5 ms·
Serious question: what is an AI coprocessor technically? Some machine learnt models burned on a chip? Or some kind of a neural net with updatable weights?
by ratbr 9y ago
Serious question: what is an AI coprocessor technically? Some machine learnt models burned on a chip? Or some kind of a neural net with updatable weights?
- arcanus 9y agoIt is not well documented by anyone. However, the expectation is that it is a matrix or convolution coprocessor, as this is a common operation in deep neural networks (for both inference and training). For instance, NVIDIA says they are supporting 4x4 convolutions with the tensor unit.
- dgacmu 9y agoOne example is: https://arxiv.org/abs/1704.04760 https://arxiv.org/abs/1704.04760 There are many potential designs for these things, but the first gen TPU is one that works, is in production, and has been described in a paper. But you have to differentiate if you mean an inference engine, or something that can also do training. For HoloLens, it's probably going to be an inference unit, which means it'll possibly look something like a TPU, perhaps with more specific hardware support optimized for convolutions (which are very important for visual processing DNNs these days), as the NVidia tensor units are.
- Eridrus 9y agoProbably also enough high speed memory to store the weights without needing to go to RAM.
- chriskanan 9y agoI was in the audience at CVPR when it was presented. They were doing semantic segmentation using resnet-18, so I'm guessing it speeds up convolutions and some linear algebra during inference. I'm guessing it won't be used for training.
- danmaz74 9y agoThe AI coprocessor is probably the first processor designed directly by the marketing department...
- gumby 9y agoOh man, that ship has sailed!
- WorldMaker 9y agoAccording to the linked article, this coprocessor seems particularly focused on Deep Neural Networks (DNN), so it does sound like a updatable weight neural network evaluator.
- hatsunearu 9y agoA whole butt ton of GPU-style FMA and low precision float multiply ALUs would be my guess
- protomyth 9y agoHow low precision can you get and still have it be useful?
- fulafel 9y agoGoogle's TPU proves that 8 bits is still good.
- sbierwagen 9y agoNote that the TPU used 8 bit integer math, not even floating point.
- fulafel 9y agoGoogle's TPU proves that 8 bits is still good.
- Govindae 9y ago1-bit, binary neural networks work.