AMD Ryzen AI 300 Pro Series Could be Equipped with up to 128 GB of Memory

[Tek-Check]

AMD isn't really competing with Intel or QC at the moment their real target should be Apple!

They already have twice as much laptop market share in comparion to Apple.
Strix Halo laptops and mini-PCs will, of course, compete with some Apple designs, but also with Intel laptops hosting discrete Nvidia GPUs.
Mini-PCs with this chip will often be more performant than full DIY systems.
So, a new opening in PC is about to happen next year with systems hosting Strix Halo chips.
There should be entire segment, with 4-5 SKUs, from top Halo to entry systems.

 

[Wirko]

Top Strix Halo SKU should have LPDDR5X at 8000 or 8500 MT/s, so 256-270 GB/s.

That will require at least 4-dies-high stacking, probably TSV because of high speed. That looks awfully expensive, like those ultra-expensive server DIMMs up to 512 GB, also with stacked dies.

 

[Minus Infinity]

Top Strix Halo SKU should have LPDDR5X at 8000 or 8500 MT/s, so 256-270 GB/s.

Halo is also moving to 256 bit bus, RDNA3.5, will support much higher speed (LP)DDR5 memory as you point out and is rated at 120W TDP. It should be able to deliver a huge uplift compared to 780M. I wouldn't be surprised to see 7600XT+ levels of performance.

 

[Vayra86]

The comparison is with mobile x70 card, not desktop. About 4070 mobile, this card is insult for the x70 tier so its very achievable, 25% slower then 7600xt which has 32 CUs, 20% less than Strix Halo. Also 7600xt has 288GB/s bandwidth vs ~200GB/s shared between the CPU and GPU.
Plus this is RDNA 3.5 vs RDNA 3, 780m 12CU is already as fast as the desktop rx 6400 with 12CU, so I expect it to be at least as fast as 4070 mobile
View attachment 352646

hahaha wow an x70 (!!) mobile is 20% slower than an x60 desktop nowadays? Oh man the sadness of marketing in mobile GPUs knows no bounds. Its like they progress backwards

AMD APU with On-package ultra low latency RAM?
Been waiting for this for *years*.

Now, do HBM

God no, please no. We've seen where that ends in consumer GPU

 

[ADB1979]

Well, despite what some people may say, I think Strix Halo will be a revolutionary product and will fundamentally shake up the product stack and offer the most competition to nvidia in years, and most interestingly, not from the above (halo products), but below (entry and mid range).

Strix Halo will be the (latest) reformation of the Workstation / HEDT market and will create a whole new entrant to the sector. Interestingly this is actually a lot like the SGI Xeon systems, their big selling point was that the graphics had access to vast amounts of RAM, far more than their competitors and for some that was an instant sell.

It will be interesting to see the various uses, custom software, tweaks, etc that people do to Strix Halo systems when they end up in the hands of consumers, this will be self perpetuating for a while as people keep on tweaking and generating more interest and competition, plus we all get to learn about the various things that can be done with huge amounts of graphics RAM, itself a preview into the future.

There should be entire segment, with 4-5 SKUs, from top Halo to entry systems.

There is a list leaked by MLID (albeit it's very old) that lists 4-5 SKU's, although he noted that one of them makes no sense except to use up every piece of silicon regardless (because everything (literally) is selling) so why not.

Personally I want to see a NUC style mini PC with 2x 128-bit LPCAMM2 (or CAMM2) modules for the DIY market, plus one in some kind of ATX form factor to make use of the (limited) PCIe lanes.

 

[Wirko]

plus we all get to learn about the various things that can be done with huge amounts of graphics RAM

Fill it with various LLMs... I mean, what else?

 

[Sound_Card]

Credit Tomshardware: but I am sticking to my internal hype that Halo is going to be at least as fast as the 6800m. The highest end modal at least. 32mb of Infinity Cache and 64mb of L3. This APU is designed to murder the industry.

EDIT: Also, didn't the 6800m only have 8x lanes of PCIe 4.0? and was limited to DDR4 system ram? That is when it ran out of its GDDR memory. I think there will be situations where the 6800m (paired with the 6900hx) comes out on top, and situations where this APU comes out on top, but I think HALO is going to come out on top more.

 

[Tek-Check]

That will require at least 4-dies-high stacking, probably TSV because of high speed. That looks awfully expensive, like those ultra-expensive server DIMMs up to 512 GB, also with stacked dies.

Videocardz says the top SKU competes with 4070 mobile.

That will require at least 4-dies-high stacking, probably TSV because of high speed. That looks awfully expensive, like those ultra-expensive server DIMMs up to 512 GB, also with stacked dies.

No stacking. Three dies, most likely.

 

[Wirko]

No stacking. Three dies, most likely.

Stacking of LPDDR dies is what I mean. If there are 8 packages, the capacity of each package has to be 16 GB or 128 Gbit. Largest dies that (almost) exist are 32 Gbit.

 

[Tek-Check]

Stacking of LPDDR dies is what I mean. If there are 8 packages, the capacity of each package has to be 16 GB or 128 Gbit. Largest dies that (almost) exist are 32 Gbit.

There must be higher capacity chips in existence for LPCAMM2 module.

I can see Samsung has on their website several 128Gbit modules.

 

[Wirko]

There must be higher capacity chips in existence for LPCAMM2 module.

I can see Samsung has on their website several 128Gbit modules.

Yes but they're always stacked. Even higher than I thought, as I see 16-dies-tall stacks are possible. It checks out:

one die = 16 Gbit = 2 GB
up to 16 dies = up to 32 GB per package
four packages = up to 128 GB.

LPCAMM2: No compromise for next-gen laptops

Micron's new memory type, LPCAMM2, enables the best aspects of no-compromise thin and light PCs. This modular LPDDR5X-based memory solution offers lower power consumption, increased performance and the ability to upgrade all while maintaining a thinner and lighter form factor.

www.micron.com

 

[Tek-Check]

Yes but they're always stacked. Even higher than I thought, as I see 16-dies-tall stacks are possible. It checks out:

one die = 16 Gbit = 2 GB
up to 16 dies = up to 32 GB per package
four packages = up to 128 GB.

LPCAMM2: No compromise for next-gen laptops

Micron's new memory type, LPCAMM2, enables the best aspects of no-compromise thin and light PCs. This modular LPDDR5X-based memory solution offers lower power consumption, increased performance and the ability to upgrade all while maintaining a thinner and lighter form factor.

www.micron.com

Also, some laptops can simply have two LPCAMM2 modules, each 64GB on 128-bit bus.

 

[Sound_Card]

There is no way the upper SKU's of Strix Halo are going to make it into Laptops. Nvidia is not going to allow it. Instead, I see Strix Halo dominating mini-pc's and quite potentially being the forerunner to the next consoles and handhelds. I do see Strix Point being the industry darling for all other laptops not radically marketed for gaming.

 

[R0H1T]

Nvidia is not going to allow it.

Ok but who the heck is Nvidia to not "allow" it?

 

[Yashyyyk]

There is no way the upper SKU's of Strix Halo are going to make it into Laptops. Nvidia is not going to allow it. Instead, I see Strix Halo dominating mini-pc's and quite potentially being the forerunner to the next consoles and handhelds. I do see Strix Point being the industry darling for all other laptops not radically marketed for gaming.

Why would consoles pay for extra CPU power they won't use when they are GPU-bound (and have not much more GPU performance than the current consoles)?

Nvidia needs to get rid of their VRAM allergy and give better memory bandwidth if they care, their cards are already more efficient

 

[Wirko]

Also, some laptops can simply have two LPCAMM2 modules, each 64GB onn 128-bit bus.

This has been demonstrated for CAMM2. Its connector is made up of two parallel sections that look identical to each other, so using two half-bit-width modules is possible, apparently. LPCAMM2 doesn't seem to be designed with this in mind.

 

[Tek-Check]

This has been demonstrated for CAMM2. Its connector is made up of two parallel sections that look identical to each other, so using two half-bit-width modules is possible, apparently. LPCAMM2 doesn't seem to be designed with this in mind.

Two full separate modules, one 128-bit interface for each, on each side of the APU.

 

[Wirko]

Two full separate modules, one 128-bit interface for each, on each side of the APU.

Yes, that's a possibility too.

 

[hsew]

Technically this is a regular GPU chip integrated into an MCM design, so if you consider this an iGPU, also consider that the mi300A is an APU that fulfills your nonsensical sarcastic dream.

Let’s completely ignore the clarification I made after that initial post and throw in yet another product in a completely different and specialized market segment… which is also operating in a power envelope an entire order of magnitude above laptop iGPUs! MCM or not is obviously irrelevant to your point, as the claim originally made in the article is speculation about an upcoming product as a whole. As it is, ahem, INTEGRATED INTO THE SOC AS A WHOLE, AS OPPOSED TO BEING A DISCRETE, STANDALONE GPU.

Christ…