Discussion AMD SoC Halo series GPU discussion

[GTracing]

99watts is maximum. Due to some kind of regulation

Asus & MSI already have 80watts

gdansk is talking about power draw (100W), not battery capacity (80WHr).

 

[gdansk]

99watts is maximum. Due to some kind of regulation

Asus & MSI already have 80watts

I mean TDP, not battery size in watt hours.

 

[jpiniero]

99watts is maximum. Due to some kind of regulation

Asus & MSI already have 80watts

It's an airline requirement. The current Macbook Pros have a battery option of 99.6 wh. The limit is 100 I guess.

 

[insertcarehere]

This is what gets me, people who don't want certain products or products lines to be available for others because they themselves wouldn't buy it.

I wouldn't buy a 18" 4Kg notebook with a RTX 4080 Mobile and desktop CPU. But they exist and I'm perfectly fine that other people are buying those.

-Its just zero sum thinking. There is a market for a $400 handheld and there is a market for a $2000 handheld.

Nobody wants one market to come at the expense of the other.

As much as we wish it is, the high-end consumer gamer-that-also-values-thin-and-light-to-not-get-dGPUs is not exactly a big market, especially given how expensive Strix Halo is going to be for AMD to make. The existence of 'good enough' solutions at the handheld market with the likes of Switch 2 and Steam Deck (+refresh) on the lower end will absolutely be a headwind to making anything higher-end with more niche chips, which Strix Halo will be.

 

[GTracing]

Does anyone have a link to the slide-deck? I can't find it anywhere.

 

[Joe NYC]

Does anyone have a link to the slide-deck? I can't find it anywhere.

Some limited marketing slides are here:

AMD Ryzen AI Max "Strix Halo" APUs Are The Ultimate AI PC APUs: 16 "Zen 5" & 40 "RDNA 3.5" Cores, 256 GB/s Bandwidth & Up To 120W

AMD Ryzen AI Max "Strix Halo" APUs are finally here & offer strong specs with the Ryzen AI 9 Max+ 395 offering up to 16 CPU & 40 GPU cores.

wccftech.com

 

[techjunkie123]

Strix halo webpage says IOD/GPU is 4 nm , not on 3 nm.... Welp.

https://www.amd.com/en/products/processors/laptop/ryzen/300-series/amd-ryzen-ai-max-plus-395.html

 

[GTracing]

Strix halo webpage says IOD/GPU is 4 nm , not on 3 nm.... Welp.

https://www.amd.com/en/products/processors/laptop/ryzen/300-series/amd-ryzen-ai-max-plus-395.html

Wow. It could be a typo. 4nm would actually make more sense from a product design perspective. All their other APUs are 4nm.

Strix Halo would still cost about as much to produce as the AD103 in the RTX 4080 though.

 

[Meteor Late]

Strix halo webpage says IOD/GPU is 4 nm , not on 3 nm.... Welp.

https://www.amd.com/en/products/processors/laptop/ryzen/300-series/amd-ryzen-ai-max-plus-395.html

Do you think AMD wouldn't brag about using 3nm? they know many people care about using the latest node, many people associate it with better performance and efficiency. It is a very good marketing tool.

 

[Joe NYC]

Strix halo webpage says IOD/GPU is 4 nm , not on 3 nm.... Welp.

https://www.amd.com/en/products/processors/laptop/ryzen/300-series/amd-ryzen-ai-max-plus-395.html

That would, in fact, be great news!

It would mean that AMD can make these ~30-40% cheaper, and there would still be a lot of room for growth from moving to N3e.

 

[Joe NYC]

Interesting preliminary specs on HP MiniPC with Strix Halo.

https://www.hp.com/us-en/workstations/z2-mini-a.html

 

[Meteor Late]

Yep AI AI AI.
This product was never gaming focused now it's clear.
While many of you already knew this, I mention it because it's hilarious that almost anyone out there in Reddit, Youtube and other channels commented about how amazing this APU would be for gaming and all that.
This product was always targeted for AI with the huge RAM capacity compared to any GPU. And that's absolutely fine, but that's not what MLID and most people were raving about.

 

[adroc_thurston]

This product was never gaming focused now it's clear

It is very much gaming-focused, but that's also the inherent nature of anything with a big nuff GPU.
It's just also their Pro offering with ISV certs and other tiny funky things that matter™.

 

[gdansk]

This product was never gaming focused now it's clear.

How can you be sure it's not a pivot?
In any case it doesn't matter. It's probably still better at gaming than ML crap because RDNA software. But it won't be priced in a way that's appealing to gamers.

 

[insertcarehere]

Wow. It could be a typo. 4nm would actually make more sense from a product design perspective. All their other APUs are 4nm.

Strix Halo would still cost about as much to produce as the AD103 in the RTX 4080 though.

Assuming the previous leaks are true, the GPU+IOD is over 300mm^2, must be pretty expensive on N4.

Apart from consoles, pretty sure that's the largest monolithic die AMD still makes at the moment?

 

[Meteor Late]

How can you be sure it's not a pivot?
In any case it doesn't matter. It's probably still better at gaming than ML crap because RDNA software. But it won't be priced in a way that's appealing to gamers.

And that's why it was never gaming focused.
This thing is supposedly competing against RTX 4060 mobile in terms of gaming performance, we will see of course, that's what I've seen in most of the rumors and data so far.
Thing is, RTX 4060, even RTX 4070 mobile GPUs, are very cheap to make, these are 160-190 mm2 cards with 128 bit bus.
AMD is using like 300 mm2 single die with like two 70 mm2 dies and then the packaging to connect it, along with very expensive motherboard because of the higher bus width and all the requirements Strix Halo needs.

 

[gdansk]

Apart from consoles, pretty sure that's the largest monolithic die AMD still makes at the moment?

Servers have bigger IOd.

 

[gdansk]

Thing is, RTX 4060, even RTX 4070 mobile GPUs, are very cheap to make, these are 160-190 mm2 cards with 128 bit bus.

Yes, but you have to look at a total system. Significant portions of that IOd are wasted on IO which aren't included in that comparison.
The problem, as usual in this board, is that RDNA3 sucks even its slightly ameliorated form.

 

[adroc_thurston]

is that RDNA3 sucks even its slightly ameliorated form.

Uh no it doesn't.

 

[Meteor Late]

Yes, but you have to look at a total system. Significant portions of that IOd are wasted on IO which aren't included in that comparison.
The problem, as usual in this board, is that RDNA3 sucks even its slightly ameliorated form.

Of course, but for gaming, you can pair an RTX 4060 with dirt cheap CPUs, like for instance 12th gen I5 or Hawk Point and it will be absolutely fine, that's what makes the system very cheap to game on.

 

[gdansk]

Of course, but for gaming, you can pair an RTX 4060 with dirt cheap CPUs, like for instance 12th gen I5 or Hawk Point and it will be absolutely fine.

Yep, that's the better value approach. But as a system Strix Halo offers something for companies trying to make a Apple-like form-over-function device. It seems few took AMD up on that.

Uh no it doesn't.

OK, let me rephrase it - as manifest in all shipping configurations it is not PPA competitive to Nvidia. And Strix Point even managed to make LNL's BMG look like not a turd. Impressive work from RDNA3.5.

 

[adroc_thurston]

as manifest in all shipping configurations it is not PPA competitive to Nvidia

Have you seen N1x?
Also N33 is about 204mm^2 of N6 and it's good nuff.

And Strix Point even managed to make LNL's BMG look like not a turd

Not comparable SoCs at all. Next.

 

[igor_kavinski]

Of course, but for gaming, you can pair an RTX 4060 with dirt cheap CPUs, like for instance 12th gen I5 or Hawk Point and it will be absolutely fine, that's what makes the system very cheap to game on.

Except when it runs out of VRAM. Then 4060 will tank. Strix Halo, even lowest configuration, may manage to provide playable experience with FSR3 or maybe even FSR4.

 

[Meteor Late]

Except when it runs out of VRAM. Then 4060 will tank. Strix Halo, even lowest configuration, may manage to provide playable experience with FSR3 or maybe even FSR4.

That is a good point. However, we are still talking about fairly low end gaming here, more VRAM will never justify such a hefty price premium.

 

[GTracing]

I think some people are overestimating how much Strix Halo costs. It isn't that expensive. Plugging die size for the GPUIO die into a calculator gives 148 good dies. At $17000 per wafer, that's $114.86 per die.



The CCDs are around $20 each. Even after the advanced packaging, the SKU costs less than $200 to manufacture.