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

[DisEnchantment]

Speculate at will

 

[DavidC1]

Wait, doesn't Windows rotate single thread workloads among cores?
On Linux I don't think I ever encountered an issue like this, every 10 seconds or so a thread is passed to a different core, at least according to my system monitor.

Windows rotating threads among cores is a remnant of the single core era when they found doing that improved performance by 1-2%.

 

[Josh128]

Wait, doesn't Windows rotate single thread workloads among cores?
On Linux I don't think I ever encountered an issue like this, every 10 seconds or so a thread is passed to a different core, at least according to my system monitor.
Perhaps the Asus laptop is doing something like this but not all cores can reach 5.1GHz?

Edit: I'm on vacation off-country since Sunday and got a serious case of food poisoning after 2 days, but I must say that the last 15 pages of this thread managed really well to keep me distracted and entertained.
One thing that keeps being repeated though is that the high cross-CCX latency between the zen 5 and zen 5C cores is hurting performance. May I ask if there is any proof of that in the actual (non-synthetic) benchmarks ran so far? From the Phoronix review nothing like that had emerged afaik.

The core jumping is done by the CPU, at least on Ryzen. I think it started with Zen 2. Its an ingenius way to keep max boost clocks. It'll usually use its best 2 or 3 cores and constantly jump between them on something like a CB 1T run. Windows may rotate different cores it starts apps on, but I think when a process is in the middle of running, its the CPU algorithm that does it, based on current, temperature, and voltage.

 

[DaaQ]

There is a point when a reasonable person would stop, before trying to convince the very last person with the opposing view in endless tit for tat.

And that point does not have to equal inactivating one's account, just simply letting go.

I'm that way almost exclusively, I decide before hand how much energy I am willing to expend.
Got one vacation here, but I was still right.
Then there is OCN where I have exhausted my effort with a profile that claims to be VP of OCZ. It was in defending someone else but eh.

Sorry posted this 4 pages too early, not caught up yet.

 

[mostwanted002]

From MLID (to be taken with a grain of salt):

 

[mostwanted002]

Full video:

 

[igor_kavinski]

From MLID (to be taken with a grain of salt):

View attachment 104273

Oh man, that sucks if true. Why disable cores on a 9950X to downgrade it to 9900X? Release it as 9950LE for Lottery Edition and price it a bit lower than 9950X, like maybe $50 less. Then if it doesn't completely meet expectations, at least it can still be more useful than a lame 9900X!

 

[Saylick]

Oh man, that sucks if true. Why disable cores on a 9950X to downgrade it to 9900X? Release it as 9950LE for Lottery Edition and price it a bit lower than 9950X, like maybe $50 less. Then if it doesn't completely meet expectations, at least it can still be more useful than a lame 9900X!

If it was not hitting clock targets reliably, they could have just saved it for a 9950 non-X, but if the issue was the core itself simply wasn’t stable then you can only fuse those cores off.

 

[igor_kavinski]

Lunar Lake is a monster in text processing, as well as in some ST tests: https://browser.geekbench.com/v6/cpu/compare/7102066?baseline=7141329



That's just 8 cores vs. 24 threads!

 

[Gideon]

I hope someone at AMD sees this and redoes their abysmal naming again:

Using the term ‘artificial intelligence’ in product descriptions reduces purchase intentions

Companies may unintentionally hurt their sales by including the words “artificial intelligence” when describing their offerings that use the technology, according to a study led by WSU researchers.

news.wsu.edu

Apparently adding "AI" to your product description makes it *less* desirable to people.

They should just remove that part from the naming scheme along with "HX". Both add nothing of value anyway

 

[mikegg]

I hope someone at AMD sees this and redoes their abysmal naming again:

Using the term ‘artificial intelligence’ in product descriptions reduces purchase intentions

Companies may unintentionally hurt their sales by including the words “artificial intelligence” when describing their offerings that use the technology, according to a study led by WSU researchers.

news.wsu.edu

Apparently adding "AI" to your product description makes it *less* desirable to people.

They should just remove that part from the naming scheme along with "HX",l. Both add nothing of value anyway

Not surprising.

The boom is in Transformer-based architectures. IE. Generative AI.

Your TV with "AI" is not the same and is just a marketing trick.

 

[gdansk]

Perhaps the AI in the name wasn't for consumers? Maybe OEMs like the sound of 'AI 9'. Or investors.

 

[poke01]

Perhaps the AI in the name wasn't for consumers? Maybe OEMs like the sound of 'AI 9'. Or investors.

It was for OEMS and investors. Take a look at ASUS’s Strix Point landing page and AI is referenced a lot.

 

[Saylick]

I caught these posts on Xitter regarding Strix Halo. Can see where the CPU and GPU is located based on hotspots. Also, die sizes.

https://twitter.com/x/status/1818894711865983044

 

[poke01]

I caught these posts on Xitter regarding Strix Halo. Can see where the CPU and GPU is located based on hotspots.

https://twitter.com/x/status/1818894711865983044

View attachment 104283
View attachment 104284

You can also see the 128GB ram config. This basically a x86 M3 Max but with a 256-bit bus. Those 128GB SKUs will be sought after LLM users. We will have 2 vendors that provide 128GB unified memory by the end of 2024.

 

[SteinFG]

Woah, 450mm² of N3 and N4 silicon... this thing is yooge

 

[Saylick]

Woah, 450mm² of N3 and N4 silicon... this thing is yooge

Yeah, roughly 66mm2 per CCD and a little over 300mm2 for the N3E IOD, which contains the GPU.

 

[jdubs03]

Here’s a review for the Asus ProArt 13 from NBC:

https://www.notebookcheck.net/Asus-ProArt-PX13-review-The-world-s-fastest-13-3-inch-2-in-1-thanks-to-AMD-Zen-5-and-RTX-4070-laptop.868429.0.html

(They also have a video up on YouTube).

Battery life isn’t what I thought it’d be tbh. Best thing I see is a solid MT uplift; but that only goes so far. The load max was surprising too (185W). Average load was 74W.

 

[DisEnchantment]

Woah, 450mm² of N3 and N4 silicon... this thing is yooge

What is N3 and N4 used for (Should be N4 for Zen 5, I guess)

Would be interesting to know if they are making any fabric changes or just route SDF/SCF over the fan out links.

 

[tsamolotoff]

Would be interesting to know if they are making any fabric changes or just route SDF/SCF over the fan out links.

I hope they won't cheap out, but knowing AMD, it'd have two single IF links with bandwidth of a dual chiplet desktop CPU 💲 💲 💲 . I wonder what prevents them from plopping the IMC into the core ringbus within monolithic APUs or these '(almost) as good as monolithic' substrates.

 

[DisEnchantment]

I hope they won't cheap out, but knowing AMD, it'd have two single IF links with bandwidth of a dual chiplet desktop CPU 💲 💲 💲 . I wonder what prevents them from plopping the IMC into the core ringbus within monolithic APUs or these '(almost) as good as monolithic' substrates.

Z4 already has dual SDP per CCD (EPYC GMI wide), so at the very least with dense fanout interconnect they can enable both SDPs and still consume less than half the energy compared to DT CCD.
So it would have double the BW of the DT CCD if they do this at least.

But that sounds like bare minimal effort, some new innovation should be there.

The other interesting aspect is the MALL, which can do aggressive prefetching for hiding memory latencies (as done on MI300), the one on RDNA2/3 is not capable of this.
I again hope this is the one they use not the one from RDNA3.

 

[Gideon]

Perhaps the AI in the name wasn't for consumers? Maybe OEMs like the sound of 'AI 9'. Or investors.

Probaby true, but even Investors are seeing that most AI stuff fails to make any real money and starting sending mixed signals:

Mixed signals on tech stocks amid debate over viability of AI boom

Fears of fresh sell-off after Nvidia and Microsoft shares dip, but other chip stocks continue to rise

www.theguardian.com

https://www.washingtonpost.com/technology/2024/07/24/ai-bubble-big-tech-stocks-goldman-sachs/

As for the naming, I just remembered how logical the naming used to be, but i spawned it off into another thread:

Discussion - What would be the most Consumer Friendly Naming scheme for (AMD) CPUs?

Spawning off the "Ryzen AI" debate, what do you guys think would be the most consumer friendly naming scheme for CPUs? Rules and notes Setting aside the corporate or OEM needs for a moment. Let's focus on the end user. The goal is to invent the most concise, readable naming scheme that would...

forums.anandtech.com

 

[Abwx]

Here’s a review for the Asus ProArt 13 from NBC:

https://www.notebookcheck.net/Asus-ProArt-PX13-review-The-world-s-fastest-13-3-inch-2-in-1-thanks-to-AMD-Zen-5-and-RTX-4070-laptop.868429.0.html

(They also have a video up on YouTube).

Battery life isn’t what I thought it’d be tbh. Best thing I see is a solid MT uplift; but that only goes so far. The load max was surprising too (185W). Average load was 74W.

The dGPU use as much as 110W measured at the main, that s 75W peak left for the rest of the system, so about 68W at the DC adapter output, wich translate to 50-55W for the CPU.

 

[SteinFG]

What is N3 and N4 used for (Should be N4 for Zen 5, I guess)

Would be interesting to know if they are making any fabric changes or just route SDF/SCF over the fan out links.

N3 SOC die is the current rumor.
And N4 for CCDs

 

[Timorous]

Oh man, that sucks if true. Why disable cores on a 9950X to downgrade it to 9900X? Release it as 9950LE for Lottery Edition and price it a bit lower than 9950X, like maybe $50 less. Then if it doesn't completely meet expectations, at least it can still be more useful than a lame 9900X!

It costs money to print boxes with new branding and for what is probably a very low volume of units that cost probably out weighs the benefits of having something that sits between the 9900X and 9950X.

 

[Rheingold]

That's just 8 cores vs. 24 threads!

The "Text Processing" subtest scores ~3070 in ST and ~3800 in MT, which basically means it uses no more than 1.25 cores. The number of available threads is meaningless in this. Even a 96-core Threadripper only gets 3076 points in MT. On the other end of the spectrum is "Ray Tracer" with ~2600 in ST and then in MT ~15500 for Lunar Lake against ~31900 for Strix Point. The "Text processing" subtest basically shouldn't be considered for MT in the first place, or the benchmark should spawn multiple instances of it for MT.