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

[DisEnchantment]

Speculate at will

 

[uzzi38]

He got actual slides for once, AFAIK these are all correct.

Strix is 12 Zen 5 cores in the same way that PHX2 is 6 Zen 4 cores FYI though.

EDIT: oh way, I missed the slide he wrote that. That's fine then.

 

[A///]

Why would making servers for internal use "count" less than making them to sell? Why should Apple buy a bunch of x86 CPUs to build servers (like they have been doing and AFAIK are still doing today) if they can build servers cheaper and/or better with Apple Silicon? Or realize some benefit of using the same ISA for their cloud (or at least part of it, for storage i.e. iCloud CPU ISA is obviously irrelevant) as they use for the Mac & iPhone?

Obviously Apple wouldn't bother if they don't feel they are getting some type of benefit from it. Maybe they don't feel they will and don't ever build any AS servers. All I'm saying is that IF they do build them on a mass scale rather than just a small pilot test, we will hear about it eventually - because they would have some reason for doing it that's going to be important enough to share with developers and/or investors at some point.

As a company almost entirely focused on consumer products it is obvious why they don't offer a server for sale, even if they felt they could make better servers than Intel and AMD. Servers aren't a consumer product, and never will be. The Xserve product never made much sense, which is probably why it was not very successful and was discontinued.

The cost to engineer the hardware only to be used internally would be great and if they don't carry it over to the mac pro what use is there to keep it among themselves? Apple's dc's run their services, mass storage is done on encrypted third parties. I can't see reason why Apple would want to use low power high performance mac silicon in their service dcs unless they were running complex operations which doesn't add up to them pushing more and more towards on device task rendering with their neural engine and ai coprocessor. idk how many service dc's they have and how much power they use, but the only way i can see them spending that kind of money and effort is to push their carbon footprint claims, which ding ding ding are fairly legitimate.

 

[A///]

who at amd and intel are aiding mlid now all of a sudden?

 

[yuri69]

Zen 5 at 5nm. No 4nm nor 3nm?

Anyway, CES 24 is gonna be hilarious with even the roadmap putting the Strix to appear like 10 months after the show.

 

[uzzi38]

AMD documentation doesn't always go into the nitty gritty details of their products ahead of launch. Zen 5C cores are sometimes referred to or included in a count alongside as Zen 5 for example (e.g. the aforementioned "12 Zen 5 core Strix" or Turin being up to 192 Zen 5 cores).

Same thing also applies to N5 vs N4. Don't focus too hard on those sorts of small details.

 

[DisEnchantment]

He got actual slides for once, AFAIK these are all correct.

Strix is 12 Zen 5 cores in the same way that PHX2 is 6 Zen 4 cores FYI though.

EDIT: oh way, I missed the slide he wrote that. That's fine then.

At least someone can vouch for said leak. But on the other hand I recollect @Kepler_L2 saying Strix is fully chiplet based.
But that slide looks doctored. I don't recollect AMD putting a release timeline against each item and its core count usually and the RDNA 2 font looks out of place.
Something like this usually is quite consistent.

Fire Range is not chiplet? I thought the DT Zen 5 in mobile version is also chiplet. Any idea what is changed with RDNA 3.5?

 

[uzzi38]

I didn't check what he wrote about chiplet vs monolithic tbh. Pretty sure Strix is monolithic (but there was at least one earlier design that was chiplet - it's been redefined multiple times) and Strix Halo is chiplet. I'd expect Fire Range to be chiplet as well but to be frank I'd never actually heard of any details of it before - I'm just expecting to be chiplet because of Granite Ridge.

As for what's changed with RDNA3.5, @Kepler_L2 can tell you some actual specifics because it's in drivers already to at least some detail. But essentially some RDNA4 improvements are supposed to be reigned in.

As for release timeline, don't focus too hard on those, it's a bit in shambles atm.

 

[yuri69]

The mobile/APU line up should be more or less:

• Zen 2-something - low end
• Phoenix rebranded - mid end
• Phoenix successor - high end
• repackaged desktop - high TDP solution
• reused console design aka APU featuring a large GPU - experimental premium line with uncertain viability?

The first ones are a continuation of the current products. The last one is just another "fast and expensive APU". AMD seems to have been juggling the idea around for ages - Kaveri&GDDR5 or Zeppelin&Greenland&HBM. It is always lurking but doesn't seem to make it to the market. Let's see if this one does this time.

 

[Mopetar]

who at amd and intel are aiding mlid now all of a sudden?

If I worked at either company I'd email him all kinds of ridiculous nonsense knowing he'd publish it. I'd try to get Pat and Lisa in on the game as well. Make for great fun during Friday after work drinks.

 

[Thibsie]

If I worked at either company I'd email him all kinds of ridiculous nonsense knowing he'd publish it. I'd try to get Pat and Lisa in on the game as well. Make for great fun during Friday after work drinks.

Maybe that's what they do 😉

 

[Exist50]

The mobile/APU line up should be more or less:

• Zen 2-something - low end
• Phoenix rebranded - mid end
• Phoenix successor - high end
• repackaged desktop - high TDP solution
• reused console design aka APU featuring a large GPU - experimental premium line with uncertain viability?

The first ones are a continuation of the current products. The last one is just another "fast and expensive APU". AMD seems to have been juggling the idea around for ages - Kaveri&GDDR5 or Zeppelin&Greenland&HBM. It is always lurking but doesn't seem to make it to the market. Let's see if this one does this time.

I think there's a better chance of seeing a "big APU" now with both Apple establishing a precedent and AMD having much more budget room than they used to. Plus, AMD's dGPU share in laptops is so low they don't exactly have anything to lose.

That said, I'm surprised they still don't seem to be making anything targeted for the premium low-power (fanless) market.

 

[A///]

If I worked at either company I'd email him all kinds of ridiculous nonsense knowing he'd publish it. I'd try to get Pat and Lisa in on the game as well. Make for great fun during Friday after work drinks.

this is a lot better than my idea and more safe.

 

[soresu]

I think there's a better chance of seeing a "big APU" now with both Apple establishing a precedent and AMD having much more budget room than they used to. Plus, AMD's dGPU share in laptops is so low they don't exactly have anything to lose.

Plus Intel is almost certain to expand into that potential segment eventually as their own GPU ambitions crank up a gear.

 

[soresu]

That said, I'm surprised they still don't seem to be making anything targeted for the premium low-power (fanless) market.

They sort of do already.

Mendocino is a 4C Zen2, 2 CU RDNA2 SoC with a TDP of 8-15w.

That replaced Stoney Ridge (APU with Excavator CPU and GCN3 GPU), which itself replaced one of the old Jaguar µArch based SoCs like Kabini.

Eventually we will probably see a Zen4/5 based SoC replace Mendocino, but likely in a coule of years given Mendocino was over 2 years after the first Zen2 APU Renoir.

 

[Exist50]

Mendocino is a 4C Zen2, 2 CU RDNA2 SoC with a TDP of 8-15w.

Mendocino is in about the right power envelope, but it has scaled back performance to match. I'm thinking something more like an Apple M-series competitor, or even for Intel's 9W U/M chips.

 

[Abwx]

Mendocino is in about the right power envelope, but it has scaled back performance to match. I'm thinking something more like an Apple M-series competitor, or even for Intel's 9W U/M chips.

7nm Zen 2 APUs series had comparable perf/watt as 7nm Zen 3 APUs, while Mendocino is fabbed with a 6nm process, and despite only 2 CUs the GPU use RDNA 2 uarch.

All in all it should be somewhat better than Intel s offerings using e cores uarch.

 

[A///]

Plus Intel is almost certain to expand into that potential segment eventually as their own GPU ambitions crank up a gear.

based on.....?????????????????? what the mad ramblings of that idiot mlid?

 

[soresu]

based on.....?????????????????? what the mad ramblings of that idiot mlid?

I don't think it's particularly likely either, just saying that if the x86 market do start following Apple down that rabbit hole then Intel would certainly be interested in it.

After all they have already been in the vicinity of higher end APUs with Kaby Lake G using AMD gfx.

It's just a matter of time before they try again with their own fully in house design now that they are getting serious about Xe, regardless of what AMD do.

Also I don't think it's mad ramblings, as stated previously in someone elses comments it's most likely that someone has fed him this to jebait the various vendors one way or another - perhaps even to assess greater market desire for such a SKU.

 

[Exist50]

All in all it should be somewhat better than Intel s offerings using e cores uarch.

It might compete well in performance against ADL-N, but that's the low end market. AMD needs something in the 9W envelope to go against Intel's M package, and of course Apple as well. I'm surprised they'd prioritize a big iGPU project higher, though obviously in an ideal world, we'd have both.

 

[Abwx]

It might compete well in performance against ADL-N, but that's the low end market. AMD needs something in the 9W envelope to go against Intel's M package, and of course Apple as well. I'm surprised they'd prioritize a big iGPU project higher, though obviously in an ideal world, we'd have both.

Given the power envelope there s not much one can do, putting 8 cores instead is not economical, higher IPC help reduce frequency and is hence an advantage but that mean more silicon wich get you again on the cost issue.

Previous 9W chips were using 2C/4T, so doubling the core count is already a big improvement, and it s already miraculous that they got to a 6nm node, wich make me think that it will be a long lived chip, dont expect more in a near future, guess that Zen 5 will be out for quite a time before they update the thing with just a 5nm Zen 3 core since Zen 4 is surely too big to be cost efficient in this price range.

 

[Exist50]

Given the power envelope there s not much one can do, putting 8 cores instead is not economical, higher IPC help reduce frequency and is hence an advantage but that mean more silicon wich get you again on the cost issue.

Previous 9W chips were using 2C/4T, so doubling the core count is already a big improvement, and it s already miraculous that they got to a 6nm node, wich make me think that it will be a long lived chip, dont expect more in a near future, guess that Zen 5 will be out for quite a time before they update the thing with just a 5nm Zen 3 core since Zen 4 is surely too big to be cost efficient in this price range.

Intel pushes 2+8 Alder Lake down into the 9W envelope, and the rumors about Lunar Lake suggest it's 4+4, presumably targeting the same. It should definitely be possible for AMD to have a PHX2 (2+4) package that can fit into fanless systems, so I'm baffled why they don't seem interested.

 

[BorisTheBlade82]

Intel pushes 2+8 Alder Lake down into the 9W envelope, and the rumors about Lunar Lake suggest it's 4+4, presumably targeting the same. It should definitely be possible for AMD to have a PHX2 (2+4) package that can fit into fanless systems, so I'm baffled why they don't seem interested.

I'd think that PHX2 might exactly be that kind of thing. If OEMs might adapt it, is another question...

 

[Exist50]

I'd think that PHX2 might exactly be that kind of thing. If OEMs might adapt it, is another question...

I hope so, but the last rumors I recall only mentioned it going down to 15W.

 

[Kepler_L2]

I hope so, but the last rumors I recall only mentioned it going down to 15W.

PHX2e/BDT is probably targeting that 9W segment.

 

[Exist50]

PHX2e/BDT is probably targeting that 9W segment.

Thanks, I missed that rumor. But an all small core chip would be kind of disappointing. That seems more like a budget successor to Mendocino. Want at least 2 big cores for the kind of web browsing, office, etc. workloads people use a fanless laptop for.