NVIDIA Fermi-Based Tesla Will Improve Supercomputing

Earlier today, NVIDIA announced the Tesla 20 series of parallel processors aimed at the high-performance computing market. Naturally, the products are based on NVIDIA's next-generation CUDA architecture, Fermi, and are actually able to deliver the same computing performance at a tenth of the cost and one twentieth of the power of standard configurations.

"Nvidia has deployed a highly attractive architecture in Fermi, with a feature set that opens the technology up to the entire computing industry," Jack Dongarra, director of the Innovative Computing Laboratory at the University of Tennessee and co-author of LINPACK and LAPACK, said.

The first Tesla 20-series products are the C2050 and C2070 single cards and the S2050 and S2070 GPU computing systems. The C2050 and C2070 come with a PCI-Express x16 interface, occupy two PCI slots, have up to 3GB and 6GB, respectively, of GDDR5 memory and provide double precision performance in the range of 520 Gflops to 630 Gflops.

On the other hand, the Tesla S2050 and S2070 GPU Computing Systems come with four Tesla GPUs in a 1U system product (meant for clusters and datacenters). They use a total of up to 12 GB and 24 GB on-board GDDR5 memory, respectively, and a double precision performance in the range of 2.1 Tflops - 2.5 Tflops.

Among the features that Tesla will provide are the support for the next-generation IEEE 754-2008 double precision floating point standard, ECC (error correcting codes) for uncompromised reliability and accuracy, multi-level cache hierarchy with L1 and L2 caches, support for the C++ programming language and as much as one terabyte of memory.

The processors will also be capable of concurrent kernel execution, fast context switching, 10x faster atomic instructions, 64-bit virtual address space and system calls and recursive functions. Among other features are the NVIDIA® GigaThread™ Engine (maximizes throughput), NVIDIA® Parallel DataCache™ (accelerates algorithms) and Asynchronous Transfer (turbocharges system performance by executing data transfers, even when the computing cores are busy).

NVIDIA has also revealed the pricing information. The Tesla C2050, C2070, S2050 and S2070 will become available in Q2 2010 priced at $2,499, $3,999, $12,995 and $18,995, respectively. More information can be found here.