The market for hardware-based accelerated bioinformatics will get a bit more crowded this summer with the release of new FPGA-based products from Mitrionics and CLC Bio.
Mitrionics, a provider of general-purpose programming tools for field-programmable gate arrays, announced this week that it is developing a suite of bioinformatics applications, including Smith-Waterman and Blast, that will run on the company's Mitrion Virtual Processor technology that sits between the hardware and an application so that the FPGA can be programmed without requiring knowledge of circuit design.
Mitrionics is developing the bioinformatics toolkit in collaboration with the Center for Biological Sequence Analysis at the Technical University of Denmark, a customer that it shares with SGI. The initial version will run on SGI's RASC (Reconfigurable Application-Specific Computing) RC200 blade, an FPGA-based system designed for SGI Altix servers that the company launched in September.
Anders Dellson, CEO of Mitrionics, told BioInform that the company is still testing its FPGA-ready version of Blast, but said he expects it to provide a 10-fold to 100-fold speedup over traditional processors.
CLC Bio, meanwhile, is taking a different approach to the same technology. The Aarhus, Denmark-based startup was founded just over a year ago to commercialize its Workbench line of bioinformatics software, but decided shortly afterwards to move into the accelerated hardware market as well.
"I think the integration of the general-purpose processor with the FPGA is going to fuel a lot more applications."
This summer, the company will launch the so-called CLC Bioinformatics Cube, an FPGA-enabled 5-cubic-inch box that plugs into a standard PC via a USB cable. The Cube, which the company developed in collaboration with an undisclosed partner, is integrated with CLC Bio's software so that users can use a drag-and-drop interface to launch Blast jobs directly from their desktops to run on the Cube with a 25-fold speedup.
Thomas Knudsen, CEO of CLC Bio, said that users can also run Blast, Smith-Waterman, and other algorithms on the Cube via a command-line interface.
Mitrion's version of Blast will be available "in a month or two," with Smith-Waterman available soon afterwards, Dellson said. The company is also mulling other bioinformatics algorithms, such as hidden Markov-based approaches, but has not finalized its development roadmap.
CLC Bio plans to launch its Smith-Waterman implementation in July, with Blast to follow in August or September.
Neither company provided pricing, but CLC Bio's Knudsen said that the company's Cube should come in below the $10,000-$15,000 range.
From Black Box to Commodity?
The companies are following in the footsteps of TimeLogic, Paracel, and Compugen, which more than a decade ago pioneered the use of reconfigurable hardware in order to accelerate compute-intensive bioinformatics algorithms.
But unlike first-generation bioinformatics accelerators dedicated "black-box" appliances that cost tens of thousands of dollars and could only run a predefined set of algorithms this new crop of FPGA-based systems promises a new level of flexibility.
CLC Bio's Knudsen said that the company's upcoming Cube will be small enough for researchers to move from computer to computer in the lab, or even carry home and use on their computers overnight or over the weekend a capability not possible with a fridge-sized DeCypher, for example.
Mitrionics' Dellson, meanwhile, noted that his firm expects to benefit from the increasing use of FPGAs in standard supercomputing systems. In addition to SGI, Cray has also launched a version of its XD1 supercomputer that combines FPGAs with standard processors, and Dellson said that he is aware of at least two other hardware vendors that are planning on launching FPGA-enabled supercomputers this year, although he declined to disclose the identity of those firms, citing nondisclosure agreements.
"Blast, and all genomics, lends itself quite well to FPGAs and lends itself so badly to general-purpose CPUs."
"These FPGAs are no longer black-box accelerators," Dellson said. "When you use the Mitrion platform with these FPGAs, they become programmable, just like the rest of the computer. So that's the big difference. It's no longer special-purpose hardware. It's actually a very, very fast processor."
Bill Mannel, director of systems marketing at SGI, acknowledged that "FPGAs have been in bioinformatics for awhile," but added that a number of factors have prevented them from finding widespread adoption in the field. The primary drawback of systems from Paracel and Time Logic, he said, is that there was "no real integration of the FPGA capability into a general-purpose server offering you bought the thing and that's all you could use it for."
Mannel said that SGI currently treats FPGAs as "just another compute paradigm within a solid general-purpose platform" a combination that "fits nicely into someone running a data center," he said.
Others agree that this trend toward integration with general-purpose systems could help drive adoption of FPGA-based systems in bioinformatics. "I think the integration of the general-purpose processor with the FPGA is going to fuel a lot more applications," said Jack Collins, manager of scientific computation and program development at the National Cancer Institute's Advanced Biomedical Computing Center. "The close coupling like you see in the Cray systems, like you're going to be seeing with the other companies coming out that's definitely going to fuel the usage."
In 2004, the ABCC licensed an FPGA-based system from Starbridge Systems, which included a development package that was designed to allow easier programming of the chips [BioInform 04-12-04]. While Collins described his experience with that system as "not a walk in the park," he said that he's pleased with what he's seen from other firms, like SRC Computers, that have developed technology that allows "a competent programmer [to] get some reasonable speedup from FPGAs for certain algorithms, without having to understand the hardware underneath it." He said that ABCC plans to acquire an SGI RASC system running the Mitrionics platform in the near future.
Although the Mitrionics bioinformatics toolkit is initially designed to run on the SGI RASC system, Dellson said that it will eventually run on all FPGA architectures that the company supports, including Cray's.
Cray announced in November that it was developing its own version of Smith-Waterman to run on the FPGA-enabled XD1 [BioInform 11-14-05]. At the time, the company said it had several beta customers for the accelerated Smith-Waterman application, but company officials could not be reached for an update on the status of the product.
Amar Shan, XD1 product manager, said at the time that the Smith-Waterman implementation served as a proof-of-principle for a suite of FPGA-enabled applications that Cray planned to release for the XD1.
Likewise, the bioinformatics suite that Mitrionics is developing will be its first "turnkey" software package for FPGA-based systems.
"Blast is one of the first applications where we, together with a customer, go all the way and provide a package with the source code to all our customers that are interested," Dellson said. "But that's quite natural because Blast, and all genomics, lends itself quite well to FPGAs and lends itself so badly to general-purpose CPUs."
The majority of Mitrionics' customers are in oil and gas, finance, and national security, Dellson said, where there isn't a "clear equivalent" to Blast. "Each customer has more specialized demands for what they want" in those markets, he said, so they tend to use the company's technology as a development platform.
Another factor, Dellson noted, is the fact that the most commonly used bioinformatics applications are open source, so Mitrionics did not have any IP hurdles in its way in developing the application suite, which will be freely available with source code to customers of its Virtual Processor technology.
Dellson said that this is an important feature in its strategy for competing with traditional accelerator vendors. "Many pharmaceutical companies run their own variants of Blast-related codes," he said, "so this is a huge opportunity for them because all of a sudden they can do that, and get this extreme performance, which was not possible before with the black-box approach because you couldn't look into that box."
The company does have some life science customers for its system, and Dellson said he's aware of customers developing FPGA-accelerated applications for phylogenetic trees and molecular dynamics algorithms like Gromacs, NAMD, and Amber.
Nevertheless, Mitrionics views its upcoming bioinformatics package as key to expanding its presence in the life science market. "This bioinformatics package based on Blast … we think is so strategic and so useful to people that it's a great door-opener to introduce FPGAs as kind of general-purpose resource in a computer system."
But it's unlikely that FPGA-based systems will ever cross over entirely into general-purpose territory. The technology is "right for certain tasks, but won't be for everything, clearly," said ABCC's Collins.
"Is it the panacea? No. But is it incredibly promising? Yes," Collins said. "The vision of many of us is that you write your program and the general processor will interpret what you want to do and figure out how to do it in the hardware whether with an FPGA or whatever new technology comes down the line and do it as fast as the clock rate will allow it to go, and you get orders of magnitude speedup over what you're getting today. That would be really, really cool, and there's a lot of people thinking about that," he said.
"I'm not going to see that probably in the next five years, but I'd like to be surprised," Collins added.
Bernadette Toner ([email protected])