Intel Axes Nervana Just Two Months After Launch

Recently, Intel also announced 3rd generation Movidius VPUs, codename Keem Bay. To that end, Intel has confirmed that there are no changes to the Movidius roadmap for vision processing.The official statement from Intel is published below.After acquiring Habana Labs in December and with input from our customers, we are making strategic updates to the data center AI acceleration roadmap. As part of this update we plan to deliver on current customer commitments for the Intel NNP-I inference accelerator (code-named “Spring Hill”) and cease development of the Intel NNP-T (code-named “Spring Crest”).This roadmap decision aligns to Intel’s AI Strategy and our commitment to deliver heterogenous AI solutions that fit our customers’ evolving power and performance needs – from the intelligent edge to the data center.The discontinuation of Nervana and the adaptation of Habana is a puzzling move for us. Habana did not disclose the size of the local memory but on a 16 nm 1-1.5 MiB isn’t unreasonable.The other part of the compute in Gaudi is the General Matrix Multiply (GEMM) engine. Habana did not disclose the size of this SRAM but we can speculate it’s a sizable amount, probably around 32-48 MiB of memory.Nervana’s design, on the other hand, has a more complex design. It’s hard to see why Intel chose Habana over Nervana based on technical merit. While Intel’s NNP-I, codename Spring Hill, (the inference version of NNP-T) is, in fact, an entirely different architecture to Spring Crest, it largely builds on Intel’s own mobile client SoC. That’s essentially the claim Intel is making with the switch to Habana.Beyond hardware, it’s also entirely possible that the problem boiled down to software. Reports on NNP-I were very positive for a low power chip.The two leading theories continue to be:1) Intel went with Habana because a number of large customers favored them, or2) The hardware complexity of Nervana made software development too difficult