Chip giant Intel has splashed out “approximately” $2 billion (£1.5 billion) to purchase Israel-based Habana Labs: a programmable deep learning accelerator specialist that recently whipped Intel’s own processors during industry-backed AI benchmark tests.
Habana Labs was founded in Tel Aviv in 2016, and currently has a staff of roughly 150 people located in offices worldwide. The firm is best known for its Goya inference card and Gaudi AI training processor, used in data centres running AI-intensive workloads.
Navin Shenoy, VP and GM of data platforms group at Intel said the acquisition “turbo-charges” Intel’s AI offerings for the data centre, amid growing pressure from cloud providers which are increasingly developing their own hardware for AI workloads.
Intel expects the AI silicon market to be worth over £18 billion by 2024, while the Santa Clara firm estimates that its own AI-driven revenue will clock in at £2.6 billion for the year 2019, a projected rise of 20 percent year-on-year.
Habana Labs Beats Intel During MLPerf
Goya AI Inference Card
One of Habana Labs’ main products is the Goya HL-1000 inference processor; a card designed specifically for neural network processing. The card incorporates a fully programmable Tensor Processing Core, alongside a SynapseAI software stack.
Habana Labs purpose-built Goya inference processor did extremely well in industry-backed AI and ML standardised MLPerf tests (a performance benchmark scoring platform).
Habana Labs HL-102-Goya PCI-board – tested in the MLPerf test division of stream ResNet-50 image classification – completed the set task in 0.24 seconds while Intel’s Xeon Platinum 9200 processors took 1.37.
Content from our partners
In another directed comparison – during the multi-stream image test – Habana’s Goya clocked a result of 700.00, while Intel’s 9200 processors took double that time as it crossed the line at 1,920.00.
Intel notes the Lab’s Gaudi AI Training Processor is currently “sampling with select hyperscale customers. Large-node training systems based on Gaudi are expected to deliver up to a 4x increase in throughput versus systems built with the equivalent number of GPUs. Gaudi is designed for efficient and flexible system scale-up and scale-out.”
Intel says that Habana Labs will remain an independent business unit within Intel’s corporate structure and retain its current management team.
Goya HL-1000 Performance on ResNet-50
