Coinciding with the launch of Nvidia's next-gen Tesla GPU computing processors, AMD has announced a competing product that also breaks the one teraFLOPS barrier. AMD claims its new FireStream 9250 can compute a trillion floating point operations per second in single-precision mode, all with a thermal envelope of "less than 150 watts"—just under the 160W TDP of the new Tesla C1060.
Like Nvidia's GeForce GTX 280, the new FireStream packs 1GB of GDDR3 memory, and it features a hardware double-precision floating point implementation. AMD's press release boldly asserts that the FireStream 9250 "occupies a single PCI slot," too, although our money is on it requiring a PCI Express x16 slot like the Tesla (and AMD's previous-gen FireStream card). AMD says it plans to release the FireStream 9250 together with a development toolkit in the third quarter of this year for $999.
Speaking of development toolkits, AMD also mentions that it has joined the Khronos Compute Working Group—a group that intends to develop "industry standards for data parallel programming and [work] with proposed specifications like OpenCL." If OpenCL rings a bell, it's because Apple plans to support it in its Mac OS X 10.6 Snow Leopard operating system due next year. Both Apple's and AMD's descriptions make OpenCL sound like a higher-level language more similar to Nvidia's CUDA than AMD's Close to Metal API. AMD says it "intends to ensure that the AMD Stream SDK rapidly evolves to comply with open industry standards as they emerge."