Supercomputers Fermi and fer you
AS Supercomputing '09 kicks off in Portland, Oregon, NVIDIA announced its new Tesla 20-series parallel processors for high-performance computing (HPC), based on its new "Fermi" architecture.
NVIDIA reckons its new cards have been specially designed from the ground-up for GPU parallel computing at one-tenth the cost and one-twentieth the power of a traditional CPU-based cluster.
"There is a big price difference between one kind of performance and another," NVIDIA's Andy Keane, general manager of the firm's Tesla business, told HEXUS, adding that most of the supercomputers in the world were actually not that fast, relatively speaking. Indeed, the 500th entry on the list is a one million dollar 17 teraflops affair.
But NVIDIA says it's planning to change all that with its Tesla 20-series, offering features that will enable a lot of applications to perform dramatically faster using GPU Computing. These include ray tracing, 3D cloud computing, video encoding, database search, data analytics, computer-aided engineering and virus scanning.
"If you can do in an hour what used to take you overnight, it can really change your life" said Keane, adding that giving regular scientists "what they couldn't have dreamed of before" was what motivated him in the business.
NVIDIA is boasting that its Tesla 20-series GPUs also combine parallel computing features which have never before been offered on a single device, including support for the next generation IEEE 754-2008 double precision floating point standard, ECC (error correcting codes), multi-level cache hierarchy with L1 and L2 caches and support for the C++ programming language.
Not to mention up to one terabyte of memory, concurrent kernel execution, fast context switching, 10x faster atomic instructions, 64-bit virtual address space, system calls and recursive functions.
Introduced as part of the Tesla 20-series are the Tesla C2050 and C2070, PCIe 2.0 cards with up to 3GB and 6GB (respectively) of on-board GDDR5 memory and double precision performance in the range of 520GFlops - 630 GFlops.
The family also includes the Tesla S2050 and S2070 with four Tesla GPUs in a 1U rack unit for cluster and datacenter deployments. The rackmount solutions feature up to 12 GB and 24 GB (respectively) total system memory on board, GDDR5 memory and double precision performance in the range of 2.1 TFlops - 2.5 TFlops.
NVIDIA has priced the C2050 and C2070 at $2,499 and $3,999, whilst the Tesla S2050 and S2070 will go for $12,995 and $18,995 respectively. All four offerings will be available from May 2010.