News Intel and Samsung Display cooperate to advance next-gen AI PCs into 'unchartered territory

[Admin]

Intel and Samsung Display ink MOU to develop screens for AI devices.

Intel and Samsung Display cooperate to advance next-gen AI PCs into 'unchartered territory : Read more

 

[Pierce2623]

Why do we use the term “AI PCs” when any processor with modern extension support can run AI workloads on the CPU? Using such a ridiculous term for a PC with a TINY piece of of fixed function silicon doing matrix math is silly. You can even get usb port matrix accelerators (generally based on the Google Coral) and effectively accelerate smaller LLMs etc as long as you understand how to get it working(which funnily enough also applies to the current integrated TPU/NPU solutions).

 

[usertests]

Sounds like a load of BS.

Using such a ridiculous term for a PC with a TINY piece of of fixed function silicon doing matrix math is silly.

I don't know if anyone has measured it yet, but supposedly a decent chunk of APU die area is going to XDNA2 (50 TOPS). Lisa Su joked about it on stage. And that could be set to increase if they are looking to, for example, double the performance with the next iteration. Newer process nodes will keep the size in check, but a lot of transistors are being devoted to this.

Better find something to do with NPUs, because they are here to stay for at least the next few years.

 

[bit_user]

I see that the source publication also used the phrase "unchartered territory", but that makes no sense. It should be "uncharted territory".

 

[bit_user]

Sounds like a load of BS.

Yeah, I can't think of anything AI-related you could do with a display, other than upscaling. Maybe they have some clever ideas, but I think it's just marketing spin put on their continued partnership.

I don't know if anyone has measured it yet, but supposedly a decent chunk of APU die area is going to XDNA2 (50 TOPS). Lisa Su joked about it on stage. And that could be set to increase if they are looking to, for example, double the performance with the next iteration. Newer process nodes will keep the size in check, but a lot of transistors are being devoted to this.

It's not small, but looks bigger than 4 full-size Zen 5 cores and their L2 cache.

Source: https://www.techpowerup.com/325035/amd-strix-point-silicon-pictured-and-annotated
​

Better find something to do with NPUs, because they are here to stay for at least the next few years.

I've tried, but pretty much everything I can think of is either something you could do with a neural network or something you could at least express in the type of processing graphs they use. It's further complicated by the fact that most of their compute power is decidedly low-precision, low-range. Also, I think they basically require data to be DMA'd in/out of local memory, so you can't really do anything with them that would require much random access.

GPUs are far more flexible for mapping onto general compute workloads. They make heavy use of SMT, as a means of hiding random access latencies, which works well, as long as you're doing something sufficiently parallelizable. They also have excellent support for fp32 and even int32, which are much better suited to doing general computation.

What's interesting about the PS5 Pro is that Sony took the approach of specializing RDNA2's compute units to better handle AI workloads (up to 300 TOPS worth!), instead of bolting on an adjunct NPU. AMD and Sony now have a joint project to better explore such architectures, which seems like it might influence UDNA.

 

[TerryLaze]

Why do we use the term “AI PCs” when any processor with modern extension support can run AI workloads on the CPU?

So that customers can easily tell what they are looking at so they can buy or avoid buying it, the same reason anything is labeled.

 

[Pierce2623]

So that customers can easily tell what they are looking at so they can buy or avoid buying it, the same reason anything is labeled.

But these “AI” PCs don’t offer any services you can’t access without an NPU. Do what makes them AI? So everything with a modern GPU is also an AI PC since they have enough shaders to easily compete with an NPU? What about just running AI models on your CPU? You can still get real time responses from stuff like Llama with no GPU acceleration? Most modern phone chip had an NPU before the AI craze even caught on. I’m afraid you’re confusing marketing with “letting you know what you’re buying”.

 

[bit_user]

But these “AI” PCs don’t offer any services you can’t access without an NPU.

NPUs offer more inferencing horsepower than either CPU or GPU, and better power efficiency as well. The amount of horsepower is important for realtime processing tasks, like image upscaling or other video processing techniques (e.g. background removal for video conferencing). Power-efficiency is important for laptops, so they don't need a bulky, heavy cooling solution and so customers can use these features while on battery.

So everything with a modern GPU is also an AI PC since they have enough shaders to easily compete with an NPU?

The main race for AI PCs is in laptops, which can only accommodate a dGPU at considerable expense, bulk, and power consumption.

 

[Gururu]

Korean news, y'all acting hot and bothered.

 

[Pierce2623]

NPUs offer more inferencing horsepower than either CPU or GPU, and better power efficiency as well. The amount of horsepower is important for realtime processing tasks, like image upscaling or other video processing techniques (e.g. background removal for video conferencing). Power-efficiency is important for laptops, so they don't need a bulky, heavy cooling solution and so customers can use these features while on battery.


The main race for AI PCs is in laptops, which can only accommodate a dGPU at considerable expense, bulk, and power consumption.

My point was a modern iGPU offers more inferencing power than an NPU. Look at Lunar Lake. The GPU offers more “AI TOPS” (which is literally just Tflops at a lower precision format) than the NPU and AMD’s would too if they rated it for “AI TOPS”.

 

[bit_user]

My point was a modern iGPU offers more inferencing power than an NPU. Look at Lunar Lake. The GPU offers more “AI TOPS” (which is literally just Tflops at a lower precision format) than the NPU and AMD’s would too if they rated it for “AI TOPS”.

Okay, if we take the example of Lunar Lake, because Intel went so big on its iGPU, it does actually overtake the NPU on raw TOPS (60 vs. 48) [1]. However, the iGPU apparently consumes about 50% more area [2] and I'm sure it burns much more power at 60 TOPS than the NPU does. In Meteor Lake, Intel had some slides showing their relative power efficiency, but I'm not seeing equivalent slides for Lunar Lake.

It's power-efficiency that's the main value-add of these NPUs, and it's why AMD said they're not in a hurry to include them in desktop CPUs.[3]

Sources:

1. https://www.tomshardware.com/pc-com...pc-gain-for-e-cores-16-ipc-gain-for-p-cores/5
2. https://www.tomshardware.com/pc-com...pc-gain-for-e-cores-16-ipc-gain-for-p-cores/5
3. I read this in an interview with an AMD exec. Don't remember where, but maybe I can find it if you don't believe me.

One thing you could do with a dedicated NPU is almost completely offloading AI-upscaling from the GPU.