Report: Intel in Talks to Acquire AI Company Habana Labs for Over $1 Billion

(Image credit score: Shutterstock)

Intel may well be in talks to procure deep studying chip start-up Habana Labs for a ticket of over $1 billion, in step with a file Tuesday. It can be Intel’s 3rd main synthetic intelligence (AI) start-up acquisition after Movidius and Nervana in 2020.

The file from CTech, the era outlet of Israeli newsletter Calcalist, cited an annoymous supply “acquainted with the topic.” Intel is in “complex negotiations”, in step with the individual, and the possible deal is valued at between $1 and $2 billion greenbacks. This would make it Intel’s biggest acquisition for the reason that $15.3 billion deal to procure Israel-based Mobileye in 2020.

Neither Intel nor Habana Labs replied to the rumor.

Habana Labs, one of many AI start-ups, used to be based in 2020 by means of two former executives of 3-D sensing corporate PrimeSense, which used to be bought by means of Apple in 2020 for $360 million. Its objective is to create processors for deep studying coaching and inference with upper efficiency, scalability price and gear “from the ground-up.” The corporate has raised a complete of $120 million and employs 120 other people. The closing making an investment spherical in November 2020 gave it $75 million and used to be led by means of Intel Capital. At the time, Habana CEO mentioned Intel had no plans for running along with Habana.

If the deal is going via, it might be a noteworthy addition to Intel’s portfolio. For the closing a number of years, Intel has been making an investment in AI features from edge to information heart and CPU to ASIC (application particular built-in circuits), in addition to instrument enablement. 

In 2020, Intel purchased Movidius for edge and imaginative and prescient AI, and Movidius just lately introduced the Keem Bay VPU (imaginative and prescient processing unit). Also in 2020, Intel bought Nervana (reportedly for $350 million) to go into the deep studying coaching chips marketplace. Intel additionally advanced the NNP-I for inference on its own. Hence, a possible acquisition of Habana Labs may – unusually – point out that Intel nonetheless sees gaps in its AI portfolio.

Gaudi: Habana’s Deep Learning Training Chip

Image 1 of 2

(Image credit score: Habana Labs)

Image 2 of 2

(Image credit score: Habana Labs)

Habana’s coaching chip is named Gaudi (PDF). Similar to Nervana, its structure is in keeping with Tensor Processor Cores (TPCs), it has 32GB of HBM2 at 1 TBps and helps PCIe 4.0. It additionally helps common DL numeric codecs, reminiscent of INT8 and BF16.

A standout function of Gaudi is its interconnect for scaling out to masses of chips. It is the primary AI chip with engines for Remote Direct Memory Access (RDMA) over Converged Ethernet with a bi-directional throughput of 2 Tbps (by means of as much as 10 100 Gbps Ethernet ports). RDMA signifies that a chip can get entry to some other chip’s reminiscence immediately (with out involving the running gadget). The good thing about Ethernet is that switches are readily to be had, and there’s no want for PCIe switches or devoted NICs. 

In phrases of efficiency, one 140W TDP chip has a throughput of 1,650 photographs according to 2d in ResNet-50, 14% upper than Nvidia’s Tesla V100 at part the ability intake. This leads to an potency of over two times the choice of photographs according to 2d according to watt. The distinction turns into even higher because the choice of chips scales. Based on MLPerf v0.5, 640 chips reach a throughput of 845,400,000 photographs according to 2d, while the V100-based gadget with 80 DGX-1s reached 218.3 thousand photographs according to 2d. A 512-chip Gaudi gadget has a relative potency of 80% in comparison to underneath 30% for a V100 gadget.

Goya: Deep Learning Inference Chip

Image 1 of 3

(Image credit score: Habana Labs)

Image 2 of 3

(Image credit score: Habana Labs)

Image 3 of 3

(Image credit score: Habana Labs)

Habana additionally claims to have the primary commercially to be had deep studying inference processor with its HL-1000 Goya (PDF) chip. Exact main points are scarce, nevertheless it options totally programmable VLIW TPCs and helps all main frameworks.

Habana claims it has as much as three occasions upper efficiency and two occasions the potency of Nvidia’s Tesla T4 whilst additionally having a lot decrease latency.