Today Intel announced that it is acquiring Habana Labs at a $2 billion valuation. First off, a piece that many have missed today is that Intel Capital was an investor in Habana Labs so it already owned a portion of the company. Still, this is a big deal given that Intel has been pushing its competitive NNP-T heavily in recent weeks.
Intel Acquires Habana Labs Stoking its AI Efforts
If you want a deep-dive into the technical aspects of Habana Labs, you can see our Favored at Facebook Habana Labs Eyes AI Training and Inferencing. When we wrote up the Hot Chips 31 technical presentation, we knew that Habana Labs was hot at Facebook. Facebook, one may recall was also a big test partner for the NNP-T. Habana Labs Gaudi is designed not just with pure compute in mind. Instead, it is designed to scale out on 100GbE RDMA fabrics. Hyper-scalers use Ethernet heavily, so Habana Labs’ approach to using Ethernet means that they are designed to scale to large numbers of chips on the common fabric in the world’s largest customers.
What is interesting here, is that we now have a fairly good idea of where the AI training market is going. Intel pushed to increase its data center silicon TAM by acquiring Barefoot Networks for its Ethernet switch silicon. This is the same silicon that can one day interconnect Gaudi servers with 10x 100GbE ports each. A typical top of rack switch has 32x 100GbE ports, so one can see how network topologies will change.
For its part, NVIDIA already has signaled that it is moving heavily toward integrating fabric with its GPU training solutions. While there is NVIDIA NVSwitch inside a chassis, for scale-out NVIDIA acquired Mellanox. In that piece, we highlighted that the Mellanox BlueField SoC was essentially what one would need to get the company’s GPUs directly onto fabrics.
Intel had been pushing the OAM NNP-T solution as recently as Intel AI Summit 2019 and we highlighted it being shown a few weeks ago in the SC19 booths of Supermicro and Inspur among others. Now, one has to question the fate of the NNP-T. Habana Labs was a direct competitor at Facebook for the kind of scale-out training that the NNP-T was designed for. Now, Intel is buying Habana Labs. It is also notable that Habana Labs chips are not manufactured in Intel fabs so this helps increase silicon TAM without putting more burden on Intel manufacturing.
While $2 billion may seem like a lot, if Intel turns this acquisition into the go-to NVIDIA alternative, then it is a bargain. Although the acquisition also signals that the NNP-T seems to have lost the AI training war, Intel is getting a second shot with Habana Labs chips.