News AMD Strix Halo Mini-ITX motherboard flaunts 128GB LPDDR5X — add a CPU cooler, boot drive, and power supply for a slim gaming or AI rig

Toggle sidebar Toggle sidebar
Notton said:
Doesn't Strix Halo have 12x PCIe 4.0 lanes? Where are the other 4x?
Click to expand...
No PCIe slot. No USB4 ports.

This is inferior to the Framework board and the HP mini PC. The Framework board exposes almost all of the I/O native to the SOC. And the HP mini PC has ECC RAM.

The only thing left for this board to compete on is price.
 
Last edited:
Li Ken-un said:
No PCIe slot. No USB4 ports.

This is inferior to the Framework board and the HP mini PC. The Framework board exposes almost all of the I/O native to the SOC. And the HP mini PC has ECC RAM.

The only thing left for this board to compete on is price.
Click to expand...
With its lack luster expansion, it had better be cheaper.
 
MobileJAD said:
With its lack luster expansion, it had better be cheaper.
Click to expand...
It does look cheap, and it has the full AI capacity of Strix Halo in a thin ITX, the kind that often goes behind a monitor.
I imagine the ram is not the greatest for gaming with a dGPU with any of these, but that won't hurt productivity. Also the iGPU may be on par with a discrete B580 if you can give it over 100w which is quite good for many. You get a real nice CPU for productivity, a ton of ram and an iGPU that ranks up there with entry level discrete. In the smallest package available on a standard sized motherboard.
If it is $700 or less and is what you are looking for it is quite the deal.

My only concern is the ram latency might make the user experience kind of jerky with some hang times like the vastly weaker older atom mini pcs. But I have no evidence at all that this will be the case and reviews should address this.
 
Li Ken-un said:
No PCIe slot. No USB4 ports.

This is inferior to the Framework board and the HP mini PC. The Framework board exposes almost all of the I/O native to the SOC. And the HP mini PC has ECC RAM.

The only thing left for this board to compete on is price.
Click to expand...
Price, overall TDP and form factor is likely where they’re looking to position and compete. There are some very compelling use cases if all of those points can be successfully met.

In the AI space, clustering these with lower overall power consumption and cost while increasing density would be appealing. And for the majority of the gaming crowd, having a gaming PC / console / emulator with 128 GB RAM and a decent SOC graphics engine that fits into a low power and small form factor profile is equally so. But it’s all contingent on meeting all those criteria.
 
BadBoyGreek said:
Price, overall TDP and form factor is likely where they’re looking to position and compete. There are some very compelling use cases if all of those points can be successfully met.

In the AI space, clustering these with lower overall power consumption and cost while increasing density would be appealing. And for the majority of the gaming crowd, having a gaming PC / console / emulator with 128 GB RAM and a decent SOC graphics engine that fits into a low power and small form factor profile is equally so. But it’s all contingent on meeting all those criteria.
Click to expand...
Honesty I wonder if that's the target audience for these, people who want to try building a AI server farm without dipping into speciality and Hella expensive enterprise hardware, the biggest giveaway for that would be if the built-in ethernet ends up being faster than gigabit.
Because if this was meant for the gaming crowd, the USB options would be looking better than they are, even people in the productivity crowd who would need the iGPU for media editing would want something faster than USB 3.0 for transferring files , and unless I read that little image that shows the USB ports wrong , all it showed was USB 2.0 and 3.0.
 
MobileJAD said:
Honesty I wonder if that's the target audience for these, people who want to try building a AI server farm without dipping into speciality and Hella expensive enterprise hardware, the biggest giveaway for that would be if the built-in ethernet ends up being faster than gigabit.
Because if this was meant for the gaming crowd, the USB options would be looking better than they are, even people in the productivity crowd who would need the iGPU for media editing would want something faster than USB 3.0 for transferring files , and unless I read that little image that shows the USB ports wrong , all it showed was USB 2.0 and 3.0.
Click to expand...
These early models could just be the groundwork for what’s possible in the gaming space. For many, the SOC, if it has good FSR capabilities, could provide “good enough” graphical performance as a minimalist console or PC where you just connect controllers / network and are not depending on an eGPU and all storage is handled internally via the M2 interface. As emulation PCs / consoles, these would be absolute beasts.

Agree though that, without USB 3.1+, it doesn’t make the most sense for content creators. But later iterations will likely address many of the early shortcomings. If it can deliver on all its initial promises, the future for these types of products is bright.
 
Why doesn't the article mention, that it uses a 256-bit RAM interface? This is kind of the most important point of this whole setup, since the 128-bit width typically seen on APUs is their Achilles' heel.
 
Notton said:
Presumably standard mounting holes, but doesn't the RAM also require cooling?
Click to expand...
As much as the RAM in your laptop does.

Strix Halo is mostly so fast, because it uses more RAM chips in parallel, which doesn't change heat density, just requires more space, overall energy and obviously more money for extra capacity.

It's LP-DDR5, LP = low power and the idea is that without the need to support DIMM sockets and long circuit board traces, you can reduce the signal voltages significantly, which may be a large part of the overall power consumption.
Notton said:
Doesn't Strix Halo have 12x PCIe 4.0 lanes? Where are the other 4x?
Click to expand...
If you look at the mainboard, it's already rather busy. I guess they could have added another M.2 slot on the backside, which a lot of these boards do. But perhaps the extra traces would have meant adding another layer to the PCB.

The idea of an APU is that it already contains most everything you'd ever want to add in terms of peripherals. And a dGPU would definitely just waste getting a Strix Halo instead of a Fire Range, which will give you practially the same CPU cores but also 24 lanes of PCIe v5 (and half the DRAM bandwidth). Also doubling the pins for RAM means workstation numbers of pins and traces unless you cut elsewhere: can't have it all and cheap.

So AMD already decided to significantly cut down on the PCIe potential of Strix Halo and this board follows the trend a bit further, probably in the interest of economy and space: from the looks of it is an OEM design they also happen to sell as a bare bone and this Slim-ITX form factor seems to be far more popular in Asia than elsewhere for some reason. In the West this never really caught on.

I sincerely hope it's in the interest and expectation of economy, because I see this as a sign of something I've expected and predicted: that Strix Halo might fail to grab a premium market share both in laptops and in "SuperNUCs" ... and thus become as cheap as they are to make.

Because apart from the obvious cost of the APU, Strix Halo is all about clever use of low-cost commodity parts.

256 bit RAM are obviously expensive in terms of the IOD beach front area, the power amplifiers and the mainboard circuit traces but otherwise 128GB of commodity RAM cost the price of 128GB, no matter if arranged as 128 bit or 256. The extra GPU performance isn't paid on the RAM side as with GDDRx or HBM, but in APU pins and circuit board traces.

But with the current 128 bit RAM bus you have to invest a similar if not bigger effort into the dGPU, which uses more expensive VRAMs, and much tighter ciruit board traces, perhaps requiring a way more expensive PCB.

So in theory Strix Halo should be cheaper to make than an equivalent CPU+dGPU combo, except that Strix Halo is currently sold as if it ran on HBM instead of commodity DDR5.

Apple and Intel Lunar Lake do something similar, but they use somewhat more expensive stacked LP-DDR dies right on the CPU die carrier, which theoretically might command a premium due to packaging complexity (surely not as much as they charge, though).

Strix Halo on the other hand practically avoids all extra cost except the raw silicon cost and the wider RAM bus (compensated via PCIe cuts). It's sacrificing the 10-15 Watt power target for ultra-thin laptops, but delivers 100 Watt workstation power at perhaps half the Wattage.

So it really should be so cheap, it's capable of replacing all entry to mid-range desktops and gamer PCs: if sold at the right price.

My benchmark is a Lenovo LOQ ARP9 laptop I bought for €750 last year. It's only 8 cores but an RTX 4060, which is somewhat faster than Strix Halo as long as 8GB of VRAM last (much better than you'd think with its 1080p resolution).

Add €200 for 128GB of RAM and novelty and a similar Strix Halo laptop should go for €1000, this board for €900 or less.

And at that price you may no longer complain about missing an extra M.2 slot or prefer a Framework for €3000.

Now it just needs to get there and since I cannot see where AMD is selling Strix Halo in volume, this practially needs to happen, because Zen 6 is coming.

I can't quite see AMD only having ordered a few of them, but I also don't quite understand what they were thinking in terms of target mass market, either.

AFAIK Strix Point is not only a bespoke IOD which might never see console use, but also bespoke CCDs not shared with Fire Range. And that violates the mass production scaling benefits their clever chip modularity and reuse provided since Zen.

It was clearly designed for mass market adoption, but half way through its life cycle that hasn't happened, so the price needs to fix that.
 
  • Like
Reactions: DS426
rluker5 said:
This is more like it. Hopefully one can dump extra power into it and get the full potential out of this chip.
I'd like to see how that compares to strix halo in a handheld.
Click to expand...
I don't think that Strix Halo can hit the 10-15 Watt power targets of a hand held with meaningful gaming performance.

It might run with that for 2D desktop work, but that's basically like turning off the dGPU in current mobile workstations with hybrid graphics.

Its TDP target is between 45 and 120 Watts and the entry level mobile workstation or desktop market, but also doesn't go beyond that. In theory you might be able to add an eGPU, but you'd need to be in a special pinch to want that.
 
abufrejoval said:
The idea of an APU is that it already contains most everything you'd ever want to add in terms of peripherals. And a dGPU would definitely just waste getting a Strix Halo instead of a Fire Range, which will give you practially the same CPU cores but also 24 lanes of PCIe v5 (and half the DRAM bandwidth). Also doubling the pins for RAM means workstation numbers of pins and traces unless you cut elsewhere: can't have it all and cheap.
Click to expand...
You don't have to plug in a dGPU, there are plenty of other viable cards to plug in there.
Top choice would be a 10Gbe NIC since the mobo only comes with 1Gbe.
PCIe to NVMe splitter card would be another. https://sabrent.com/products/ec-p3x4
NVMe 2280 to SATA RAID card isn't too difficult to find.
 
abufrejoval said:
I don't think that Strix Halo can hit the 10-15 Watt power targets of a hand held with meaningful gaming performance.
...
Click to expand...
Exactly. Regular Strix Point (up to the HX 375) has a configurable TDP of 15 - 54W -- exactly why Z2 Extreme (comparable to Ryzen 7 AI 360) is feasible in a handheld as it is a Strix Point APU. Strix Halo has a minimum cTDP of of 45W, which pretty much makes it a non-starter. Now, with Zen 6 going to the TSMC 2nm node, I don't doubt that the best "Z" part might be able to utilize 20 or 24 RDNA4 or 4.5 cores in the some power envelope.

I know Zen 6 is finally increasing core density and therefore core counts as the CCD goes from 8 cores to 12, so we can almost guarantee that the top "Z" part will be a 12-core APU... almost have to guarantee that having 50% more cores means less thermal and power headroom for the iGPU, depending of course on clock speeds as well. Most games would still prefer higher-clocked eight cores over lower-clocked twelve, but that's where the fun comes in with AMD's tuning on these chips.
 
  • Like
Reactions: abufrejoval
Notton said:
You don't have to plug in a dGPU, there are plenty of other viable cards to plug in there.
Top choice would be a 10Gbe NIC since the mobo only comes with 1Gbe.
PCIe to NVMe splitter card would be another. https://sabrent.com/products/ec-p3x4
NVMe 2280 to SATA RAID card isn't too difficult to find.
Click to expand...
I couldn't agree more as to what I'd put in there, especially since I got one of those Sabrents, which I used with 4 Samsung 970 Evo+ in my Alder-Lake 12700H, that has practically identical PCIe capabilities, but also an x8 PCIe v4 slot.

It's actually using a (non-Broadcom) PCIe v3 switch, but offers only bifurcation bandwidths (4x1), it's main advantage is price. And aggregate bandwidth is limited by the x4 uplink to the host anyway.

I pilfered the NVMe drives for other systems (Topton NAS variants built on Zen APUs with lots of SATA ports), so the slot is currently holding a B580 instead, while the Sabrent is on the shelf.

Got plenty of M.2 SATA adapters, too, as well as M.2 to PCIe x4 adapters to reuse my Aquantia AQC107 10Gbit NICs.

There is a also a French company that sells M.2 AQC107 NICs with a tiny connector cable to a slot bracket, which helps in a crowded system but cost a bit of premium.
 
  • Like
Reactions: Notton and Li Ken-un
abufrejoval said:
There is a also a French company that sells M.2 AQC107 NICs with a tiny connector cable to a slot bracket, which helps in a crowded system but cost a bit of premium.
Click to expand...
https://www.tomshardware.com/networ...t-slots-into-your-m2-port-costs-only-dollar86

BadBoyGreek said:
In the AI space, clustering these with lower overall power consumption and cost while increasing density would be appealing.
Click to expand...
Framework is selling their AI cluster solution using Thunderbolt/USB4 to link all the units. For the capabilities of Strix Halo, there isn’t possibly a better solution than that.
 
  • Like
Reactions: abufrejoval and Notton
  • Like
Reactions: rluker5
Li Ken-un said:
https://www.tomshardware.com/networ...t-slots-into-your-m2-port-costs-only-dollar86
Click to expand...
That's certainly cheaper than the Kalea variant, but that has a truly tiny cable you can squeeze through near anything:
https://www.kalea-informatique.com/...-ethernet-network-adapter-aquantia-aqc113.htm
Li Ken-un said:
Framework is selling their AI cluster solution using Thunderbolt/USB4 to link all the units. For the capabilities of Strix Halo, there isn’t possibly a better solution than that.
Click to expand...
Nothing better, I'll agree, but that doesn't mean that the result will even be good.

Splitting an LLM means that your performance will be almost entirely dictated by the bandwidth of the interconnect.

You can try that even without a network, just use pretty near any model, LLM or diffusion that's larger than your dGPU and you'll see the performace drop off to basically CPU performance, because it's limited by GPU to [CPU} RAM bandwidth via PCIe.

And that bandwidth and latency is already rather good compared to what you can get on a network or fabric.

Admittedly direct Thunderbolt networking gets you much better latencies than normal Ethernet (and avoids some extra cost if you're happy to accept the scale-out limits of direct connections), but you only get 10GBit bandwidth with IP over Thunderbolt: that's a 1GByte/s bottleneck smack in the middle of your 270GB/s of Strix Halo RAM.

And even if you were able to find anything that connects several Strix Halos latency free, the bandwidth can never exceed what the APU's PCIe interfaces are capable of, which is 16 lanes of PCIe v4 or 32GB/s. Some might want a bit of storage, so half that might be the actual limit.

And at that point you may be just better off using a big CPU that has good support for your weights format, because even at desktop dual channel they can offer twice that speed, more if you go EPYC.

These clusters only make sense if you can split your models in such a way that they won't suffer as much from the node latencies. If that were easy Nvidia wouldn't make billions selling NVlink and DeepSeek wouldn't have to invest man-centuries in inventing bespoke model topologies and software libraries to make things work.

Note that the somewhat Nvidia DIGITS uses ConnectX-7 Infiniband adapters at 400GBit/s or 40Gbyte/s per port. But that doesn't provide transparent scale-up either, so you can't just run inference from bigger LLMs and expect a quad DIGITS to perform as if you'd put its Blackwell bits into a single GPU with 1000GByte/s VRAM.

Please resist AMD trying to sell Strix Halo as an AI wunderkind, just because they really don't have any other market to sell it to. It's really an entry level workstation and gaming system design that's optimized for economy but lacks expandability and a big enough market.

It's great when sold at the right price for the right purpose, both of which AMD wants to avoid somehow, since they were hoping to compete with the Fruity Cult on this one.

I'm pretty sure AMD is loosing big money on these chips right now, because contrary to AMD's typical approach they use bespoke CCDs and IODs that need massive scale to pay off. But their real trick is that outside that APU they are designed for cheap commodity parts, unlike Apple's and Intel's die carrier DDR stacks. That should leave more money for AMD, if users accepted the limitations of Halo and the price premium that goes with it.

And that just doesn't seem to happen when we are already half way through Strix Halo's life cycle.
 
DS426 said:
Most games would still prefer higher-clocked eight cores over lower-clocked twelve, but that's where the fun comes in with AMD's tuning on these chips.
Click to expand...
They auto tune 😉

The thing to remember is that as you light up more cores in a TDP limited setup, per-core TDP allocations will have to go down. By the time you're exceeding 4 cores running at full tilt, you already have to lower the clocks to match the shared TDP pool pretty close to the point where the compact cores would max out. So if you were to put 12 'un-dense' cores in the same setup, you'd wind up with very much the same clock speeds as more cores drink from the same heat limited power faucet. And AMD is using the dense variants, because at those lower clock speeds they no longer need to spread out those transistors for cooling at top frequencies (yes, I'm ignoring the smaller L3 cache).

That's the whole engineering genius behind that iteration, you effectively don't loose anything, while you potentially gain extra capacity for multi-threaded workloads, because the CMOS performance/clock curve is a knee and not linear.

And while games are actually very different, this design provides the maximum attainable benefits for both, those games that hog a few (often still just one) high frequency core, while ignoring most of the others, as well as those that really spread their load evenly across as many cores that you can provide.

I've seem some of those modern games run equally well even on my 22-core Broadwell Xeon, which has pretty near the same multi-threading computational power as my 8-core Zen 7 5800X, while single core hogs do as badly as the IPC and clock differences between those two architectures suggest.
 
John Nemesh said:
Interesting! Will have to see if it costs an arm and a leg like their last model...and what the battery life is like, but it is certainly intriguing! Personally, I think I would be more interested in the ASUS tablet/laptop with the Strix Halo if it is over $1000.
Click to expand...
They might cost having to train your arm or using your leg to support the big hefty battery they aren't showing in the picture.

At 45 Watts with current battery tech something's got to give: size, play time, or neding an AC cable.

Of course, if you connect it to your power delivery TV, you'd never notice.
 
  • Like
Reactions: John Nemesh
You must log in or register to reply here.

TRENDING THREADS

Latest posts

Moderators online

Share this page

COMPANY
RESOURCES
FOLLOW
Top Bottom