News AMD's game-changing Strix Halo APU, formerly Ryzen AI Max, poses for new die shots

[Admin]

New die shots offer a glimpse of the intricacies that went into designing AMD's flagship Strix Point or Ryzen AI Max APUs.

AMD's game-changing Strix Halo APU, formerly Ryzen AI Max, poses for new die shots : Read more

 

[heffeque]

There are benchmarks of the Asus laptop.

TDP is limited below what the chip can use, but performance is still great.

What worries me most is pricing.
$2500 for the 32 GB version seems extremely excessive, even if performance is as good as a 4070 laptop on same TDP.

Hopefully there will be 64 GB and 128 GB versions with full TDP usage for mini-PC out there that aren't stupidly expensive.

 

[Pemalite]

This makes me moist... Happily snap one up in a laptop if it's the right price.

 

[GenericUser2001]

Looks really good; personally I am hoping they will get the 8 core, 32 CU version into mid-range gaming laptops soon enough. 16 cores is overkill for my needs.

 

[Gururu]

I never understood dedicated graphics in laptops. This is the future.

 

[HideOut]

Holy worn out expressions bat man. I feel like Im listening to some BS from Steven A Smith on ESPN. Continuuously saying "game changa". Its a new laptop that in every review today is sitll a bit behind the M4 in many categories. Keep in mind, apple has likely either a "bigger" M4 if needed, but more than likely just a fem months from the m5 reveal. This is the most powerfull X86 moble to day, but its not a game c hanger. its maybe at best "somewhat catching up, for 20 more watts".

 

[JamesSneed]

I never understood dedicated graphics in laptops. This is the future.

You could only fit so many transistors on a die so dedicated it was. Now we have density and power reductions due to the latest TSMC nodes and its all possible. Apple did this a bit sooner than AMD but still all of it hinged on being able to make a CPU and GPU powerful enough that fit in the power envelope of a laptop. This will make laptops what they should have been all along(we know because we have the mac's to prove it).

 

[usertests]

Hopefully there will be 64 GB and 128 GB versions with full TDP usage for mini-PC out there that aren't stupidly expensive.

What would be great is if we could add memory ourselves. If it's limited to LPDDR5X, then LPCAMM at least exists. How can quad-channel be handled? Maybe two stacks of two CAMMs, but I don't know if stacked CAMM has been implemented anywhere or merely suggested by JEDEC.

Hopefully AMD learns some lessons from Strix Halo and comes out swinging with Medusa Halo a couple years from now.

 

[markhahn]

I hope the yields are terrible, so it becomes worthwhile to sell bins with fewer cores and CUs - heck, even half-functional LLC.

Gamers may wet themselves, but the prospect of two CCXs over a lower-latency link, and 256b memory is exciting for a much TAM, even with the GPU binned way down.

 

[Gillerer]

I hope the yields are terrible, so it becomes worthwhile to sell bins with fewer cores and CUs - heck, even half-functional LLC.

Gamers may wet themselves, but the prospect of two CCXs over a lower-latency link, and 256b memory is exciting for a much TAM, even with the GPU binned way down.

They wouldn't waste two functional CCXes on a heavily cut down I/O / graphics die. You'd get the gaming version; just 8 cores on a single CCX. And only two memory channels for the lowest model.

 

[bit_user]

Its a new laptop that in every review today is sitll a bit behind the M4 in many categories.

Sounds like someone is jealous of the attention it's getting?

It has more P-cores than even a M4 Max, which is a 12P + 4E configuration. And it has full AVX-512, whereas the M4 has a basically non-functional subset of SVE + 128-bit NEON. So, it's got some things going for it. Not to mention properly-supported native Linux.

 

[bit_user]

You could only fit so many transistors on a die so dedicated it was.

Nonsense. The PS4 had a monolithic APU and it launched 12 years ago!

An AMD exec actually said why they never did it before. It wasn't for any technical reason, but simply because they didn't believe there was a market for laptops with a large iGPU. That is, until Apple proved otherwise.

https://www.engadget.com/computing/...ly-wouldnt-exist-without-apple-220034111.html
​

 

[Mr Majestyk]

$2.5KUS for only 32GB models, so full phat 16+40 models with 128GB in Australia will probably hit $8K, even worse than I thought possible. No laptop is worth that sort of money outside of a paid for by your company workstation, where they were the depreciation.

 

[usertests]

I hope the yields are terrible, so it becomes worthwhile to sell bins with fewer cores and CUs - heck, even half-functional LLC.

Gamers may wet themselves, but the prospect of two CCXs over a lower-latency link, and 256b memory is exciting for a much TAM, even with the GPU binned way down.

Obviously the CCDs are going to have high yields, so subtract that from the overall size to get ~307mm^2 for the I/O+GPU. Which is big but not outlandish.

It sounds like Strix Halo is the test run for silicon bridge interconnects to be used across all Zen 6 products. So if there are latency benefits found here, those are carrying over to all desktop and mobile products next generation. Unfortunately, we are stuck at 128-bit for most products as long as the AM5 socket is around.

 

[suryasans]

If the SoC is going to power $ 1000 laptop segment with 32gb LPDDR5 8****, I will buy it.

 

[Alvar "Miles" Udell]

I don't see how this is "game changing". RTX 4070 based laptops with 16 core AMD CPUs, like the ASUS Strix G17, can be found for under $1300, and the Acer 14" Predator Helios Neo 14 is currently $1250 with an Intel 16 core and 32GB RAM.

The ASUS ROG Flow Z13 is $2300.

 

[kealii123]

Rumors are that the yields for this sort of thing are terrible, hence the initial high price and seeming last minute price hike as well. IDK how Apple does it since they are all made by the same TSMC. Apparently thats the Achilles heel of this type of device: Since its all together on the same die, its harder to QC and sort , mix and match, etc. One failure on any part renders the whole thing worthless.


For those saying this is "game changing", remember that Dell laptop with an Intel CPU that had an AMD graphics chip on the same die?

Dell XPS 15 2-in-1 review: Meet the child of Intel and AMD’s unholy union

A chipset that combines Intel’s CPU with AMD’s GPU makes the XPS 15 thinner.

arstechnica.com

 

[bit_user]

Apparently thats the Achilles heel of this type of device: Since its all together on the same die, its harder to QC and sort , mix and match, etc. One failure on any part renders the whole thing worthless.

What is all together on the same die? You know the CPU cores are on separate chiplets, right?

For those saying this is "game changing", remember that Dell laptop with an Intel CPU that had an AMD graphics chip on the same die?

Dell XPS 15 2-in-1 review: Meet the child of Intel and AMD’s unholy union

A chipset that combines Intel’s CPU with AMD’s GPU makes the XPS 15 thinner.

arstechnica.com

No, they shared a package, but different dies. In that case, the GPU even had its own HBM memory and was only connected to the CPU via PCIe.

In this case, the iGPU is on the same infinity fabric and shares the same memory as the CPU cores.

 

[usertests]

Rumors are that the yields for this sort of thing are terrible, hence the initial high price and seeming last minute price hike as well.

I've only seen these rumors today, is Intel spreading them?

It's a 307mm^2 die, basically the same size as the Navi 31 GCD, and a lot smaller than what Nvidia has made on these nodes.

 

[thestryker]

What would be great is if we could add memory ourselves. If it's limited to LPDDR5X, then LPCAMM at least exists. How can quad-channel be handled?

I've seen no implementations over 128-bit so far, but since there are enterprise versions of the CAMM2 specifications it must be possible. I would assume this would require wiring multiple channels similarly to DIMMs, but none of the documentation I've seen has diagrams of multiple CAMM2 modules being used*.

*There is a single channel CAMM2 module design which allows two modules to be stacked on top of one another to create dual channel. The implication being that you could get higher capacity this way than just a single dual channel CAMM2 module, but I've not seen this implemented at all.

Maybe two stacks of two CAMMs, but I don't know if stacked CAMM has been implemented anywhere or merely suggested by JEDEC.

Stacked as you mean is 100% impossible it would require two completely separate connectors and there doesn't appear to be any allowance in the specification to have one taller than the other so you could mount them facing each other to save space.

If you want to look through the JEDEC specifications it is free with registration and the document is JESD318A.

 

[thestryker]

These seem like a test vehicle for both a big APU and silicon interconnect. I really like the idea and would love to see a 385 in at the very least a minipc as it would be great if the price was right. I wonder how hard of a limit AMD's configurable TDP is for manufacturers because I know it can go below the 45W minimum, but I don't know if that means it could be sold like that. If it could what a handheld that could potentially make (The Phawx overrode the limits on the 390/395 for his benchmarking) although entirely possible it would be cost prohibitive.

 

[usertests]

Stacked as you mean is 100% impossible it would require two completely separate connectors and there doesn't appear to be any allowance in the specification to have one taller than the other so you could mount them facing each other to save space.

Sorry I wasn't being clear enough, I did mean two entirely separate CAMM stacks, each with two CAMMs as that is the apparent limit for now.

Since it doesn't appear we will get Strix Halo with upgradeable memory, there's another 1-2 years for the industry to hash out the future of memory before the Zen 6 successor comes out. Even Nvidia is in the mix with "SOCAMM" in Project Digits.

These seem like a test vehicle for both a big APU and silicon interconnect. I really like the idea and would love to see a 385 in at the very least a minipc as it would be great if the price was right. I wonder how hard of a limit AMD's configurable TDP is for manufacturers because I know it can go below the 45W minimum, but I don't know if that means it could be sold like that. If it could what a handheld that could potentially make (The Phawx overrode the limits on the 390/395 for his benchmarking) although entirely possible it would be cost prohibitive.

We've seen ROG Ally with an official 10W mode, so I think these companies can go below what AMD recommends.

The 8-core 385 certainly looks like the sweet spot, with the use case depending on the amount of RAM included. A lot of people scoffed at the idea of a Strix Halo handheld but it doesn't necessarily need 45 Watts for gaming, and other handhelds have gone the opposite direction.

The FP11 socket is 37.5mm x 45mm, while FP8 (Phoenix, Strix Point) is 25mm x 40mm. It also requires enough memory modules around it to reach quad-channel, otherwise there's little point. Some of the "handhelds" (which are part of a bubble that will burst) are quite large so fitting Strix Halo in one shouldn't be impossible.

 

[Notton]

There are benchmarks of the Asus laptop.

TDP is limited below what the chip can use, but performance is still great.

What worries me most is pricing.
$2500 for the 32 GB version seems extremely excessive, even if performance is as good as a 4070 laptop on same TDP.

Hopefully there will be 64 GB and 128 GB versions with full TDP usage for mini-PC out there that aren't stupidly expensive.

Where are you seeing $2500 for the 32GB model?

On Asus' website, they have the Z13 (2025/GZ302) listed as
$2200~$2300 USD for the 395+/32GB/1TB model. (Prices went up by $100 since a few days ago, for some reason)
$2000~$2100 USD for the 390/32GB/1TB model. (ditto above)
$2800 for the 395+/128GB/1TB model.

On BestBuy they have the 395+/64GB/1TB listed for $2200. Although, weirdly, they have the 395+/32GB/1TB model also listed for $2200.

 

[thestryker]

The 8-core 385 certainly looks like the sweet spot, with the use case depending on the amount of RAM included. A lot of people scoffed at the idea of a Strix Halo handheld but it doesn't necessarily need 45 Watts for gaming, and other handhelds have gone the opposite direction.

Yeah the results in The Phawx's video seemed to indicate 20W is when the 395 just wins everything gaming wise so I assume the rest would likely be the same despite fewer CUs. Though the improvement from 32CU to 40CU is very real just the same and it'd be great to see that on lower core counts (probably next generation).

 

[dada_dave]

Where are you seeing $2500 for the 32GB model?

On Asus' website, they have the Z13 (2025/GZ302) listed as
$2200~$2300 USD for the 395+/32GB/1TB model. (Prices went up by $100 since a few days ago, for some reason)
$2000~$2100 USD for the 390/32GB/1TB model. (ditto above)
$2800 for the 395+/128GB/1TB model.

On BestBuy they have the 395+/64GB/1TB listed for $2200. Although, weirdly, they have the 395+/32GB/1TB model also listed for $2200.

Some of the listing at stores, maybe even Asus, seem to be placeholders (why they keep changing or don't make sense). The reviewers I've seen for the Asus device have all mentioned that the review units they received, the 395+/1TB/32GB model, were priced at $2500 or 2500 Euros:

e.g.
https://www.notebookcheck.net/Asus-...d-neck-with-the-RTX-4070-Laptop.963266.0.html

So what the price actually is seems to be an issue.

Sounds like someone is jealous of the attention it's getting?

It has more P-cores than even a M4 Max, which is a 12P + 4E configuration. And it has full AVX-512, whereas the M4 has a basically non-functional subset of SVE + 128-bit NEON. So, it's got some things going for it. Not to mention properly-supported native Linux.

I'm not 100% sure what the non-functional SVE subset you refer to is, could you clarify? The M4 does have SSVE and SME using the AMX units which is Apple's currently preferred 512bit solution.

Also ... yes the Halo has more cores than the M4 Max, but not the performance of one. I'm not denigrating the Halo, AMD themselves are positioning it as an M4 Pro competitor which is probably about right and it should net a share of performance (though not efficiency*) wins - especially in the GPU where, with a few exceptions, it seems to perform in-between an 20-core M4 Pro and an 32-core M4 Max**.

I'm not in agreement with the other user either. As you point out in a later post yourself, Apple may have changed the PC game first*** and created the permission for AMD to try this themselves in a PC chip, but I agree that AMD creating is still a huge step, a game-changer for non-Macs certainly, and based on reviews it seems to be successful technically****. Hopefully it will be commercially as well.

*While Cinebench R24 is a benchmark Apple does particularly well in, the Halo requires more power than the 16-core M4 Max to achieve slightly less than the performance of the 14-core Pro - it's using 30W more than the M4 Pro.

**Obviously gaming, especially the huge library of non-native Mac games, is going to be a massive advantage for the Halo over an M4 even beyond what the benchmarks say.

***Although as you also point out: consoles exist. Heck even in the PC space APUs like this is what AMD so clearly wanted for years to make a reality. Unfortunately AMD just didn't have the necessary ingredients until recently or the budget to really push the concept for PCs. Apple did.

****I have to admit I thought it'd be a touch more efficient. But I think it is using the N4X node rather than the N4P for the CPUs, yes? People keep saying they want to see it run at higher power settings, but I'd like to see more tests and wall power measurements at lower power setting. I have a feeling that 60-70W wall power for the CPU (not TDP) might be the sweet spot for this chip, especially if it primarily meant for laptops. Also, for the GPU it's a pity the timing couldn't have worked out to make this RDNA 4. Obviously that wasn't possible, but, like with the M3/4, the extra ray tracing capabilities would've also been very cool to have with all the potential VRAM.