Analytics Chip Startup Speedata Emerges from Stealth
There is no shortage of chip startups hoping to claim a piece of the emerging AI pie. But instead of accelerating machine learning workloads, Speedata is looking to speed up the performance of SQL analytics running in column-oriented databases, which the Israeli company says is bigger opportunity than AI.
Speedata is developing an Analytics Processing Unit (APU) that is designed specifically to address the bottlenecks that are common in traditional analytics. The APU itself will plug into a PCIe slot on an X86 server, and give companies a 10x to 20x speedup, depending on the workload, the company claims.
Like a host of emerging AI-specific chips, the Speedata APU is not based on a von Neumann architecture, which is the classic architecture that is traditionally used by CPUs and GPUs, says Speedata CEO and co-founder Jonathan Friedmann. However, that’s where the similiaries end.
“In AI, you would optimize for floating point multiplications, while in analytics and databases you would optimize for strings, which have their own challenges,” Friedmann tells Datanami.
According to Friedmann, Speedata’s APU (which you can see in the feature image at the top of this story) targets the three main bottlenecks in analytics and databases: I/O acceleration, compute acceleration, and memory. It does this both in terms of bandwidth and capacity, he says.
“We take a look at important functions within SQL that are very popular and very hard for CPUs to process, and we actually build specific hardware pipelines for it,” he says. “The whole architecture is so different from a CPU that actually everything is being done pipelined within the chip itself, so there is much less going out to the DRAM and getting back. So there’s a huge benefit in terms of both DRAM capacity and DRAM bandwidth which is needed.”
Friedmann declined to provide specifics about the chip design, such as the amount of memory or gate-width. Suffice it to say, a prototype is already running in cycle-accurate simulators and on a FPGA in the Speedata lab. The chip will be manufactured by TSMC, and should be in customers’ hands within a few quarters, Friedmann says.
When the APU is available, customers will be able to plug it into their existing X86 architecture to speed up analytics workloads, much as GPUs can be plugged in to augment the compute-intensive workloads inherent in high-end graphics, machine learning, and modeling and simulation, according to Friedmann.
Speedata is currently working with several Fortune 500 companies to determine how the chip design will function with their SQL workloads. The company’s target is data that’s stored in a columnar data format within a column-oriented relational database.
The company has shown the low-end chip will work with Apache Spark and its SQL processing engine, Spark SQL, and work is underway to support other processing engines, including Presto and PostgreSQL, Friedmann says.
The potential performance gains of having a dedicated analytics engine are substantial, Friedman says. “We have shown a multiple of anywhere from an order of magnitude going all the way to two orders of magnitude improvements,” he says. SQL analytic functions like joins and aggregations, among others, will benefit from having hardware specifically designed for it, according to Friedmann.
The exact improvement that each customer will see depends on the workload and the data, he adds. The company has run a chip prototype against the TPC-DS benchmark, a common benchmark used to determine the performance of analytics database. However, results were not available.
While speeding up machine learning workloads has attracted lots of attention from established and startup chipmakers, the analytics side of the equation, i.e. running good old fashioned SQL queries, is an overlooked opportunity, Friedmann says.
“The databases and analytics are a workload which is actually bigger in terms of dollars spent in the data center,” he says. “While Nvidia and Intel are looking into it…still not many startups are doing that and the big companies are only making the first steps. So we feel like we have a nice two year advantage here.”
Speedata, which is based in Netanya, Israel, today announced a $55 million Series A round of venture capital funding, giving the company $70 million in total funding (there was a previously undisclosed $15 million round). The Series A was led by Walden Catalyst Ventures, 83North, and Koch Disruptive Technologies (KDT), with participation from existing investors Pitango First, Viola Ventures as well as Eyal Waldman, the co-founder and former CEO of Mellanox Technologies.
Speed data was founded by a multi-disciplinary group of men, including Friedmann, Dan Charash, Rafi Shalom, Itai Incze, Yoav Etsion, and Dani Voitsechov. Friedman and Charash both built chips at Provigent, which was acquired by Broadcom for $360 million in 2011, while Etsion and Voitsechov spent six years researching the current APU design. Rafi Shalon brings networking expertise, while Itai Incze is its chief software architect.
“In order to solve [processor challenges] you need the experts in multi disciplines and as such the founding team is like that,” Friedmann says.