June 22, 2007

NVIDIA Takes Direct Aim at High Performance Computing

by Michael Feldman

For the past year and half, NVIDIA has been putting together the product strategy for the company's high performance computing platform. On Wednesday, NVIDIA announced Tesla, a GPU product line targeted squarely at HPC customers. The new NVIDIA products are designed to act as computational accelerators for workstations and servers that host high performance technical computing applications.

Tesla represents an evolution of NVIDIA's thinking about serving HPC customers. Last year, the company entered the arena of general-purpose computing with GPUs (GPGPU) in earnest with their high-end GeForce and Quadro GPUs. For software support, they introduced their CUDA C compiler to offer relatively low-level access to the computing capabilities of their GPUs. According to NVIDIA, the CUDA tools have been downloaded by 3000 to 4000 developers since it was introduced in November 2006. For those interested in higher levels of abstraction, a GPGPU MATLAB library plug-in will soon be released.

With these early tools, technical computing users were able to demonstrate application performance increases of between 40 and 240 times compared to traditional x86 platforms. The applications ranged from neuron simulation and seismic modeling to MRI processing.

But the GeForce and Quadro products are designed mainly for visualization applications in a personal workstation or PC setup. There is no reasonable way to scale these devices across a cluster of servers to achieve a more generalized HPC solution. Nor was there a technology roadmap for NVIDIA's mainstream GPU lines that emphasized computing performance over graphics performance. Tesla now makes this possible. With the three separate GPU product lines, NVIDIA is able to target distinct application areas that reflect the company customer base. The GeForce products are geared for consumer/entertainment computing and visualization applications; the Quadro boards, for professional design and creation applications; and now the Tesla products, for traditional HPC applications.

Tesla was designed with the kind of form factors, power profiles, reliability levels and interconnect types that are compatible with high performance computing workstations and server platforms. There are three initial offerings: a 4-GPU server board, a 2-GPU workstation board, and a GPU computing processor. All the initial products will be based on the current high-end Quadro GPU, offering over 500 gigaflops of single precision performance per processor.

The Tesla S870 server board is really the big breakthrough for NVIDIA, since it represents their first product designed for the HPC datacenter. It fits in a 1U chassis, contains four GPUs, and communicates with the server host using a Gen 2 PCI Express switch. Temperature sensors and system monitoring are included to provide the level of reliability expected in datacenter hardware. The board dissipates 550 watts. Add another 10 watts for a PCI Express host adapter card. That might seem like a lot of juice for an accelerator, but for 560 watts you get over 2 teraflops of single-precision performance. MSRP for the server board is $12,000.

The Tesla server also comes in a 2-GPU version, and an 8-GPU version is in the works. The latter configuration is expected to improve upon the performance per watt ratio somewhat.

The other two initial Tesla products are targeted for workstations or PCs. The Tesla D870 is a 2-GPU board that connects to a deskside workstation. Like the server product, it connects to the host via PCI Express. The D870 uses 550 watts of power and lists for $7500. The Tesla C870 is a single 170 watt GPU processor that fits in a PCI Express slot in a workstation or PC. It lists for $1,499.

Andy Keane, general manager of GPU Computing at NVIDIA, thinks most of the company's early technical computing customers will migrate from the current GeForce and Quadro platforms to Tesla. Customers that are using the current products for both visualization and computing may stick with them if the computing side of their application doesn't outrun the GPU performance. But Tesla is clearly meant to be the future of technical computing at NVIDIA.

Although the initial offerings are based on NVIDIA's 8-series devices, as Tesla evolves it will sport its own GPU variants, which may run with faster clock speeds (but perhaps slower on-chip memory) than GPUs whose primary focus is to drive visual displays. More significantly, Keane says that double-precision floating point capability will be added to the entire Tesla product line by the end of 2007.

The addition of double-precision capability will open up the entire technical computing market for NVIDIA, since the inherent limitations of single precision arithmetic will be removed. So unless AMD comes out with a double precision GPU in the next few months, NVIDIA will be the vendor to pioneer 64-bit floating point in GPGPU computing. As such, it becomes a more direct competitor with ClearSpeed boards, a math co-processor offering that also targets the HPC market. Although NVIDIA has not released power or performance specs for their upcoming double-precision devices, one can surmise that ClearSpeed will be able to claim a performance per watt advantage, but perhaps not a performance per dollar advantage. Depending on how Intel's Larrabee processor development plays out, NVIDIA could eventually run into additional competition there as well.

In any case, there may be plenty of acceleration opportunities to go around. The commercial HPC market is growing rapidly — even faster than the general IT market. According to IDC, technical computing revenues will reach $14.2 billion by 2010. Currently, the oil & gas and financial services segments represent two of the highest growth areas right now. But manufacturing, biotech and government HPC are also expanding. NVIDIA thinks its new HPC line can ride a lot of this growth as users start to figure out that Tesla-equipped workstations can replace decent sized clusters and Tesla-equipped clusters can match the raw performance in some high-end supercomputers.

“Now when we go into an IT department and they ask us how to put GPUs into their datacenter, we have a specific answer and a product that exactly fits what they expect to buy,” says Keane.