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

[John Nemesh]

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.

Well, the main issue I have for Strix Halo is the price...it really just doesn't make sense for gaming...especially PORTABLE gaming, given it's power requirements.

I am FAR more likely to build a desktop PC, with a 9070 or 9070XT, that will be much more powerful than the Strix Halo (and include FSR4!) and similarly priced to the Strix Halo boxes I have seen being sold. I really WANT an APU that will be just as powerful as a 9070XT with a 9800X3D, but I may have to wait a couple more gens for that.

 

[abufrejoval]

Well, the main issue I have for Strix Halo is the price...it really just doesn't make sense for gaming...especially PORTABLE gaming, given it's power requirements.

I agree, the price must match both its capabilities and the actual cost.

I really WANT an APU that will be just as powerful as a 9070XT with a 9800X3D, but I may have to wait a couple more gens for that.

That may be impossible, or unrealistic, or just not worth the money.

The base issue is that CPUs and GPUs have very diverging memory requirements.

CPUs are really designed for workloads with strong code and data locality: that's why they use more surface area on caches than anything else, while desktop and mobile designs stick to 128 bit data bus width (dual DDR 64-bit or quad LP-DDR 32-bit). The cost of going wider won't pay in better performance until you go to EPYC core counts.

GPUs have far lower data locality and need tons of bandwidth, especially for AI. So going wide here helps, but while some newer chips like the Halo will now do 256 bit, with DRAM that's still below what 128 bit of GDDR6 on an RTX 5050 will deliver.

High-end consumer GPUs need around 1TByte/s and that means 1024 bit of DDR5 or 16 channels, which is quite a few of pins, traces, sockets etc... beyond what even the biggest EPYC mainboards do just yet.

Apple's Ultras do 512 bit, but they keep that to a die carrier stacking the DRAM there so it's expensive on the SoC packaging and not very flexible, but it keeps the PCB to something pretty near a NUC.

But they still fall behind consumer dGPUs on bandwidth.

I can't really see those two converging because their demands remain different.

The main change needs to come from RAM that offers far more bandwidth but with DDR economy.
While GDDR prices are coming down, they may never come close enough and then they add latencies to CPU workloads, ...which might be compensated with bigger V-caches.

Some consoles still just go with VRAM for everything, but CPU benchmarks won't be pretty and the PC fan base might be outraged.

I haven't heard of anyone working at an intermediate RAM type, most likely since that would require redesigns for near everything.

If any of this were easy, they'd have done it, because it's pinching everyone in the same spots.

 

[John Nemesh]

I agree, the price must match both its capabilities and the actual cost.

That may be impossible, or unrealistic, or just not worth the money.

The base issue is that CPUs and GPUs have very diverging memory requirements.

CPUs are really designed for workloads with strong code and data locality: that's why they use more surface area on caches than anything else, while desktop and mobile designs stick to 128 bit data bus width (dual DDR 64-bit or quad LP-DDR 32-bit). The cost of going wider won't pay in better performance until you go to EPYC core counts.

GPUs have far lower data locality and need tons of bandwidth, especially for AI. So going wide here helps, but while some newer chips like the Halo will now do 256 bit, with DRAM that's still below what 128 bit of GDDR6 on an RTX 5050 will deliver.

High-end consumer GPUs need around 1TByte/s and that means 1024 bit of DDR5 or 16 channels, which is quite a few of pins, traces, sockets etc... beyond what even the biggest EPYC mainboards do just yet.

Apple's Ultras do 512 bit, but they keep that to a die carrier stacking the DRAM there so it's expensive on the SoC packaging and not very flexible, but it keeps the PCB to something pretty near a NUC.

But they still fall behind consumer dGPUs on bandwidth.

I can't really see those two converging because their demands remain different.

The main change needs to come from RAM that offers far more bandwidth but with DDR economy.
While GDDR prices are coming down, they may never come close enough and then they add latencies to CPU workloads, ...which might be compensated with bigger V-caches.

Some consoles still just go with VRAM for everything, but CPU benchmarks won't be pretty and the PC fan base might be outraged.

I haven't heard of anyone working at an intermediate RAM type, most likely since that would require redesigns for near everything.

If any of this were easy, they'd have done it, because it's pinching everyone in the same spots.

Not saying its EASY, but it CAN be done. A lot of the bandwidth requirements go bye bye when you have huge amounts of cache and unified memory...just sayin'