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

[DisEnchantment]

Speculate at will

 

[itsmydamnation]

mods, can someone please delete all the crap that does not have anything to do with Zen5? thanks.

It's funny this thread is degenerating the closer we get and it's still crazy no real leaks

 

[DrMrLordX]

At least ROCm is AMD-related. Also I think Strix Point and Strix Halo are (or will be) ROCm-supported at launch, possibly even in Win11.

 

[Mahboi]

At least ROCm is AMD-related. Also I think Strix Point and Strix Halo are (or will be) ROCm-supported at launch, possibly even in Win11.

Linux 6.10 Improves AMD ROCm Compute Support For "Small" Ryzen APUs - Phoronix

www.phoronix.com

"Small APUs(i.e., consumer, embedded products) usually have a small carveout device memory which can't satisfy most compute workloads memory allocation requirements.

We can't even run a Basic MNIST Example with a default 512MB carveout. https://github.com/pytorch/examples/tree/main/mnist.

Though we can change BIOS settings to enlarge carveout size, which is inflexible and may bring complaint. On the other hand, the memory resource can't be effectively used between host and device.

The solution is MI300A approach, i.e., let VRAM allocations go to GTT."

Definitely working towards it.

 

[Tuna-Fish]

The rumors are that the IO die is unchanged from Zen 4. If That's the case running above 1:1 is essentially useless, as bandwidth will be limited by fabric clock and latency will be worse.

Is it known if the IF frequency is limited by the CCD or the IOD on Zen4?

If it's limited by the CCD, then it would be possible for Zen5 to both use the exact same IOD, and also for it to potentially clock it higher at 1:1.

 

[Det0x]

The rumors are that the IO die is unchanged from Zen 4. If That's the case running above 1:1 is essentially useless, as bandwidth will be limited by fabric clock and latency will be worse.

Even single CCD Zen4 cpus benefit from higher bandwidth in 2:1 mode, while having lower latency when running in 1:1 mode
The same is true with dual CCD cpus, the difference is just larger

Whats "fastest" depends on what the app/benchmark/game prefer, bandwidth or latency (for example y-cuncher loves bandwidth)

I can provide real numbers from clam cache/memory benchmark showing this, for both single and dual CCD's in a 1:1 vs 2:1 mode comparison to back up my claims if requested. (6666/2222 vs 8080/2222 with 7800X3D and 6600/2200 vs 8000/2200 for 7950X3D)

Btw the same is true for "synced FCLK" vs "unsynced FCLK".. I also have numbers for the 8700G @ 2200/8800 vs 2600/8800
(for vanilla Zen4 "synced FCLK" would be 2000/6000, 2133/6400, 2200/6600 and 2000/8000 for example)

 

[eek2121]

Can't blame them. Getting proficient in a language is not easy. It's kind of masochistic for a normal person. And even if a person may know multiple languages, he may not be a master in any of them. Our own Dr. Ian Cuttress doesn't have a blogpost about ROCm, despite his Ph.D being related to CUDA so he's the sort of person who should be exploring ROCm more and comparing and contrasting it to CUDA.

You don’t understand it: we autistic nerds love learning new languages and frameworks. I have a few languages that see daily use, depending on the type of project, but I have learned over 20 in my lifetime.

Even single CCD Zen4 cpus benefit from higher bandwidth in 2:1 mode, while having lower latency when running in 1:1 mode
The same is true with dual CCD cpus, the difference is just larger

Whats "fastest" depends on what the app/benchmark/game prefer, bandwidth or latency (for example y-cuncher loves bandwidth)

I can provide real numbers from clam cache/memory benchmark showing this, for both single and dual CCD's in a 1:1 vs 2:1 mode comparison to back up my claims if wanted (6666/2222 vs 8080/2222 with 7800X3D and 6600/2200 vs 8000/2200 for 7950X3D)

Btw the same is true for "synced FCLK" vs "unsynced FCLK".. I also have numbers for the 8700G @ 2200/8800 vs 2600/8800
(for vanilla Zen4 "synced FCLK would be 2133/6400 and 2000/8000 for example)

Interesting. I haven’t tried to push IF clocks on my own chip. Right now it is set at 2133, and my 7950X suffered thanks to the voltage issues that killed chips early on, but when Zen 5 comes out and I upgrade, I will definitely explore pushing the limits.


(I chose not to RMA because the chip works fine at stock, it just doesn’t tolerate any type of overclock, which is fine because I need stability)


Hoping to snag the 16 core part again, and when Strix Halo finally drops I want to upgrade my laptop.

Anyone else have plans to upgrade?

 

[leoneazzurro]

Is it known if the IF frequency is limited by the CCD or the IOD on Zen4?

If it's limited by the CCD, then it would be possible for Zen5 to both use the exact same IOD, and also for it to potentially clock it higher at 1:1.

My guess would be that package may play here a role bigger than the silicon.

 

[Aapje]

It's always the same story, don't give devs more work. They already had too much before they woke up today.

Yes, but it's also a lot of work to have to rebuild something because the company can save a lot of money with different hardware or because they have to switch hardware for other reasons. We also see that the big companies are really getting into building their own chips.

It seems to me that the most common sense thing to do for Google, Amazon, Meta, IBM, MS, etc is to create an open source competitor to CUDA, or adopt and expand upon the open source ROCm. Then they can share the costs of building software for their own proprietary chips, but also pit hardware vendors against each other.

And it's beneficial to smaller/upcoming hardware vendors as well, since they can make drivers for the open source compute stack and then they are automatically compatible with an existing stack of software. If Tenstorrent/Jim Keller makes a great chip, it could be used with the same software stack. And AMD could use it to force open the compute market.

In general, developers tend to much prefer standards and open source software, over propietary vendors and software, so they don't dig themselves in a hole, where they may be forced into doing major rework or even a complete rebuild of the software.

 

[MadRat]

Its going to be very relevant when CUDA-dependent code gets replaced using Intel-based AI accelerators to port code to AMD ROCm.

 

[static shock]

EDIT: Nevermind.

 

[SteinFG]

For comparison, 190 is the brazilian 911.

Wrong thread

 

[DAPUNISHER]

mods, can someone please delete all the crap that does not have anything to do with Zen5? thanks.

Done. Thanks for trying to get people to stop squirreling. If I missed some hit me up with a link to the post numbers and I'll clean it up. I am still on my firt cup of tea.

Wrong thread

Indeed it is.

A quick heads up on both the rules and culture here on the Anandtech boards for those not in the know. We require that discussions stay on track. The vast majority of members here voted for it ages ago. The membership at the time, a fair number of whom are still here, were given a strong voice in how the moderation would work moving forward. That was in fact, how I became a mod, I was nominated by other members.

Please stay on topic. Attack the posts, not the posters (first rule) and do not shake pom poms in other vendors threads. And profanity is also not allowed. Hell and versions of damn are fine though. As mods we don't like restricting your access to the forums. Follow those rules/embrace that culture and you can do your thing. Happy posting.

Mod DAPUNISHER

 

[blackangus]

Another quick question - Is the general consensus still no new chipset?
I have seen a couple articles in the last few days talking about an 8xx series chip set for AMD, but it all still seems to be guess work.

 

[SteinFG]

Another quick question - Is the general consensus still no new chipset?
I have seen a couple articles in the last few days talking about an 8xx series chip set for AMD, but it all still seems to be guess work.

Well, new name at least.
B850, X870 are almost confirmed. There is some talk about B840.
But they'll be using old silicon for the chipset.
B850 = B650, X870 = X670. There's some chance AMD will require PCIe 5.0 now instead of leaving it for the EXTREME CHIPSETS

 

[Kepler_L2]

Another quick question - Is the general consensus still no new chipset?
I have seen a couple articles in the last few days talking about an 8xx series chip set for AMD, but it all still seems to be guess work.

The actual chipset (Promontory21) is unchanged. 800-series motherboards will just be a rebrand of 600-series, like 300-series to 400-series.

 

[blackangus]

Are new MB's going to be released in tamdem with ZEN5 availability?

 

[SteinFG]

Are new MB's going to be released in tamdem with ZEN5 availability?

Yes, maybe a couple weeks apart

 

[leoneazzurro]

In theory every AM5 motherboard should be already Zen5 compatible, they need only a BIOS update, possibly we will see mostly rebranded refreshes.

 

[soresu]

The actual chipset (Promontory21) is unchanged. 800-series motherboards will just be a rebrand of 600-series, like 300-series to 400-series.

Hopefully cheaper regardless.

Mobo costs have become ridiculous for anything beyond a really basic option.

 

[gdansk]

Hopefully cheaper regardless.

Mobo costs have become ridiculous for anything beyond a really basic option.

Introducing "new" product at lower prices? No, they must be more expensive.
But maybe 600 get sales.

 

[CakeMonster]

Jumping from 6** to 8** is a bad look when its not even iterative, just rebadged. Gives the impression that its a new generation.

 

[linkgoron]

Jumping from 6** to 8** is a bad look when its not even iterative, just rebadged. Gives the impression that its a new generation.

AMD just want to align all of their numbers, Ryzen 8xxx, Radeon 8xxx and now 8xx mobos.

 

[gdansk]

AMD just want to align all of their numbers, Ryzen 8xxx, Radeon 8xxx and now 8xx mobos.

But won't it be Ryzen 9000?

 

[linkgoron]

But won't it be Ryzen 9000?

I assumed that It'll be Ryzen 8xxx because of the "yearly" thing where the first number now represents "the year" that the part was released in - but I haven't been following the naming so I may have missed that Zen5 will be Ryzen 9xxx.

 

[Timorous]

Jumping from 6** to 8** is a bad look when its not even iterative, just rebadged. Gives the impression that its a new generation.

Not really.

The main thing is that if I buy a 6xx mobo now the board won't support Zen 5 and I need to update the bios. Shouldn't be an issue with usb flashing even without a CPU but still a pain for a new builder.

Having 800 series means OEMs get something new to sell, can iterate on designs to fix any niggles and ensures the board comes with a zen 5 compatible bios from the factory. So a bit of a win win.

It will also allow for some 600 boards to go on fire sale which is good for those who don't mind the faff of updating bios or who want a cheaper zen 4 build.