/tech/ - Technology

>>31577
it is hard to tell without more information; computer vision for example is considered AI but uses different hardware (conventional GPUs) than chatbots. or maybe it is a GPU with additional hardware for AI features, or whatever. your CPU has a lot of stuff in it besides the ALU and the decoder, and yet we still call it a CPU

the real discussion is that huawei should focus on euv development and stop trying to cope with duv designs

>they should call them AIPUs and market them as that.
They are though??? It's right there in your picture. Huawei called it an inference card.

Have heard the term LPU, language processing unit from Groq.
And NPU used by Microsoft.
Seems there's just not a standard yet.
Language seems over restrictive, NPU sounds closer.

>>31597
AIPU too clunky to say, IPU/inference card or NPU will prolly win in the long term

Sorry kiddo those babies pump out 4 bit calculations at Petaflops, if you want to use it for vidya start looking at making your own motherboard inspired by the Sony Playstation 2

https://venturebeat.com/ai/nvidia-researchers-unlock-4-bit-llm-training-that-matches-8-bit-performance

>>31599
APU was unfortunately taken by AMD's marketing team pretending that an integrated GPU was novel or useful in a new way.

My money is on NPU.

GPU but the G stands for Gay

Machine learning specific processors have been around for a long while. Tensor processing units. Mostly good for regression, running cnns, rnns.

>>31631
yeah intel went really hard with their neural sticks, google had their tensor cores, apple had their own name for it too, mostly for computer vision, when AI was still cool in the eyes of consumers.

Excellent news kiddo
https://www.tomshardware.com/tech-industry/quantum-computing/new-chinese-optical-quantum-chip-allegedly-1-000x-faster-than-nvidia-gpus-for-processing-ai-workloads-but-yields-are-low
Prices on those babies should now eventually come down in a short time, around two years. You can get what is essentually a devkit for your forthcoming vidya console from orangepi starting at, too much for a mass console … for now

Eventually a lowmem board won't cost too much at which point have fun

>>31727
ps. Softbank just dumped NVDIA
Those who know know

Office hours I guess, any questions?