Neurala?s Technology Platform significantly increases the execution speed of highly parallelizable brain-based algorithms
BOSTON, MA -- April 8, 2008 ? Neurala today posted results confirming that combining the Neurala Technology Platform with AMD?s Firestream Stream Processor? and SDK can run AI algorithms up to 230 times faster than a single core AMD Opteron CPU [1]. The Neurala Technology Platform enables AI researchers and developers to easily take advantage of the parallel architecture of GPUs and multi-core CPUs for implementing computationally intensive AI algorithms.
"Our vision is to realize the full potential of brain-based applications in both scientific and business communities throughout the world by using state-of-the-art parallel programming technology" said Massimiliano Versace, one of the founders of Neurala LLC, and Assistant Director of the CNS Technology Lab. "Neurala focuses on innovation and developing strategic partnerships aimed at helping programmers to harness the power of parallel processing in an inexpensive and easy to use technology platform.?
Dr. Versace explains that many brain-based algorithms, including neural networks, can benefit from parallel hardware when their computations are performed on many processing elements at once. This is especially true for architectures that use all-to-all connectivity between elements, as found in a wide variety of neural networks used in both academics and industry. Conventional CPU implementations require computations to be performed on one element at a time, which in the worst case leads to quadratic complexity and extremely slow execution speed. Users can greatly increase execution speed by implementing these networks on multicore CPUs and GPUs including the AMD Firestream Stream Processor?.
AMD?s Firestream Stream Processor brings developers one step closer to the GPU hardware through the use of the Firestream SDK, solving problems caused by the limited hardware access provided by current GPU implementations of neural networks. Neurala has been a part of AMD?s Stream Processor Initiative since July 2006, where Neurala?s engineers have been working to adapt Neurala?s Technology Development Platform to run with the AMD Firestream Stream Processor? and SDK.
These exciting results confirm the tremendous speed increases that researchers and engineers can achieve by using the Neurala Technology Platform in conjunction with AMD?s Firestream Stream Processor? and SDK for their highly parallelizable neural network applications. Neurala will release the Neurala Technology Platform to beta customers in the fall, and plans to collaborate with local research labs and companies in the Boston area to pilot test the platform in selected applications.
Opteron, ATI Radeon and CTM are registered trademarks of Advanced Micro Devices, Inc.
[1] The machine used for the simulations was a dual AMD Opteron 248 (using only a single processor for comparison), 2 GB SDRAM DDR 400 (PC 3200) ECC Registered dual channel, SUSE Linux 10 with custom kernel, currently 2.6.18.1 and ATI drivers with the upgrade date: 2007-02-15 fglrx-8.31.200 + 2007-02-15 ati-ctm-sdk-1.0.5beta5.x86_64. The simulations were performed using recurrent competitive field neural network with all to all connectivity. The largest speedup of 231 times was achieved with a network size of 150x100 neurons. The table below shows the results achieved with a few select network sizes.