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

[DisEnchantment]

Speculate at will

 

[jamescox]

The cores can be hidden from the OS pretty easily. The thing is that you don’t want to do this. You WANT Windows/Linux to see a normal core. Ideally, you would provide an instruction for an OS to call that describes the topology of the system and let the OS figure out scheduling. Things like core grouping (for shared caches, thread grouping, etc), speed relative to other cores, latency, power limits, etc. would ideally all be provided. However, this is not the case currently. You can sniff out a lot of this stuff using other methods, but they can be error prone.

Similarly, applications should be able to request a specific type of thread (low speed/low power/etc). Unless Microsoft has changed things, you don’t have much control over where your thread ends up.

(note my knowledge on both x86 instructions and general low level or systems programming is out of date, so it is possible this situation could have improved. I don’t develop at that level any longer)

I thought there was some patent about transparently switching a thread between a high and low performance core based on what the thread actually needs. This can be done in a transparent manner where the high performance and the low power core look like the same core to the OS. This would remove the need for scheduling tweaks on the OS side. I don't know how else you have a hidden core unless it is something like a gpu accelerator. The OS doesn't schedule threads in a gpu, the hardware/driver handles that.

Hardware switching theads wouldn't allow all of the cores to be used at the same time though. It would seem better to just use the preferred cores, or whatever scheduling hints are available in the OS, to get the high performance threads on the correct cores. This allows all cores to be used at the same time. Modern cpus have performance counters in them for all manner of things, probably including cache usage, so the cpu should be able to tell whether a thread would benefit from running on a higher performance core, so it seems like there should be something to allow the cpu to give the OS scheduler hints. I don't know if something like that already exist?

We already have a heterogeneous situation with some cores being able to clock higher than others in all cpus and some having more cache (v-cache parts). The dense cores don't really change that since they are the exact same architecture, just with smaller cache an lower clocks. The v-cache parts are large cache and slightly lower clocks.

 

[adroc_thurston]

I thought there was some patent about transparently switching a thread between a high and low performance core based on what the thread actually needs. This can be done in a transparent manner where the high performance and the low power core look like the same core to the OS. This would remove the need for scheduling tweaks on the OS side. I don't know how else you have a hidden core unless it is something like a gpu accelerator. The OS doesn't schedule threads in a gpu, the hardware/driver handles that.

iirc MS now mandates all CPU cores to be OS-visible.

 

[gdansk]

super not allowed (a whole lot of cloud vendors are locked into it).

I mean it doesn't delay desktop parts if Zen 5c CCD isn't done at the same time as Zen 5 CCD.
Does their roadmap suggest they'll arrive simultaneously? I didn't get that impression:

 

[adroc_thurston]

I mean it doesn't delay desktop parts if Zen 5c CCD isn't done at the same time as Zen 5 CCD.

Desktop parts are of frankly little relevance.

I didn't get that impression.

They're very close.

 

[gdansk]

Desktop parts are of frankly little relevance.

Tsk, that's all I care about...

They're very close.

Yeah, but my point was unlike Intel they do not have to get both Zen 5 and Zen 5c done exactly simultaneously to ship their part like Intel needs to do.

 

[adroc_thurston]

Tsk, that's all I care about...

Well, it is what it is.

but my point was unlike Intel they do not have to get both Zen 5 and Zen 5c done exactly simultaneously to ship their part like Intel needs to do.

Eh, Strix isn't far off, and it's both Z5 and Z5c, so there's that.

 

[bakyt115]

https://twitter.com/x/status/1685162762442588160

new info about zen5 apu

 

[RTX2080]

https://twitter.com/x/status/1685162762442588160

new info about zen5 apu

Strange. Maybe I would be the last one who would still think 'C' suffix cores are for server/datacenter only. The sources are fighting each others until these products release in the future.

 

[moinmoin]

iirc MS now mandates all CPU cores to be OS-visible.

Ain't that sweet. Offering the worst scheduler in the industry but demand that hardware must not circumvent it.

 

[adroc_thurston]

Offering the worst scheduler in the industry

It's OK.
Works for AMD and Qualcomm both.

 

[Timorous]

Strange. Maybe I would be the last one who would still think 'C' suffix cores are for server/datacenter only. The sources are fighting each others until these products release in the future.

Also for lower power and area saving.

I would expect by the time you are loading up more than 4 cores the all core clock is probably low enough that the c cores are more efficient than the standard cores and when doing basic stuff like web browsing the c cores will be more power efficient.

 

[dacostafilipe]

My understanding is that for Zen 1-4 this has been a single core ARM Cortex A5. Any chance they move to RISC-V for Zen 5 or a different ARM core?

RISC-V less likely, new ARM cores yes possible, but there has to be a valid reason to change that old trusty A5

 

[Gideon]

What are the current rumors on the launch window for different Zen 5 products?

Has there been any speculation/leaks that's more precise than the confirmed "in 2024"?. Is Strix really arriving before Granite Ridge?

 

[SteinFG]

What are the current rumors on the launch window for different Zen 5 products?

Has there been any speculation/leaks that's more precise than the confirmed "in 2024"?. Is Strix really arriving before Granite Ridge?

Not confirmed but current estimates are Q1-Q2 for granite, Q2-Q3 for strix. I don't think strix is arriving earlier than granite.

 

[eek2121]

What are the current rumors on the launch window for different Zen 5 products?

Has there been any speculation/leaks that's more precise than the confirmed "in 2024"?. Is Strix really arriving before Granite Ridge?

No idea, but if AMD sticks with normal cadence, mobile parts will be announced in January and gradually release throughout the year, while desktop parts will release late Q3 or sometime in Q4 of next year.

 

[coercitiv]

No idea, but if AMD sticks with normal cadence, mobile parts will be announced in January and gradually release throughout the year, while desktop parts will release late Q3 or sometime in Q4 of next year.

If AMD sticks with normal cadence then desktop parts can launch even relatively early next year. AFAIK Zen 4 was intentionally delayed to market by 1-2 quarters and AMD's "usual" target between gens is lower than 18 months.

I'm not saying this as being optimistic about an early 2024 launch, I'm merely pointing out that historical trend does not point towards a late 2024 release.

 

[jamescox]

Strange. Maybe I would be the last one who would still think 'C' suffix cores are for server/datacenter only. The sources are fighting each others until these products release in the future.

I don't think we will see a mixed chiplet with both types of cores in the same die, as some have talked about, but they appear to be going to mix dense and regular cores in APUs, if the rumors can be trusted. I was wondering if the dense and regular chiplets are actually made on the same process. Could they all have slightly different processes, even if it is only tweaks for that specific part? I was assuming that the dense core chiplet would be on something more like a mobile optimized process, not just a different layout on the same process.

I don't know if they will do a dense chiplet in a consumer product. The market for more than 8 cores is very small. As enthusiast and professionals, our view of this is often a bit skewed. Most people are probably running 6 core parts since it is probably the most economical. Steam survey shows a large percentage on 6 core and smaller percentages on 4 and 8 core. Probably a lot of people still running old intel 4 core systems. 4, 6, and 8 may still be more than 90% of the market in general. I don't have time to go looking for specific figures. Even the 12 and 16 core is likely a tiny market share, so going to 24 with a mixed chiplet situation seems like it would have an exceedingly small market. I don't know how often 8p + 16c cores would outperform just 8p + 8p anyway.

I have wondered if they might make a bigger push with a Ryzen Pro type product. A lot of low end products and such don't actually need the full Genoa or even Siena package. A 32-core Ryzen Pro type part based on dense cores seems like it would have a lot of uses in the professional market. Perhaps some storage appliances and other low end servers. A workstation like that would make a lot of sense for a lot of developers since you could have 32 core / 64 thread. Such a system would be very fast for compiles if the memory can keep up.

 

[soresu]

The market for more than 8 cores is very small

This to me seems unfathomably odd.

Given that the first 8 core console CPUs caem into play 10 years ago roughly it seems extremely off that there isn't more of a baseline around that spec by now, for gaming if nothing else - even if the 8 Jaguar cores of that generation were so weak in IPC and clock frequency.

 

[RTX2080]

I don't think we will see a mixed chiplet with both types of cores in the same die, as some have talked about, but they appear to be going to mix dense and regular cores in APUs, if the rumors can be trusted. I was wondering if the dense and regular chiplets are actually made on the same process. Could they all have slightly different processes, even if it is only tweaks for that specific part? I was assuming that the dense core chiplet would be on something more like a mobile optimized process, not just a different layout on the same process.

Your reply makes me wanna speak more in anand.
I don't know the process differences between nonC and C cores. Some recent infos from Greymon(s) and his(their) friends(witnesses), nonC cores and C cores are embedded at different pipelines inside packaging facility, mixing nonC and C cores together would vastly increase the complexity and cost. Sounds like Amd's facilities are either not ready for a mixture of two kinds of core yet, or only design for game console because of high cost. (Maybe we need different CCDs to seperate two kind of cores?)
I still waiting for their re-confirmation in the near future before these product release. The present stage I see these "hybrid or not hybrid for Zen5" as 50:50 gamble. Amd's statement about hybrid is still vague.

What are the current rumors on the launch window for different Zen 5 products?

Has there been any speculation/leaks that's more precise than the confirmed "in 2024"?. Is Strix really arriving before Granite Ridge?

Zen5 would enter mass production in October, release could happen during 2024H1.

 

[SteinFG]

This to me seems unfathomably odd.

Given that the first 8 core console CPUs caem into play 10 years ago roughly it seems extremely off that there isn't more of a baseline around that spec by now, for gaming if nothing else - even if the 8 Jaguar cores of that generation were so weak in IPC and clock frequency.

It's not like games today require X set amount of cores, instead they require Y tasks to be running at the same time at certain speed to maintain set amount of fps. And beyond Y tasks, devs can't add more tasks to increase fps. A 6-core zen2 chip would surely be able to run all the tasks that the 8-core jaguar chip does, at higher speed. That's why 6 cores is the standard now.

Even compared to today's chips, 6-core 7600 (Zen 4 at 5 GHz, 32MB L3) would handily outperform 8-core console chip (Zen 2 at 3.8 GHz, 8MB L3). As long as IPC is increasing at ~15% per generation, this trend will continue.

(I named them tasks here, but they are more like OS threads, but thread is a different term in the OS and in the CPU, so I don't want to confuse someone)

 

[moinmoin]

Given that the first 8 core console CPUs caem into play 10 years ago roughly it seems extremely off that there isn't more of a baseline around that spec by now, for gaming if nothing else - even if the 8 Jaguar cores of that generation were so weak in IPC and clock frequency.

Two caveats: Jaguar didn't have SMT. And while consoles have 8 cores for ages now, depending on gen and OS version up to two of them were reserved for the OS and all the additional system level features consoles offer. So 6 cores may well be more of a realistic baseline, and that's rather well matched in the market nowadays I think.

 

[eek2121]

If AMD sticks with normal cadence then desktop parts can launch even relatively early next year. AFAIK Zen 4 was intentionally delayed to market by 1-2 quarters and AMD's "usual" target between gens is lower than 18 months.

I'm not saying this as being optimistic about an early 2024 launch, I'm merely pointing out that historical trend does not point towards a late 2024 release.

Zen 3 (desktop) and Zen 4 (desktop) launched in 2H. Zen 2 (desktop)launched in July.

AMD only launched Zen and Zen+ earlier in the year.

Cezanne (Mobile) launched in January. As did Rembrandt (Mobile). Renoir launched in March, but IIRC was announced in January.

We likely won’t see Zen 5 desktop prior to September, especially as AMD won’t be pressured by Intel’s next refresh.

 

[coercitiv]

Zen 3 (desktop) and Zen 4 (desktop) launched in 2H. Zen 2 (desktop)launched in July.

AMD only launched Zen and Zen+ earlier in the year.

This is not the pattern you're looking for.

Dr. Ian Cutress: So far AMD’s rate of new products is on track to produce a new core almost every year. The roadmaps quite proudly showcase Zen 3 as almost ready, Zen 4 in development, and Zen 5 further out. Is this cadence sustainable?

Mark Papermaster: We’re on a 12-18 month cadence, and we believe that is sustainable. It’s what the industry demands from us.

Source.

 

[MadRat]

Why fight AMD when you have a niche that actually is unassailable? Seems like Intel does exceptional on IPC. If the IPC for top games was the focus they would achieve an undisputed performance crown.

We are reaching a point where single thread performance is hamstrung by throwing too much into one CPU. To get great IPC you need high dollar systems loaded with a bunch or extra cores that do not benefit your gaming experience. For consumers they should focus on raw IPC in 2 or 4 cores and then throw multiple core operations on chiplets or co-processors.

Focus bleeding edge foundry technology into the high performance parts that leverage simpler, low core count goals. Cuts down your refresh cycle time to market and reduces risk. Move distributed work, where core count is king, to the periphery so general tasks don't pump heat into your performance core. Leverage older proven foundry technology on higher core counts that sit on the CPU's edge, or on a co-processor. Your laptop may get one co-processor while a workstation might support a dozen. Maybe they plug in like memory sticks, in parallel rows with interlocking heatspreaders, and all cool down directly with a common fan.

If a consumer wants more general computing power they can throw in an extra co-processor or three. Even better if that co-processor can be upgraded annually to drive sales down the road. Your gamers will focus on bleeding edge main CPUs. But your co-processor pulls in your science-based workstation consumers that want 100+ general cores. The workstation consumer can care less how Fortnite plays on the main CPU and more about performance across many cores. Maybe co-processors with 4x as many e-core is better for them that 1/4 as many p-cores. Mix and match co-processors at different speeds and from different generations, and let the scheduler in the OS juggle the workload.

 

[v.strix]

Zen 3 (desktop) and Zen 4 (desktop) launched in 2H. Zen 2 (desktop)launched in July.

AMD only launched Zen and Zen+ earlier in the year.

Cezanne (Mobile) launched in January. As did Rembrandt (Mobile). Renoir launched in March, but IIRC was announced in January.

We likely won’t see Zen 5 desktop prior to September, especially as AMD won’t be pressured by Intel’s next refresh.

See, Zen2 -> Zen3 was 16 months.
Zen3 -> Zen4 was 22. If you subtract aforementioned two quarters from it you'll arrive exactly at 16 months, coincidentally.
So, 16 months from Zen4 is next January. 22 is July.
Finally, if it's more than 22 it'll end up being the biggest delay in Zen era yet.

Make of it all whatever you will...