Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)

[DisEnchantment]

Speculate at will

 

[Nothingness]

I couldn't get any configuration of prime95 to draw 162W PPT. I also tried ycruncher with AVX 512 enabled, blender benchmark, cinebenches. I am running through these workloads again to make sure I'm not just misremembering.

Edit: Confirmed, the maximum PPT I can get under any of these workloads is about 145W. Most are 125-135W. No thermal, power, or current limiting hit at any point.

Thanks a lot for checking, much appreciated 😀

 

[Hail The Brain Slug]

Thanks a lot for checking, much appreciated 😀

Under the same conditions my 7950X would routinely run in the 220W range, hitting against the 230W limit frequently. Enabling PBO let it spill over into the 230's and occasionally, 240's. All at stock otherwise (no CO, no frequency offset)

I think some people don't realize just how lean AMD tuned the V/F curve on the 7950X3D.

 

[Nothingness]

Under the same conditions my 7950X would routinely run in the 220W range, hitting against the 230W limit frequently. Enabling PBO let it spill over into the 230's and occasionally, 240's. All at stock otherwise (no CO, no frequency offset)

I think some people don't realize just how lean AMD tuned the V/F curve on the 7950X3D.

That's exactly the reason why I don't want to get an X chip but an X3D. I hope Zen5 X3D will be similarly power efficient.

 

[Hail The Brain Slug]

That's exactly the reason why I don't want to get an X chip but an X3D. I hope Zen5 X3D will be similarly power efficient.

I am in the same boat. Especially since it seems gaming performance won't match previous gen X3D, there's not a great inscentive to jump from a 7950X3D to a 9950X when the 9950X3D seems to be just around the corner.

 

[eek2121]

If you know of a workload that will get my 7950X3D to actually hit its PPT cap, please tell me. I tried several different workloads and found nothing that took it meaningfully beyond around 140 watts. I've got it on water so it is not clocking down due to thermal limits.

yeah, don't think you'll be able to do it with Prime95. and dual the dual CCD X3D chip is trickier because of the differences in power/voltages.

Actually, if you run Prime95 on CCD1 and something else on CCD0 that could probably get it close.

EDIT: If I still had access to an X3D chip I would play with it, but I don't. I've seen others hit power limits, however.

 

[blackangus]

I am in the same boat. Especially since it seems gaming performance won't match previous gen X3D, there's not a great inscentive to jump from a 7950X3D to a 9950X when the 9950X3D seems to be just around the corner.

My thought here is that they know how well the x3d stuff sold to gamers and are just getting it out the door early as possible to make the full range available and not have people just waiting.
The longer x3d is available the more they will sell!

 

[Jan Olšan]

If they were designing for N3 at first and had to backport (continuing the assumption made in the post I originally replied to, I'm not endorsing that assumption) they might have originally planned on 10 cores per die, and had to back off to 8.

What is unrealistic is redesigning a core to be smaller because of a change to the process roadmap. Such a major redesign would set your schedule back so far that it would obviate the reason for doing it.

That makes sense then, but kind of I don't see them ever planning either higher L3 cache or core count bump in the CCDs. Mostly intuitive guess based on how the company has worked in last X years so I can't offer some nuanced explanation.

Cutting some uarch features might not necessarily set you back in development if the project basically works with the possible variants until a deadline where the process node needs to be locked in, and the 3nm to 4nm swap happened before that deadline (or put it another way, they made sure to have the question of node solved by the time they had to decide on things like dies size and what to include). Meaning such changes would not be made late, but in time. The ditching of N3 was decided well in advance IIRC.

Not sure if I think they did this, myself. It's so remote from the sorts of things we have data and info on available... My default assumption is probably that Zen 5 didn't really lose "stuff" from getting ported to N4. Doesn't mean there weren't planned but dropped/postponned features, but that probably happens regardless, for all designs.

 

[soresu]

wheres mah zen6 6ghz 2nm 32core ultraX3D 1gb L3 cache mobile

9955HX3D in Q1 '25 hopeful?

TSMC just cant print enough money

It's not quite so rosy as all that.

They have to sink a lot of money into developing each new major node, and that's without even considering a device change like finFET to GAA/Nanosheet.

It takes a while to break even on that investment and then start profiting on it, even with major players like Apple buying up shed loads of capacity before the node fab has even broke ground.

 

[Joe NYC]

Impressive. I wish I were able to make eight paragraph long articles that say next to nothing out of a single tweet and get paid for it.

That's part of the interviewing process, to become a Tech Radar writer.

 

[Joe NYC]

Some don't use twitter and say it asks them to sign in to see the posts, so I did not direct link.

Any comments besides sarcasm and jelly?

My question, reading it: "How do we know that the V-Cache optimization (in the driver) is related to Ryzen 9000 specifically or just general V-Cache optimization"?

 

[fastandfurious6]

It's not quite so rosy as all that.

They have to sink a lot of money into developing each new major node, and that's without even considering a device change like finFET to GAA/Nanosheet.

It takes a while to break even on that investment and then start profiting on it, even with major players like Apple buying up shed loads of capacity before the node fab has even broke ground.

is there some major FOMO in the air that the next ARM supermiracle will outperform X86?

strong investment into the absolute BEST CPU should be a no-brainer no? it will buy the whole market

 

[soresu]

is there some major FOMO in the air that the next ARM supermiracle will outperform X86?

strong investment into the absolute BEST CPU should be a no-brainer no? it will buy the whole market

I was talking about semiconductor fab economics, not CPU core performance.

AFAIK the recently announced Cortex-X925 µArch should clock for clock beat any current x86 core in raw perf for scalar workloads, probably including Zen5.

As to perf/watt - less certain about that, we'll have to wait and see on that score for Zen5 and X925.

Notice I said scalar workloads.

None of the consumer ARM cores I know about come close to matching the vector compute capacity per core of either Zen5, or Intel's current *Cove core.

The best currently announced being the X925, which has 6x 128 bit SIMD units, while I'm pretty sure that Zen5 has at least 4x 256 bit units, possibly 2x 512 bit units.

Apple M4 I believe is still just 4x 128 bit units, which is decent enough for mobile and lower end desktop/DCC work, but not enough for hardcore DCC demands, let alone HPC these days.

Oryon is apparently also 4x 128 bit units.

 

[adroc_thurston]

while I'm pretty sure that Zen5 has at least 4x 256 bit units, possibly 2x 512 bit units.

4x512b, they're 2 FMAs and 2 FADDs.

 

[soresu]

4x512b, they're 2 FMAs and 2 FADDs.

Noice.

 

[CouncilorIrissa]

RPCS3 will see some obscene gains thanks to this.

 

[fastandfurious6]

I was talking about semiconductor fab economics, not CPU core performance.

AFAIK the recently announced Cortex-X925 µArch should clock for clock beat any current x86 core in raw perf for scalar workloads, probably including Zen5.

As to perf/watt - less certain about that, we'll have to wait and see on that score for Zen5 and X925.

Notice I said scalar workloads.

None of the consumer ARM cores I know about come close to matching the vector compute capacity per core of either Zen5, or Intel's current *Cove core.

The best currently announced being the X925, which has 6x 128 bit SIMD units, while I'm pretty sure that Zen5 has at least 4x 256 bit units, possibly 2x 512 bit units.

Apple M4 I believe is still just 4x 128 bit units, which is decent enough for mobile and lower end desktop/DCC work, but not enough for hardcore DCC demands, let alone HPC these days.

Oryon is apparently also 4x 128 bit units.

yeha that makes total sense

TBH for majority of consumers ARM is just better, then x86 vector processing for data centers/HPC/professionals and gaming

and TBH a local working small LLM is gonna be useful for everyone eventually, we're close to that

 

[itsmydamnation]

TBH for majority of consumers ARM is just better,

Hardly ..... overall single thread performance is still king, especially for things like GUI's , Web/JS and Games. so onces you hit a TDP for a chip of like 15watts your statement is objectively wrong.

 

[adroc_thurston]

small LLM

Oxymoron.
Either way ML stuff isn't useful.

 

[soresu]

and TBH a local working small LLM is gonna be useful for everyone eventually, we're close to that

The whole point of NPUs is to lift workloads like that off the CPU/GPU at the consumer level once you have the models trained and ready for much lower power/compute inference.

 

[fastandfurious6]

Oxymoron.
Either way ML stuff isn't useful.

small LLM is close, those 7B models getting better

cant deny the usefulness of an LLM, it's still early days (brute force method) but its the holy grail of computing

also AI for OS will ultimately be very useful but thye have to figure out how to implement that properly without being super creepy with privacy hazards

 

[soresu]

Either way ML stuff isn't useful.

Not much in phones excepting voice recognition algos and such.

VR HMDs could possibly benefit for foveated rendering (filling in the pixel gaps in low sample periphery) or inside out tracking though.

 

[adroc_thurston]

those 7B models getting better

Nope.

but its the holy grail of computing

I don't think piles of matmuls were ever that.

cant deny the usefulness of an LLM

But what does it do?

 

[jpiniero]

But what does it do?

Raise nVidia's stock price?

 

[poke01]

Hardly ..... overall single thread performance is still king, especially for things like GUI's , Web/JS and Games. so onces you hit a TDP for a chip of like 15watts your statement is objectively wrong.

I kinda get what he’s getting at. Phones are people’s computing devices theses days. Intel would has a much larger market to compete in if x86 was still relevant in phones.

 

[itsmydamnation]

I kinda get what he’s getting at. Phones are people’s computing devices theses days. Intel would has a much larger market to compete in if x86 was still relevant in phones.

My hot take

phones are poor peoples primary computing device, and a not poor persons convenience. Take my 15 and 17 year old children , They both have mobile , iPad and laptop , we have desktops and PS4 + PS5. They do all there real work on the laptop , do most of their creative work on the laptop ( they have drawing tablets ) but some on the ipad and all their games on the Desktop + PS5. Phone is primarily music + video player.

i dont see AMD or x86 caring to race to the bottom against budget arm SOC's that go into budget phones........