Nvidia launched its second-generation GPU computing architecture on Wednesday, code-named "Fermi". Oak Ridge National Laboratory will design a supercomputer based on the Fermi, an executive said.
The "Fermi" announcement kicked off Nvidia's GPU Technology Conference, which will run the remainder of this week in Silicon Valley.
Fermi will serve as next generation of CUDA, a programming architecture Nvidia developed in 2006 in conjunction with the GeForce GeForce 8800 architecture, which included a unified compute processor that could either perform graphics-specific tasks or could be programmed in C to be used for general-purpose computing. Because they do specialized tasks – computing graphics – extremely quickly and efficiently, scientific applications that do a number of tasks over and over again could take advantage of the DSP-like graphics-chip architecture.
The GT200, announced in 2008, added dual-precision floating point technology, which the GeForce 8800 lacked, according to Nathan Brookwood, an analyst with Insight64.
Now, the Fermi adds additional features that those apps should be able to take advantage of: C++ programming, error-correction code (ECC), and what Nvidia calls Parallel DataCache, an optimized cache hierarchy to speed up raytracing and physics calculations through a configurable shared L1 memory and unified L2 cache. A GigaThread Engine allows different kernels of the same application to concurrently execute, improving performance further. The memory address space has been expanded from 4 Gbytes to a full terabyte, accessed via six 64-bit interfaces to external DRAM.
Nvidia said that Fermi would include 512 cores supporting IEEE 754-2008 floating point accuracy, a step up from the dual-precision FP used by the GT200 architecture. Double-precision floating-point math can be performed up to eight times faster than existing GPUs, according to Tom Halfhill, an analyst with In-Stat.
The Fermi architecture will also be integrated in to Microsoft's Visual Studio 2008 SP1 via Nexus, a development environment that includes a debugger for source code, including breakpoints and direct GPU memory inspection; an analyzer or performance tool for GPU and CPU events, and a graphics inspector to debug profile frames.
"This is the first PC GPU architecture designed specifically for supercomputing," said Andy Keane, general manager of the TESLA business unit at Nvidia. "We're delivering the features and performance at one-tenth the cost and one-twentieth the power of the previous generation."
Interestingly, even though the AMD Radeon HD 5800 series has received rave reviews, from a computing perspective Halfhill characterized Nvidia's Fermi as uncontested.
Fermi surpasses anything announced by Nvidia's leading GPU competitor, AMD," Halfhill wrote. "Conceptually, Fermi's closest competitor is Intel's Larrabee. Expected to debut next year, Larrabee chips will attack the market from the opposite direction. They will use general-purpose x86-architecture cores for both 3D graphics and GPU computing – a radically different approach than adapting a specialized 3D-graphics architecture for general-purpose computing. Until the first Larrabee chips appear, it's not clear whether the x86 can make such a leap."
Copyright 2014 © Godem Online Inc. | Web and server solutions by NewTech Solutions.