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

[DisEnchantment]

Speculate at will

 

[jpiniero]

That's the only metric that really matters in client CPUs.

Not anymore. It's all about AI, and that's the only thing that matters

 

[Markfw]

Not anymore. It's all about AI, and that's the only thing that matters

I think that Intel and Microsoft are the only ones that think AI is the next coming.

 

[adroc_thurston]

I think that Intel and Microsoft are the only ones that think AI is the next coming.

Qualcomm and AMD and Nvidia and a whole lot of other vendors.
It's the next big thing (until the bubble bursts).

 

[H433x0n]

+20-25% multithreaded improvement would be disappointing imho. Many people including me would be happy not paying the zen5 tax and simply buy a discounted 7950X(3D), or an equivalent Intel.

So hopefully the MT improvement is very close to the ST one, or it's at least reachable with a bit of undervolting and overclocking. Not everyone spends hundreds of $ to brag about single core scores a bit higher than those of an ipad (or to game at uselessly high fps).

MT performance is almost completely useless and is on the path to becoming even more useless as more workloads are accelerated by GPU. I can’t think of a time where I’ve actually been bottlenecked by MT performance recently and I’m a developer. My bottlenecks have been memory bandwidth and PCIe lanes.

Stuff like cinememe is basically a useless benchmark for this reason. If you’re running laughably parallel workloads like this on a client PC you’re basically wasting electricity.

 

[adroc_thurston]

as more workloads are accelerated by GPU

Okay that's dumb.
No, SIMD machines aren't all that useful.

 

[H433x0n]

Okay that's dumb.
No, SIMD machines aren't all that useful.

Talking about client in particular.

Do the popular MT benchmarks like Vray, blender or video encoding performance really matter?

 

[adroc_thurston]

Talking about client in particular.

Client is javascript and video games.

Do the popular MT benchmarks like Vray, blender or video encoding performance really matter?

Sort of? Blender is a useful proxy for nT FP perf

 

[jpiniero]

Client is javascript and video games.

And Office 364.

 

[adroc_thurston]

And Office 364.

And teams 3*3 conferencing yea.

 

[carancho]

It doesn't matter.

That's the only metric that really matters in client CPUs.

A whole lot higher.

A whole lot higher than 4,000 GB 6.3 ST? 😲

 

[gdansk]

A whole lot higher than 4,000 GB 6.3 ST? 😲

With Ln2 sure

 

[H433x0n]

A whole lot higher than 4,000 GB 6.3 ST? 😲

Let’s say that’s a score of 4200, that’s ~40% increase from 7950X.

ARL-S is supposed to be ~3750 or so. That’d be a really good result if Zen 5 got a full +10% or more lead. Only 2 weeks left to get proper answers.

Since earlier today a bunch of people were making predictions on Zen 5 performance I’ll go out on a limb and predict it won’t outperform the M4 by a meaningful margin (if at all).

 

[Goop_reformed]

Let’s say that’s a score of 4200, that’s ~40% increase from 7950X.

ARL-S is supposed to be ~3750 or so. That’d be a really good result if Zen 5 got a full +10% or more lead. Only 2 weeks left to get proper answers.

Since earlier today a bunch of people were making predictions on Zen 5 performance I’ll go out on a limb and predict it won’t outperform the M4 by a meaningful margin (if at all).

4200 is not possible.

 

[H433x0n]

4200 is not possible.

Well, the hype has been +40% 1T perf and a score of 4200 is what it’d take to reach that.

 

[poke01]

i am thinking higher than 4000 as well

 

[Thunder 57]

Well, the hype has been +40% 1T perf and a score of 4200 is what it’d take to reach that.

You shouldn't listen to the hype train. The highest I remember anyone claiming was @adroc_thurston at 32% I believe. I'll just go ahead and say 22-25% (leaning towards the lower end, I can see 20-21% too) and see how off I am in a couple weeks.

 

[poke01]

if that 40% is true then AMD will have the fastest CPU core on the market and that will trickle down to their laptop SKU. This means that AMD can beat M4 even with the lower clocks than a supposed 9950X in a laptop Zen 5 SKU in every performance benchmark.

But we must manage our expectations like Thunder said. We can get excited after the June keynote.

 

[adroc_thurston]

You shouldn't listen to the hype train

the reddit atheism of tech discourse.

if that 40% is true then AMD will have the fastest CPU core on the market and that will trickle down to their laptop SKU. This means that AMD can beat M4 even with the lower clocks than a supposed 9950X in a laptop Zen 5 SKU in every performance benchmark.

yeeeeeeeeeeesssssssssss why do you think it won so many slots kek.
Perf is the selling point.

 

[Fjodor2001]

You shouldn't listen to the hype train. The highest I remember anyone claiming was @adroc_thurston at 32% I believe.

From previously in this thread:

Adroc's claim has always been 32% IPC iso/clk.

Looks like you missed it when he admitted that his 32% IPC Claim was made up BS because the real number is so high that people would think it's made up BS. And no, this is not a joke.

I remember it, but the claim not being a concrete number makes it impossible to evaluate whether the statement ends up true or not.

 

[Goop_reformed]

if that 40% is true then AMD will have the fastest CPU core on the market and that will trickle down to their laptop SKU. This means that AMD can beat M4 even with the lower clocks than a supposed 9950X in a laptop Zen 5 SKU in every performance benchmark.

But we must manage our expectations like Thunder said. We can get excited after the June keynote.

I believe the M4 is still faster in GB5 and 6

 

[itsmydamnation]

From previously in this thread:

Why do you keep doing this , spec int

Then it's quote it as IPC
Spec int doesnt do simd fp (funny that). so non avx512 simd workload will be less IPC then spec int and your avx512 enabled could be up to 100% more.

 

[Fjodor2001]

Why do you keep doing this , spec int

Then it's quote it as IPC
Spec int doesnt do simd fp (funny that). so non avx512 simd workload will be less IPC then spec int and your avx512 enabled could be up to 100% more.

I’m not claiming anything myself. I only quoted what others wrote previously. Don’t shoot the messenger.

According to those three quotes it was IPC. But like the quotes also say, the 32% number was made up BS anyway, so it doesn’t matter whether it was IPC or SpecINT. If you want to argue about that, bring it up with the persons I quoted or adroc.

 

[itsmydamnation]

I’m not claiming anything myself. I only quoted what others wrote previously. Don’t shoot the messenger.

According to those quotes it was IPC. But like the quotes also say, the 32% number was made up BS anyway, so it doesn’t matter whether it was IPC or SpecINT. If you want to argue about that, bring it up with the persons I quoted or adroc.

Thats not the way the real world works , otherwise i can just Quote Nazi's and say dont shoot the messenger , yep thats right i just went full Godwins Law.

So if you look at what adroc said he claimed his initial position was extrapolated from Server SIR and after that i haven't seen him post anything more specific. There is the opportunity for more numbers to become known over time etc , server , ridge , strix etc. Im not going to hyper play the man , its boring and small. But i am still yet to see someone explain a Low IPC position given all the bits and pieces we know about the core.

 

[adroc_thurston]

According to those three quotes it was IPC. But like the quotes also say, the 32% number was made up BS anyway, so it doesn’t matter whether it was IPC or SpecINT. If you want to argue about that, bring it up with the persons I quoted or adroc.

You would've known Zen5 perf targets for all segments if you were obsessed with that as much as you're obsessed with me. Please don't.

Either way, two more weeks!

 

[StefanR5R]

[Multithread performance] matters to me and every single person I know irl that uses a desktop or workstation. […] The vast majority of people that do any kind of serious work on a PC can't afford a Threadripper or an Epyc.

So how well does "any kind of serious work" done on desktops and workstations, in IRL, scale past ≈eight threads?

I am an engineer and my company lets me use a 6c/12t dual-channel PC. I, and those who pay me, and those who pay those who pay me, consider my work kind of… serious. At those occasions at which I use an application which is not single-threaded, this application alternates between 1T and 6T usage for almost 50% and 50% of its run time (the software developers determined to leave the HyperThreads unused in these nT parts), and uses 100% of the hardware threads for just brief stretches of the run time.

If your application scales easily to high thread counts but workstation and server are too expensive to you, then buy a dozen of 2nd hand PCs and a cheap Ethernet switch and run the CPU intensive part of your well scaling application on the Ethernet cluster. This instantly gives this type of applications greatly more nT performance than any generational CPU update will ever do, and for little money to boot. (Or if you do streaming vector arithmetic, get your application ported to GPU, as @H433x0n pointed out.)

But if your computing intensive desktop application is a more common one and scales to about 8 threads and not much more, then Zen 5 desktop CPUs will have you covered: On the top end SKU, the amount of cores which is actually useful to your application apparently will have plentiful 230 W PPT limit as power budget — that's the same as Zen 4's, but these cores are distinctly wider than Zen 4's and thus able to make more effective use of this high power budget —, and presumably can be driven to similar peak clock speeds as Zen 4. In other words, the generational 1T performance update pretty much directly translates into an nT performance update in this application scenario.

In other other words, even though you, as end user with nT performance needs but without budget do not matter to AMD and to OEMs, you still get to reap much of the benefits of those CPU architecture updates which are driven by requirements of customer groups which do matter to AMD and OEMs. Nice!