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

[DisEnchantment]

Speculate at will

 

[Io Magnesso]

Will there be a TURIN with V-cache?
Do you prioritize manufacturing resources on the packaging of cDNA GPUs?

 

[MS_AT]

View attachment 126238

https://twitter.com/x/status/1937731500977606951

I wonder how much of this is due to real silicon differences and how much due to BIOS, I think when he tested HX370 he had a lot of issues keeping the Asus he had from throttling, he eventually succeeded but I don't remember if that was thanks to replacing the machine with other Asus model. Anyway what I wonder about is did he had a chance to test other laptop SKUs with HX370, or does he base his investigation on the early samples he had access to.

 

[CouncilorIrissa]

I wonder how much of this is due to real silicon differences and how much due to BIOS, I think when he tested HX370 he had a lot of issues keeping the Asus he had from throttling, he eventually succeeded but I don't remember if that was thanks to replacing the machine with other Asus model. Anyway what I wonder about is did he had a chance to test other laptop SKUs with HX370, or does he base his investigation on the early samples he had access to.

I asked him about it:

https://twitter.com/x/status/1937846575277179240

He had two samples, his STX v/f curve is based on the retail ZenBook S16.

 

[LightningZ71]

I suspect that there was some tailoring of the lithography process between STX and KRK. As has been said multiple times before, both AMD and TSMC have other customizations that can be implemented in a given chip within a lithography node. While both of them may be manufactured on N4P, KRK is very likely going to be optimized for use in laptops that have less cooling and lower power limits as compared to STX. It will be interesting to see if the STX refresh will have any further process optimizations using lessons learned from KRK.

 

[DrMrLordX]

Will there be a TURIN with V-cache?

No.

 

[Io Magnesso]

No.

It's sad Well, it can't be helped because the packaging is given priority to cDNA GPUs...

 

[poke01]

AMD needs to wider. Intel and Apple are already 6 and 8 wide respectively.

Bad bot, that’s out of topic

 

[ToTTenTranz]

There it is. Strix Halo handheld spotted.

https://twitter.com/x/status/1938979516921385106

3x Strix Point performance on what seems to be a GPD Win prototype (35W?).

 

[fastandfurious6]

thats crazy, RTX 4060 performance on handheld

this thing can run everything 1080p ultra 60fps

 

[ToTTenTranz]

thats crazy, RTX 4060 performance on handheld

* without the 8GB VRAM limit. Minimum system RAM on a Strix Halo will probably be 32GB.


Just shut up and take it 💸💸💸💸💸💸💸💸

 

[LightningZ71]

Well, according to a prolific poster here, there's no way that thing can cost anything less than $2000+. They should sell dozens!

 

[poke01]

Well, according to a prolific poster here, there's no way that thing can cost anything less than $2000+. They should sell dozens!

It’s probably $1800+ and the starting config is 16GB

 

[DrMrLordX]

Well, according to a prolific poster here, there's no way that thing can cost anything less than $2000+. They should sell dozens!

You're lucky to find a Strix Halo mini PC for $1500 (granted they usually start at 64GB). Strix Halo ain't cheap.

 

[AMDK11]

AMD needs to wider. Intel and Apple are already 6 and 8 wide respectively.

Core Zen4: Decode 4-Wide, 4ALU, 4FP
Core Zen5: Decode 2x4-Wide(8-Wide SMT), 6ALU, 4FP
Core GoldenCove: Decode 6-Wide, 3ALU/FP+2ALU
Core LionCove: Decode 8-Wide, 6ALU, 4FP
Apple Core: Decode 10-Wide, 8ALU, 4FP

EDIT:
PantherCove: ?Decode 10-Wide, 8ALU, 6FP?

Intel emphasizes in the LionCove animation that splitting the unified scheduler (GoldenCove and earlier) into separate INTs and separate FPs allows scaling to at least 10ALU and 8FP. It is very likely that this was spread over at least two generations. The next big one after LionCove is PantherCove.

 

[LightningZ71]

You're lucky to find a Strix Halo mini PC for $1500 (granted they usually start at 64GB). Strix Halo ain't cheap.

I never said it was cheap. I am certain that they are pricing a very healthy margin into products with it though. They price it like it's a 9950x combined with a 4060ti marked up by 20%, which is similar to how it performs, but way more than the COGS. I also get that there aren't a lot of them in the market yet, so they are still punishing the wallets of early adopters. It's a great way to demolish word of mouth marketing and enthusiasm for the product.

Also, for a gaming part, there's no real need for both CCDs to be present. A custom SKU with a full house iGPU, but with only 8 cores would be perfect for a handheld on the higher end of the power range. It's not like they haven't done similar things before for HP and Lenovo or retuned the bios of an existing part to better fit low power usage ala the Z series.

 

[DrMrLordX]

Also, for a gaming part, there's no real need for both CCDs to be present.

It's not a gaming part. It's being sold as an AI self-hosting solution.

 

[Io Magnesso]

It's not a gaming part. It's being sold as an AI self-hosting solution.

No, it would be nice to have CCD on both sides.
It's not that having a large number of cores is for AI.

 

[moinmoin]

You're lucky to find a Strix Halo mini PC for $1500 (granted they usually start at 64GB). Strix Halo ain't cheap.

You are lucky, Framework not only has it as one of the first sellers, they even sell it starting from $1099 for the computer or $799 for the contained mini ITX mainboard.

I really don't understand these complaints about pricing and availability while you got these offers by Framework.

 

[ToTTenTranz]

You're lucky to find a Strix Halo mini PC for $1500 (granted they usually start at 64GB). Strix Halo ain't cheap.

Most likely, Strix Halo wasn't cheap until AMD failed miserably at getting design wins in the laptop space. If they sell a version with single 8c CCD + full 40CU IOD for like $400 they're probably already getting a markup well over 50% for it.


It wouldn't get them the M4 Max levels of profits they were hoping for, but it's still better than the alternative of letting the chip be forgotten into complete irrelevance IMO.

 

[adroc_thurston]

until AMD failed miserably at getting design wins in the laptop space

Says you?

 

[branch_suggestion]

Most likely, Strix Halo wasn't cheap until AMD failed miserably at getting design wins in the laptop space. If they sell a version with single 8c CCD + full 40CU IOD for like $400 they're probably already getting a markup well over 50% for it.

Strix Halo could've been sold for free with 10x the supply and adoption wouldn't be 10x of reality, not even remotely close.
It is far from that simple, it requires newly engineered platforms to support it, that takes time.
OEMs will sell for what they think they can get away with, and they are able to use the very overpriced green sticker machines to justify the high pricing for Halo machines.
NV has higher average GMs in mobile than probably any Strix Halo SKU.
To get cheaper Strix Halo machines, AMD would need MDF and force OEMs to sell them at very aggressive pricing for share to set up Medusa Halo...

It wouldn't get them the M4 Max levels of profits they were hoping for, but it's still better than the alternative of letting the chip be forgotten into complete irrelevance IMO.

Or just wait for others in the space to also make 256b SoCs so there is a healthy market to actually build in, which is precisely what N1X is.
Strix Halo is the Lakefield equivalent, Medusa Halo will be an actual product in an actual market since NV who owns nearly the whole laptop dGPU market is entering said market.
There is no credible market without NV or Intel being in it too, AMD just isn't going to bet big on client anymore all on their own tangent, it is not really what the company is good at. They excel at simply out engineering the competition in established markets. A simple pure mantra, build the best and they will come.
Medusa Halo obliterates N1X into nothingness, that is a pretty simple way to establish the pecking order, after all AMD are actually the early movers this time.

 

[poke01]

There is no credible market without NV or Intel being in it too, AMD just isn't going to bet big on client anymore all on their own tangent, it is not really what the company is good at. They excel at simply out engineering the competition in established markets. A simple pure mantra, build the best and they will come.
Medusa Halo obliterates N1X into nothingness, that is a pretty simple way to establish the pecking order, after all AMD are actually the early movers this time.

N1X will release late 2025/early 2026 and Strix Halo will do fine against it.

Medusa Halo’s competition is M6 Pro/Max at least in the AI space in early 2027. Will Nvidia even have N2X ready by then? AMD will own the 256b Windows space for a long time.

 

[branch_suggestion]

N1X will release late 2025/early 2026 and Strix Halo will do fine against it.

Yes, but we will need to see how many laptop design wins they get.

Medusa Halo’s competition is M6 Pro/Max at least in the AI space in early 2027.

Apple does not compete against Windows PC's, they just don't.
And also not merchant Si, which is kinda the big reason why.

Will Nvidia even have N2X ready by then?

Would need a very quick turnaround knowing how terribad N1X development has been.

AMD will own the 256b Windows space for a long time.

In every metric but design wins and market share, yes.
NV will be ahead early on through sheer force of entrenched share and lotsa monies, even knowing how ehh WoA is.
Glymur is a weird middle ground part, not sure what market it really addresses as the GPU will remain awful vs the comp.

 

[jpiniero]

If they sell a version with single 8c CCD + full 40CU IOD for like $400 they're probably already getting a markup well over 50% for it.

The IO die is really big though, and N4 is very expensive. There's a reason AMD is trying to sell SH on AI instead of gaming.

 

[poke01]

Yes, but we will need to see how many laptop design wins they get.

Dell, HP and Lenovo are ready for N1X.

Glymur is a weird middle ground part, not sure what market it really addresses as the GPU will remain awful vs the comp

No one sane will be doing AI work on a Qualcomm laptop.

Apple does not compete against Windows PC's, they just don't.
And also not merchant Si, which is kinda the big reason why.

Agree but they do compete in the laptop space. And Apple’s laptops do sell well in the $1800+ range.

In every metric but design wins and market share, yes.
NV will be ahead early on through sheer force of entrenched share and lotsa monies, even knowing how ehh WoA is.

It’s not easy, laptops are very hard. Don’t expect NV to flood the market. These aren’t dGPUs.