One person particularly happy about Intel’s announcement last week that it is buying semiconductor vendor Altera was Subbu Rama.
Intel is spending $16.7 billion to buy Altera primarily to bring the company’s programmable chip technology in-house. Field-programmable gate arrays (FPGAs) can be programmed through software and are becoming increasingly important ways to accelerate applications in cloud and Web-scale environments.
Rama, a former Intel employee, is now CEO of Bitfusion, a startup that offers software that can help boost workload performance even more in systems that use accelerators like FPGAs and GPUs from Advanced Micro Devices and Nvidia. Organizations are increasingly using accelerators for a broad array of applications—such as big data analytics, bioinformatics, pharmaceuticals, and oil and gas—to speed up performance while keeping a lid on power consumption.
Intel’s pursuit of Altera proves the soundness of that approach, according to Rama.
“We just got a huge market validation for what we’re doing,” he told eWEEK. “There is a lot of interest.”
Bitfusion debuted last month at the TechCrunch Disrupt NY event, with Rama laying out his vision for the company. Based in Austin, Texas, Bitfusion was started by Rama and two other one-time Intel employees, COO Maciej Bajkowski and CTO Mazhar Memon. The company recently raised $1.45 million in seed money from investors Data Collective, Geekdom Fund, Resonant Venture Partners and Techstars.
Bitfusion also has entered into a partnership with Rackspace Hosting, enabling potential customers to access its software technology through the cloud and try it out. Initially the Rackspace offering will be used as a test bed for customers, though there is the possibility of creating a managed service via the cloud server provider down the road, Rama said.
The company’s technology comes as organizations are looking for ways to better address the challenges brought on by the massive amounts of data that is being generated, Rama said. There not only is more data, but it’s coming in different formats through various hardware form factors, software applications and social media avenues, he said.
Most organizations currently are dealing with all this by adding server nodes to their data center, which only adds to the cost and complexity, he said. However, accelerators like FPGAs, GPUs and Intel’s x86-based Xeon Phi chips enable businesses to get more application performance from their existing hardware infrastructure, something the high-performance computing (HPC) and supercomputing fields already understand.
Bitfusion’s software is designed to enable organizations to get even more workload performance from the accelerators by enabling the applications to better leverage the capabilities of not only the CPU but also the accelerators, in large part by building optimized versions of open-source libraries, Rama said. In addition, that can be done automatically without having to rewrite the code. The software enables the application to access multiple accelerators and the CPU at the same time, Rama said.
“People really don’t rework their code,” he said. “They want their code to work and want to leave the magic to the software producers. … We do it through libraries that we’re building.”
Startup Bitfusion Wants to Speed Up Server Workload Performance
Rama said his company’s use of recompiled libraries will help speed up application performance in systems with accelerators and those using only CPUs. However, while doing so will give workloads a speed increase, a challenge is the lack of a clear standard for interfaces between the accelerators and application, which means having to deal with multiple standards, from CUDA to OpenCL to AVX2. However, Rama said that OpenCL appears to be becoming the “base layer” for accessing devices.
Bitfusion officials see a time when the software will lead to a 100-fold performance boost in workloads.
Accelerators have been used in HPC systems for several years, though for the most part they have been GPUs. However, FPGAs—with their ability to be reprogrammed through software to meet the demands of particular workloads—are getting more attention. While GPUs are found in many HPC environments, FPGAs are better used in situations where software is apt to change.
Bitfusion is starting off working with CPUs, GPUs and FPGAs from Intel, AMD and Nvidia, and has plans to support ARM-based server chips in the future, Rama said. The company also is looking into IBM’s Power architecture.
Customers will be able to access the technology by putting the software onto their own systems using Bitfusion Boost, or buying systems from Bitfusion that already have the accelerators and software installed in them. The company is working with Dell and Supermicro to build the Bitfusion Appliances. The other way is through the Bitfusion Supercloud via the company’s partnership with Rackspace.
“The key is software,” he said. “The PC became popular because of Microsoft. … If you don’t have good software, the hardware will not work.”
Bitfusion’s direction is similar to that of SRC Computers, a 16-year-old company started by the founder of supercomputer Cray that has been selling HPC systems to defense and intelligence agencies for about a dozen years. SRC in May announced its entrance into the commercial server space with the Saturn 1 systems, a dynamically reconfigurable server based on a new computing architecture that leverages FPGAs in a way that enables customers to significantly speed up performance and use 1 percent of the power and space of traditional servers at about 25 percent of the cost.
