that this is the most important contributor to AMD's recent success.

Desktop in general and DIY in specific are decreasing markets. While its great that AMD has raised their ASP in desktop, they are focusing on DC:

https://www.reddit.com/r/hardware/comments/1fg0g4o/ceo_lisa_su_says_amd_is_a_data_centerfirst

You are going back to mixing datacenter GPU and CPU. Speaking of "datacenter" in general makes the comparison skewed to GPU.

Your quote is from 9 months ago, at which time, AMD fortunes in GPU business still appeared to be going to the moon. In subsequent 9 months, those fortuned have somewhat soured for AMD and datacenter GPU has been underperforming for AMD.

My post, about gains in desktop and increasing ASPs in client are more recent, from the most recent quarterly earnings report.

I doubt that workstation people, in general, care about gaming on PC's (or at all).

Then you would be wrong. When my daughter entered college studying interior design (closely adjacent to architecture), they were told to buy a "gaming notebook", which is what gives you closest thing to graphics design workstation on a modest budget.

 

[Joe NYC]

It's not wishful thinking it's a leak from Raichu 🙂

https://twitter.com/x/status/1922618865948782720

For people who don't follow him.


It is for new products lol.

View attachment 125112

I am still confused (and so is Raichu, it seems) how this high end NVL is supposed to be configured:
- two identical 8p 16e tiles with big LLC on a separate tile
- one 8p 16e with small L3, 2nd one with big LLC

Edit: NVL SKUs are new products, so 50% GM should apply to them.

 

[Joe NYC]

And this is any different between TR and NVL-S or any other CPU how exactly?

It TR and some potential high core count NVL are different from more typical client system in that the higher core count and higher price deliver close to zero performance improvement.

Meaning, performance / $ is closer and closer to zero with each additional core (thread) and the money spent on each additional core (thread) is closer and closer to flushing money down the toilet.

Also, it’s actually any additional thread beyond 1, assuming you’re talking about Amdahl’s law. But I don’t think you recommend we should only use 1C CPUs.

There gains, decreasing gains which are greater than zero in client applications. But you are quickly starting to approach zero gains past ~16 threads.

 

[511]

Edit: NVL SKUs are new products, so 50% GM should apply to them.

They said new products and Nova Lake is not a new product it's middle of cycle inside Intel

 

[eek2121]

I'm curious, what's the problem? Stutters? I've been playing a lot of D2:R lately on an 9950X3D W10 system. That can also have automatic scheduling problems, but in either case it's playable.

It locks up for 5-20 seconds at a time, eventually hard locking the system. If I reset the settings of the game bar, it is temporarily fixed…only to break the next time I launch.

Thankfully, I have other machines to play it on. They just aren’t as beefy (Steam Deck, laptop, mini-PC)

I haven’t figured out any other workaround. Some on reddit had the same issue, but their workarounds didn’t work for me on Win112H24.

 

[Thibsie]

It locks up for 5-20 seconds at a time, eventually hard locking the system. If I reset the settings of the game bar, it is temporarily fixed…only to break the next time I launch.

Thankfully, I have other machines to play it on. They just aren’t as beefy (Steam Deck, laptop, mini-PC)

I haven’t figured out any other workaround. Some on reddit had the same issue, but their workarounds didn’t work for me on Win112H24.

Can't kill the game bar service?

 

[OneEng2]

Then you would be wrong. When my daughter entered college studying interior design (closely adjacent to architecture), they were told to buy a "gaming notebook", which is what gives you closest thing to graphics design workstation on a modest budget.

This is not causal at all. The purpose of using a "gaming notebook" for interior design programs is simply another way of saying "High graphics capability" and "Higher performance" laptop. It had nothing to do with gaming.

I also suspect that when your daughter is done being a student and is actually a professional, she will be provided a workstation class computer for doing her work (a real workstation).

 

[inquiss]

This is not causal at all. The purpose of using a "gaming notebook" for interior design programs is simply another way of saying "High graphics capability" and "Higher performance" laptop. It had nothing to do with gaming.

I also suspect that when your daughter is done being a student and is actually a professional, she will be provided a workstation class computer for doing her work (a real workstation).

Yeah. It's really niche. True story of course, but the percentage of home users needing a cheaper equivalent to a graphics workstation is a rounding error.

When she can make money with her profession (and I wish her all the best in that), then we're into paying real money for real machines

 

[Thunder 57]

I really wanted to crap on this thread but I didn't, because I'd get a vacation. I'd like one to Norway please.

 

[Thunder 57]

I really wanted to crap on this thread but I didn't, because I'd get a vacation. I'd like one to Norway please.

Not because of you 511. I think we all know who though.

 

[igor_kavinski]

I am still confused (and so is Raichu, it seems) how this high end NVL is supposed to be configured:
- two identical 8p 16e tiles with big LLC on a separate tile
- one 8p 16e with small L3, 2nd one with big LLC

Ideally, the tile with the bLLC should only have P-cores. Forcing them to mingle with the peasant cores on the same ring is bad for their morale.

 

[fastandfurious6]

when i bought the 8250u it was not crap lol i bought it in 2019 i think.

oh believe me it was total crap, skylake U was so slower than ivy bridge m despite more cores.... 5 generations apart lolllll

My interpretation of Zen 6 leaks is that there may be a Medusa Halo(ish) with Zen 6 cores and Strix Halo (LPDDR5x) IOD, and a real Medusa Halo, with new IOD supporting LPDDR6.

is it economically viable if properly scheduled?
get medusa v1 with ddr5 first out of course because quicker time to market, then release medusa halo v2 with ddr6 when its ready

 

[fastandfurious6]

dupe delete

 

[LightningZ71]

While my experience was more with Whiskey Lake than with Kaby Lake R, I have to say that, for office use, not many people complained about them in production. They did what they needed to, and as long as they were configured with SSDs, they were fast enough.

Medusa Halo with the Strix Halo IOD should be a decent enough upgrade. There should be more than enough memory bandwidth to go around on computationally heavy tasks to allow the new cores to stretch their legs.

 

[Joe NYC]

Yeah. It's really niche. True story of course, but the percentage of home users needing a cheaper equivalent to a graphics workstation is a rounding error.

When she can make money with her profession (and I wish her all the best in that), then we're into paying real money for real machines

She actually just graduated and has been interning at a very good design firm in NYC, that is hopefully leading to a full-time job.

Most of the interns don't do a lot of design work, so the companies have a "hand me down" approach with laptops for Interns. They do have a remote desktop capability to log into more powerful workstations / servers.

When the company found out my daughter can do real design work, they put her in with full time junior and senior designers, and that hand me down laptop is not quite cutting it. In the meantime, she sorted out the software licensing last night, and she can use my home desktop (in sig) when she works remotely.

 

[Joe NYC]

Ideally, the tile with the bLLC should only have P-cores. Forcing them to mingle with the peasant cores on the same ring is bad for their morale.

That would be ideal, but that would mean they would have to tape out 2 separate dies (one with all P cores and big L3 and one with all Atom cores), these dies will not be useable anywhere else. So, it would be a vanity project.

 

[Joe NYC]

is it economically viable if properly scheduled?
get medusa v1 with ddr5 first out of course because quicker time to market, then release medusa halo v2 with ddr6 when its ready

Well, a lot depend how backward and forward compatible the new Strix Halo die-to-die link is.

If it is fully compatible, then the design cost is extremely well utilized. Existing Strix Halo IOD would be paired with the universal Zen 6 CCD (no extra design cost).

And then, the design resources are put into a next gen IOD with LPDDR6, new process node (N3P?), new version of RDNA (5?).

In MLID leaks, he had Medusa Halo with 2 sockets - FP10 (Strix Halo compatible) and FP11 (new). So that would be the supporting argument for this.

 

[511]

is it economically viable if properly scheduled?
get medusa v1 with ddr5 first out of course because quicker time to market, then release medusa halo v2 with ddr6 when its ready

Where is the leak for Medusa halo ddr6 it's DDR5 afaik

 

[Tigerick]

Where is the leak for Medusa halo ddr6 it's DDR5 afaik

It's actually from MLID . Combination of 256-bit LPDDR5x and 384-bit LPDDR6 seems odd at first with extra 128-pin just reserved for LPDDR6 if LPDDR5x is being used. However, if you count in the 48CU, then it makes sense for AMD to support LPDDR6 especially with Medusa Halo launching in H22027. By that time, LPDDR6 should be widely available, then supporting LPDDR5x is more like backup plan.

I have created a table in the frontpage of Soundwave page to include Medusa Little Halo and Medusa Point R9, kind of make sense to have Little Halo, isn't it?