But how will the AI Habana Labs' data center chips fit with existing product lines?
Intel has acquired Israeli AI chip startup Habana Labs for approximately $2 billion.
Habana Labs, whose products are AI training and inference chips for data centers, was one of the first to deliver working AI accelerator silicon and its Goya inference chip performed impressively in the recent round of MLPerf benchmarks. The company is based in Tel Aviv and employs around 150 people in several offices worldwide.
“This acquisition advances our AI strategy, which is to provide customers with solutions to fit every performance need — from the intelligent edge to the data center,” said Navin Shenoy, executive vice president and general manager of the Data Platforms Group at Intel, in a statement. “More specifically, Habana turbo-charges our AI offerings for the data center with a high-performance training processor family and a standards-based programming environment to address evolving AI workloads.”
Habana Labs will remain an independent business unit reporting to Intel’s Data Platforms Group, and will keep its current management team. Habana chairman Avigdor Willenz has agreed to serve as a senior adviser to the business unit as well as to Intel.
Habana vs Nervana
Questions remain about what this acquisition means for Nervana’s product line, which competes directly with Habana’s offering.
Nervana was acquired by Intel back in 2016 for a sum believed to be around $400 million. It launched its own data center training and inference processors, the NNP-T and NNP-I last month.
While two competing chip acquisitions may seem like a strange strategy, there is an awful lot at stake; Intel expects to generate over $3.5 billion in AI-driven revenue in 2019 (up 20% year over year). It therefore cannot afford to back the wrong horse. Backing two horses, or even combining the best horse with the best jockey (likely Habana’s chip with Nervana’s software in this scenario) may be a better bet.
Gaudi and Goya
Habana’s Goya inference chip launched in September 2018 and is commercially available today. It can process 15,000 ResNet-50 images/second with 1.3-ms latency at a batch size of 10 while running at 100 W (more than 5x the number of images than competing platforms).
Habana unveiled its Gaudi training chip in June 2019, which is currently sampling with hyperscale customers. Gaudi can process 1,650 images per second at a batch size of 64 when training a ResNet-50 network. This throughput is delivered at 140 W power consumption.
Both chips have eight VLIW SIMD (very long instruction word, single instruction multiple data) vector processor cores, which Habana calls tensor processor cores (TPC), that are specially designed for AI workloads.
One of the key aspects of Gaudi’s architecture is its on-chip RoCE (Remote direct memory access over Converged Ethernet) network. The chip offers 10 ports of 100 Gigabit Ethernet directly on the processor silicon, a feature Habana said at the time was unique in the world of AI accelerators (competing solutions need extra chips for connectivity; Nvidia previously acquired Mellanox for exactly this technology).
On-chip RoCE makes for easy scalability; huge training systems with dozens of chips can be built using standards-based non-proprietary interfaces. This capability is no doubt attractive to Intel’s data centre customers. Habana’s technology combined with Intel’s access to the hyperscalers, backed by the computing giant’s extensive resources, will likely prove to be a very successful combination.