Silicon Valley’s venture capital firms had largely pulled out of the semiconductor market before Google introduced the custom tensor processing unit in 2016. The announcement sparked billions of dollars of investment into startups building chips that could be installed in data centers to handle machine learning algorithms or embedded in Internet of Things devices like security cameras and drones.
The venture capital industry opened its checkbook again last week with the startup Habana Labs raising $75 million led by Intel Capital. Silicon Valley has started acknowledging the argument that the main computer processors powering data centers and other systems today are insufficient for training and running algorithms that, for instance, can tell the difference between a photograph of a black cat and one of a crow.
The first product based on Habana’s Goya architecture is targeting machine learning inference. That means running algorithms that have been trained to detect malfunctions in factory equipment before they happen or allow autonomous cars to tell the difference between street signs. The chip’s die is composed of tensor processing cores that each have local memory and access to shared memory, the company said.
The company said the Goya processor supports lower latency, higher throughput and lower power than alternatives from Intel and Nvidia, which are currently leading the inferencing and training markets, respectively. The company said that it can beat Nvidia's Tesla V100 with throughput of 15,012 images per second, latency of 1.3 milliseconds, and the ability to process 150 images per second per watt using the ResNet-50 algorithm.
Thursday was the company’s second round of the funding, bringing its total amount raised to $120 million. It pushes Habana into the club of machine learning hardware startups that have raised in excess of $100 million of venture capital. The funding was rounded out by Bessemer Venture Partners, Battery Ventures and WRV Capital—an investment firm founded by Cadence Design Systems chief executive Lip Bu Tan.
But unlike many of the other companies taking advantage of Silicon Valley’s shifting views on semiconductors—including Graphcore, Cerebras Systems, and Wave Computing—Habana said that it is ready to start production. The company is currently shipping the Goya processor to customers, while Habana’s 16-nanometer Gaudi training processor is expected to become available in samples by the second quarter of 2019.
The company, which is also building the accompanying software tools, currently employs around 120 people globally. “The funding will be used to execute on our product roadmap for inference and training solutions, including our next generation 7-nanometer artificial intelligence processor [and] to scale our sales and customer support teams,” David Dahan, Habana’s chief executive, said in a statement.