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

[DisEnchantment]

Speculate at will

 

[Josh128]

Zen 5, being a different design, would obviously make some difference.

It would, for sure, but 4nm is inherent to Zen 5 as we know it already. If its got +16%IPC over Zen 4 at iso clocks, then the only thing that would boost it to +30% over Zen 4 that would be attributable to process node would have to be clock speed. I really dont know, maybe all core clocks on 64 core Zen 5 are significantly higher than 64 core Zen 4. Thats why I posed the question.

On desktop, maybe it all cores +100 or +200 MHz faster, so if its the same percentage for HEDT, thats not enough to make up for the difference.

 

[StefanR5R]

Apropos...

Desktop and server do not share the CCD, just the design itself.

very very different xtor-level optimizations.

...AMD's f_max specifications:
5.0 GHz for Turin (9575F, 9175F),
5.4 GHz for Shimada Peak (any one of them),
5.7 GHz for Granite Ridge (9950X, and 9950X3D vanilla CCD).

I wonder which of the CCD flavors they put into Shimada Peak.

Edit: Also, are the CCDs fed by GMI-wide? It'd work up to 64 cores.

 

[StefanR5R]

TSMC claimed 22% power reduction or 11% more performance from N5 to N4P.

Keep in mind that TSMC provides such figures for certain ARM cores ARM IP on the vanilla process. AMD has got AMD cores, and DTCO is at play too. IOW, their own gains from process upgrade necessarily differs somewhat.

 

[adroc_thurston]

As the third major enhancement of TSMC’s 5nm family, N4P will deliver an 11% performance boost over the original N5 technology and a 6% boost over N4
So TSMC is lying?

TSMC Expands Advanced Technology Leadership with N4P Process|Taiwan Semiconductor Manufacturing Company Limited

Hsinchu, Taiwan, R.O.C., Oct. 26, 2021 - TSMC (TWSE: 2330, NYSE: TSM) today introduced its N4P process, a performance-focused enhancement of the 5-nanometer technology platform.

pr.tsmc.com

numbers in a vacuum that have no relation to modern DTCO'd to hell designs

 

[Mopetar]

Some of it could be tweaks to the process over the past year. I don't think that would explain the entire difference, but it has been a little over 12 months since Zen 5 launched and nodes do mature over time.

 

[adroc_thurston]

and nodes do mature over time.

It's TSM, their nodes are fully mature on ramp.

 

[igor_kavinski]

I can definitely glue a monitor to a case, plug it into power station and call it a day.

Like this?



Just needs wheels and a nice seat and it could double as a transport vehicle too!

 

[mostwanted002]

Like this?

View attachment 127962

Just needs wheels and a nice seat and it could double as a transport vehicle too!

Igor here casually leaking Tesla prototypes.

 

[fastandfurious6]

a week ago I thought about the future is small devices (just APU and ports, maybe handheld form with small screen and buttons) and then monitor/peripherals being in very portable form i.e. foldable OLED

yesterday I saw this lenovo "thinkbook" with rollable OLED transforms from 14'' to 17'' screen lmao https://www.notebookcheck.net/Think...-becomes-available-to-purchase.1073152.0.html

killer feature, will likely become norm if easy/cheap to implement, here in action

lenovo really at forefront in mobile, no other company comes close

 

[mostwanted002]

They have a "better" I/O die on TR chips. because even 2 channels can clear 100 GB/s at 6400.

As Wendell tested only 2 sticks here.

 

[igor_kavinski]

Wendell did a pretty great job, especially with gaming benchmarks on HEDT CPUs.

 

[511]

Wendell did a pretty great job, especially with gaming benchmarks on HEDT CPUs.

Wendell always does the best job in this regard

 

[511]

I agree with Adroc here. Do we have any evidence Zen 5 is clocking 11% faster at the same power draws as Zen 4? Because thats what this would imply.

Not necessarily but again this is what the half node provides even that is for ARM is it impossible for AMD to get few% from the node?

 

[gdansk]

Not necessarily but again this is what the half node provides even that is for ARM is it impossible for AMD to get few% from the node?

The process isn't explaining anything. Because Zen 5 has appeared elsewhere and has not faired this well compared to its Zen 4 predecessor when at the same power level.
I suspect membw is contributing heavily here as Phoronix includes a good number of AVX-512 workloads many of which enjoy membw.

 

[adroc_thurston]

I suspect membw is contributing heavily here

It's a dinky 4ch platform with no real men speed bumps.

 

[mostwanted002]

Wendell did a pretty great job, especially with gaming benchmarks on HEDT CPUs.

I've just started the Linux video from Level1Linux for the same CPUs, as he emphasizes that AMD has brought in a lot of upgrades and changes regarding TRs on Linux.

 

[gdansk]

It's a dinky 4ch platform with no real men speed bumps.

I already posted the test system memory speeds from Phoronix. It was 1.33x membw increase for the results people are talking about. 🤔 Not big, but enough to explain some of the "extra" +14%.

 

[igor_kavinski]

It was 1.33x membw increase for the results people are talking about.

Nimo Direct Inc. MME2S vs Micro-Star International Co., Ltd. MS-7E51 - Geekbench

browser.geekbench.com

That Strix Halo vs. 9950X3D comparison also shows some weirdly great multicore wins for the Halo chip and I think membw is the only explanation.

 

[gdansk]

Nimo Direct Inc. MME2S vs Micro-Star International Co., Ltd. MS-7E51 - Geekbench

browser.geekbench.com

That Strix Halo vs. 9950X3D comparison also shows some weirdly great multicore wins for the Halo chip and I think membw is the only explanation.

Possibly, who knows.
Here's a comparison to my stock 9950X3D at least running the same OS since the Linux builds of GB consistently put up better numbers.

Nimo Direct Inc. MME2S vs ASRock X670E Steel Legend - Geekbench

"HTML5 Browser" subtest being bandwidth limited I would not have expected

 

[igor_kavinski]

"HTML5 Browser" subtest being bandwidth limited I would not have expected

So if Zen 6 "fixes" this membw issue of Zen 5 due to better fabric and then adds +15% IPC and +10% frequency on top, it's gonna look really impressive compared to Zen 5, at least in GB6.

 

[poke01]

Possibly, who knows.
Here's a comparison to my stock 9950X3D at least running the same OS since the Linux builds of GB consistently put up better numbers.

Nimo Direct Inc. MME2S vs ASRock X670E Steel Legend - Geekbench

"HTML5 Browser" subtest being bandwidth limited I would not have expected

A SoC with 14 core 512GB/s has no effect on the score vs a 14 core 276GB/s SoC.

I guess past a certain point more bandwidth in this subtest is useless

 

[ToTTenTranz]

a week ago I thought about the future is small devices (just APU and ports, maybe handheld form with small screen and buttons) and then monitor/peripherals being in very portable form i.e. foldable OLED

yesterday I saw this lenovo "thinkbook" with rollable OLED transforms from 14'' to 17'' screen lmao https://www.notebookcheck.net/Think...-becomes-available-to-purchase.1073152.0.html

killer feature, will likely become norm if easy/cheap to implement, here in action

lenovo really at forefront in mobile, no other company comes close

Future ia EMG bracelets that detect fine finger and wrist movements, being used to interface with AR glasses.

Nimo Direct Inc. MME2S vs Micro-Star International Co., Ltd. MS-7E51 - Geekbench

browser.geekbench.com

That Strix Halo vs. 9950X3D comparison also shows some weirdly great multicore wins for the Halo chip and I think membw is the only explanation.

STX Halo has a faster and more expensive inter-CCD connect.

 

[yuri69]

His geomean results put 9980X as +30% over 7980X. Very interesting as it should have similar clocks as the 7980X. This holds true for 9970X vs 7970X as well (+28%). Whats accounting for the additional +14% overall perf over the claimed +16% IPC?

* There is a lot of throughput oriented benchmarks in that suite - 4800 vs 6400 memory
* There might be more AI/HPC benchmarks than "in the average mix" - full AVX-512 shines there
* More advanced DVFS algos on Zen 5?
* More mature platform/firemware/BIOS?

 

[DrMrLordX]

Yes really. Strix Halo is an experiment for AMD, they're testing the waters. The same applies to handheld makers, they're tempted by the capabilities of the product and willing to experiment despite it's obvious drawbacks.

The only real drawback is that it's way outside of the power range for handhelds. But hey whatever, maybe Asia has a market for these things. More sales for AMD isn't a bad thing.

Moreover, Strix Halo is such a known quantity that up until a few days ago you didn't even know it had a single CCD SKU.

Like my knowledge about a single-CCD 380/385 matters. I could be dead in a ditch tomorrow and it wouldn't make this any more or less interesting. Has anyone seriously not benchmarked the 380 or 385? We already know what they can do, don't we?

A real experiment would be someone going to AMD's semi-custom division and hacking together a part that would actually target handhelds. Something that would hopefully wind up being better than a Z2 Extreme.

 

[MS_AT]

His geomean results put 9980X as +30% over 7980X. Very interesting as it should have similar clocks as the 7980X. This holds true for 9970X vs 7970X as well (+28%). Whats accounting for the additional +14% overall perf over the claimed +16% IPC?

Sustained clocks should be higher. Just compare on Epyc https://www.amd.com/en/products/pro...ation-9004-and-8004-series/amd-epyc-9554.html vs https://www.amd.com/en/products/processors/server/epyc/9005-series/amd-epyc-9555.html (all core boost speed, this is not specified for threadripper parts) this is also what they were talking about around Zen5 release. That Zen5 parts don't boost higher but are able to hold higher clocks under heavier loads compared to Zen4. I mean it's few hundred MHz but if you add IPC increase and memBW increase it will all add up.