News AMD pushes Ryzen to the Max — Ryzen AI Max 300 Strix Halo reportedly has up to 16 Zen 5 cores and 40 RDNA 3+ CUs

[abufrejoval]

AMD totally missing out on an opportunity to claim theirs goes up to 11.
Instead, they copy Apple now.

AMD has Intel as an excellent example how aiming for 11 can have you shoot yourself in the foot: they could never afford that type of mistake and Intel will fall because they did it once too many.

80432, i860, Itanium, Xeon Phi etc. the list of Intel moonshots that fizzled is rather impressive and it seems like Intel engineers aimed far to much at internal competition to see what was coming to their "home turf".

AMD innovates in a different manner and as a whole seems to target only outdoing outside competitors.

Both going wider and going with co-packaged DRAM aren't exactly pure fruity cult innovations. Going with a scalable 1/2/4 chip design using a common base probably is: but that's not even what is happening here, so I don't see the fruity cult imitation.

 

[abufrejoval]

BGA only
There's no way it will go into an AM5 socket when it has a 256-bit memory bus. That would require more pins.
Although I wouldn't be surprised if someone hacks it onto a socket near the end of its life cycle.

256 LPDDR5 means 8x 32-bit LPDDR5 channels or 4x 64-bit DDR5 channels: no idea how costly it is to support both in terms of IP block size.

Four LPDDR5 channels could be serviced via on-package dies, leaving four to go off socket to CAMM or even to two DDR5 DIMMs.

That's not so far from an Alder-Lake notebook I run, which has one channel soldered with 8GB and the other using 32GB in SO-DIMM socket, which isn't perfect but much better at running multiple VMs with 40GB while the most bandwidth hungry part of the SoC, the iGPU is serviced from the bottom part of the physical RAM using both channels.

While bandwidth is absolutely crucial for GPUs, on CPUs access patterns are typically so chaotic, that caches do the job better than ultra wide RAM (not counting LLMs).

Now, of course, there is a non-trivial real-estate issue on AM5: that new SuperGPU IOD is bound to be huge, no matter what they use to fab it, because there is tons of I/O that's really hard to squeeze and plenty of SRAM within the GPU which isn't easy either. And that's before adding on-packge DRAM.

But if you go wild and into another layer, like the notebook SoCs being fitted to desktop sockets via an organic die carrier, things could be done in such a manner, that only half of the channels go through the socket, enabling an AM5 form factor.

But realistically such desperate measures seem unlikely, because the balance between that extra effort and just having Strix Halo APU soldered to a mainboard is too little value in a niche within a niche.

However, mainboards with such an asymmetric design, where 50% RAM in soldered LPDDR5 and 50% ready for upgrades as CAMM or DDR5 SO-DIMMs may perhaps not be that far off, if the memory controller on the Strix Halo IOD actually supports that; or at least I hope so, because I tend to want far more RAM than mainstream hardware provides.

 

[Guardians Bane]

That's ~$150 more than the prior gen version of the same thing which is a pretty heavy premium to pay contextually speaking. It also explains the high cost floor for the laptops currently using the 370 which are on the market now. It's not that these devices are all bad deals by any means just that the price has to go up for something that has low margins (or in the case of huge companies raising to make their minimum acceptable).

Gotchya. I am going to go for a Point mini for now. I just hope (probably in vain) a Halo based mini isn't priced out of reach for me. I already came to terms with not getting a laptop with Point or Halo. Point laptops are already too pricey for me. And with the AooStar system I hope to get, I'll probably just have to be happy with. I do like mini PCs and spend way too much already on shiny new tech! Lol...

 

[abufrejoval]

So up to 96GB of ram for GPU allocation means that we can expect 96GB and 128GB memory configurations for any systems using these new SoCs?

The main thing with AMD APUs since Llano is that RAM used for the iGPU and the CPU is the same.

Now, if only for backward compatibility, the portion actually used by the iGPU was typically cordoned off and not given to the OS to manage e.g. including paging, but could have been any part and any portion of physical RAM.

And then most APUs and their drivers have mechanisms by which OS apportioned RAM might actually be re-purposed for the iGPU, locked in place and excluded from running applications.

On Strix Halo, like on all APUs since Llano, the very same RAM can be used by the GPU, the CPU and probably the NPU and basically anyone on the PCIe bus laying claim, so 96GB in my eyes just expresses that in a system using 48GB DIMMs or CAMMs, nearly all of that RAM is also available for GPU usage.

In AMD's original HSA concepts RAM managed by the OS could still be used by the G[P]GPU: it supports function call level control transfer (no OS involvement) between GPU and CPU cores in shared RAM.

I have no idea how much of that survives and which frameworks or APIs make use of that, but any APU "reservations" or allocations are mostly done for convenience and legacy support: anything really modern shouldn't have to care and make no distiction in a unified memory space.

What this means in terms of memory capacities or upgrade potential is a whole new can of worms and mostly wild guessing at this point in time.

 

[thestryker]

I do like mini PCs and spend way too much already on shiny new tech! Lol...

I do this with passively cooled router PCs, but I don't think I'll need to move beyond my current N305 box... but it is only 2.5gbps sooooooooooooo there's always 10gbps

 

[Guardians Bane]

I do this with passively cooled router PCs, but I don't think I'll need to move beyond my current N305 box... but it is only 2.5gbps sooooooooooooo there's always 10gbps

Yep, tech enthusiasts or tech addicts? I traded alcohol (7 years sober) for tech, sooo....

 

[thestryker]

I thought controller is built into the SOC , not suggesting to use AM5 but build a board that can support this processor utilizing the required socket pinning. The point someone could build an atx eatx … board to use in desktop, soldering it appropriately to the board in the appropriate bga format. Ultimately, the question is there a market for it … considering the cost of discrete graphics and cpu today. Seems this could be a viable product but if I truly were AMD may considered doing it?

There are companies (as I mentioned Erying) who do exactly that, but they usually use DIMMs which would require quad channel (if Strix Halo's memory controller even supports DIMMs) and there aren't any current boards using quad channel memory topologies except workstation. Not to say it's impossible, but that the deck is very much stacked against this sort of thing happening.

 

[systemBuilder_49]

Hopefully they're the very last Radeon GPUs that suck at ray tracing (3.5) ....

After extensive study I realized the hx370 is too bottlenecked by ram access so I bought an hx365 laptop because most of the 800m series uplift is from the architecture - the extra 4 CUs in the 890m are mostly wasted...

 

[systemBuilder_49]

>If the performance is high enough I could see this ending up in gaming laptops.

It'll depend on SoC pricing. The gaming laptop segment is already well served by dGPUs, ..

Building a high-powered gaming laptop with one single fan would be a breakthrough - cheaper and better than today's two and three fan laptops!

 

[mac_angel]

It will definitely show up in gaming laptops, which are hardly a niche unless you are just comparing it to the overall laptop market. That 32 MiB of Infinity Cache isn't for the AI workloads. (Consumer) Mini-PCs are an afterthought, coming after laptops. I mentioned it because in those, battery life doesn't matter.


Handhelds come later than laptops. AMD has confirmed the Z2 Extreme is coming. That's highly likely to be a carbon copy of the HX 370, as the Z1 Extreme was just the top Phoenix chip.


The problem is that it's likely to be very expensive, more than CPU+GPU combos of similar or better performance.

If compactness is the goal at any price, maybe. Soldered to a Mini-ITX board might even be larger than necessary.

looks like you're right
https://www.tomshardware.com/deskto...unt-from-the-regular-dollar1249-listing-price

 

[usertests]

looks like you're right
https://www.tomshardware.com/deskto...unt-from-the-regular-dollar1249-listing-price

That's Strix Point, not Halo.

Unless you meant Halo is going to be expensive since we have a top Strix Point Mini-PC going for $1,000, in which case.... yeah.

 

[Hartemis]

Bring this architecture to the next gen desktop APUs.

No market for this. Obviously, 16c/32t + 40CU + x NPU should cost more than 9950X. Enthousiast gamers prefer to byy a 9950X (or a 9850X3D) + a dedicated RTX.
AMD thinks his desktop APU as low end /entry level market. Ryzen AI Max will not fit it.

After extensive study I realized the hx370 is too bottlenecked by ram access so I bought an hx365 laptop because most of the 800m series uplift is from the architecture - the extra 4 CUs in the 890m are mostly wasted...

Strix Halo has Infinity Cache.