Thanks to the event Hot Chips 31, the week is very busy for the Intel. In addition to officially announcing the 10th generation Comet Lake processors and giving more details about Lakefield, the company introduced the family chips Nervana Neural Network Processor (NNP), aimed at artificial intelligence.
For now, there are two chips: the NNP-I 1000 (code-named Spring Hill) and the NNP-T (Spring Crest). Both were designed to accelerate systems based on artificial intelligence, but each takes a different approach.
Let’s start with NNP-I. This chip was developed with a focus on inference (hence the ‘I’ in the name), which means it should optimally perform tasks from already trained neural networks.
The NNP-T, on the other hand, focuses on training, that is, the job of teaching the neural network to process data according to certain sets of rules and conditions.
Often, training neural networks requires great computational capacity, so the NNP-T was designed to be highly scalable. Several units of the chip can be placed to work together with relative ease.
The NNP-T is based on a 16 nanometer chip from TSMC (as it is). It can handle up to 32 GB of HBM2 memory (four modules of 8 GB) and comes with 24 TPCs (Tensor Processing Clusters), in addition to 60 MB of cache (if adding the caches of all its cores).
The NNP-I has simpler specifications. It is based on a 10 nanometer Ice Lake processor: here, the GPU and two CPU cores have been removed to make room for 12 ICE (Inference Compute Engine) accelerators that, as the name implies, are targeted at inference tasks.
Completing the specifications of the NNP-I are two Ice Lake cores (of Sunny Cove architecture) for artificial intelligence that are accompanied by 24 MB of L3 cache. Both deal with AVX-512 VNNI instructions, which are widely used in neural networks.
The logic here is relatively simple: NNP-I handles simple inference tasks and therefore requires less energy – TDP varies between 10 and 50 W here; the NNP-T, on the other hand, must take action in the heaviest tasks, which involve deep learning, for example. This explains the fact that its TDP varies between 150 and 250 W.
Samples of the Nervana chips have already been sent to companies like Baidu and Facebook. Large-scale production is expected to begin by the end of the year, at least for NNP-I chips.
It is worth mentioning that this project was possible because Intel bought Nervana Systems in 2016. The chips were designed at the Intel unit in Haifa, Israel.