By Meredith W. Salisbury
Simchon Faigler was one of the founders of Compugen and a member of the team that developed the company’s hardware accelerator — but if you ask him, he will probably recommend that you don’t buy it.
Faigler, who nearly three years ago purchased the accelerator technology from Compugen and used it as the basis for a new company called Biocceleration, says that these days, he and his colleagues have stopped advertising the hardware accelerator and may in fact actively advise potential customers against buying it.
That may seem like a strange marketing technique, but it makes sense to Faigler. In the time since he helped build the hardware accelerator, he says, several changes in computer technology have made that tool overpriced and less effective. His team at Biocceleration has spent the past few years designing a software-based accelerator that he says gives the same performance as a hardware accelerator at a fourth of the price.
The software, released last fall and known as GenCore 6.0, is in beta testing at a few academic sites and was licensed late last year by the United States Patent and Trademark Office. USPTO is “one of the largest sites from the point of view of high-throughput sequence searches,” Faigler says, because of all the prior art records patent examiners have to query for each new patent application. The agency has also been his customer since 1998, back when Compugen sold the hardware accelerator.
The accelerator tool speeds up computational power for search algorithms such as Smith-Waterman, HMM, and GeneWise, Faigler says. Optimized for a single processor, the tool shows 10-fold speedups; that drops to three-fold for dual-processor machines, but Faigler says that “speed improvement is linear with the number of nodes in the cluster — up to the point where we get I/O bottlenecks.” It’s no FPGA acceleration, but for a tool that’s free for academics, that may not be a dealbreaker. In general, Faigler adds, “if the query length is relatively long and the job is relatively heavy, there’s a nice improvement.”
Prices and processors
Faigler says the writing was on the wall years ago for the fate of hardware accelerators. “The main issue was that the general-purpose processors were becoming more and more powerful and less and less expensive,” he says. Many universities or organizations have Linux clusters that can cost less than $100,000 — the price for what Faigler calls “an entry-level” hardware accelerator. “We felt that the market was ready and the technology was ready for a shift in the way sequence searches could be accelerated,” he adds.
GenCore relies on programming essentially at the processor level to get speedups out of everything from single computers to large, multiserver clusters. It uses the MPI interface to support sending single jobs to multiple nodes, and takes advantage of the hyperthreading technology in Intel chips to help ramp up speed.
Currently, Faigler’s crew is compiling a version of GenCore for AMD processors. “We thought that the market would be largely dominated by Intel-based processors,” he says, “but we’ve noticed in the last few months that the AMD-based servers occupy a significant percentage of the market.” The GenCore technology does not support the Apple system, he says.
Biocceleration offers a free license for the tool to scientists at academic and research institutions. Faigler says that move was “a strategic decision” — it gets the technology in the hands of people who may not otherwise have the budget to buy it, and he believes the product serves as its own best promotion. As people move from academia to industry, they’re already familiar with the technology, giving Biocceleration a ready entrée into the commercial sector.
For those already in the private sector, GenCore will run several thousands of dollars. The base price is between $10,000 and $20,000, and there’s a per-server price that decreases with more and more servers, Faigler says. Hardware accelerators, by comparison, start around $100,000 and can run up to $500,000 for top-of-the-line versions. “From the benchmark that we did,” Faigler says, “the price-to-performance is around four-fold. For a hardware accelerator that costs $100,000, you can buy a software accelerator for $25,000 and you get the same performance.”