Sign up for our newsletter and get the latest HPC news and analysis.

Matrox intros GPU platform aimed at industrial imaging

Over the email transom, news that at a tradeshow last week Matrox introduced what it is calling its new high performance computing platform for industrial imaging: the Matrox Supersight e2.

What’s industrial imaging? Wafer inspection, CT scanning, that sort of thing. What does the card do?

Matrox Supersight e2 offers OEMs unprecedented performance and flexibility for machine vision and medical imaging applications that generate extraordinary amounts of data. Applications such as semiconductor wafer and mask inspection, flat panel display (FPD) inspection, and CT scanning benefit from performance gains with Matrox Supersight e2 by leveraging multiple clusters of CPUs, GPUs and FPGAs. Matrox Supersight e2 provides an environment for considerable data and task-level parallel processing through the interconnection of CPUs, GPUs and FPGAs using a unique PCI Express (PCIe) x16 2.0 (Gen2) switched fabric that removes I/O bottlenecks between the multiple processors.

“We have seen tremendous improvements in imaging application performance with multiple CPU cores, pipelines in FPGAs, and GPUs as accelerators,” says Dwayne Crawford, Product Manager, Matrox Imaging. “But to get maximum performance, the technologies must be judiciously applied in a system.”  Imaging applications differ significantly from traditional IT datacenter or HPC applications where the application is typically compute-bound. With imaging applications, the gigapixels per second of data leave most platforms I/O-bound. Matrox Supersight e2 offers a solution to this unique architectural challenge as traditional Blade servers and 1U “pizza boxes” simply do not have the I/O bandwidth to distribute images to the processors and accelerators.

Sadly, no OpenCL. You program these using the Matrox Imaging Library. My guess is that using one of these in a traditional HPC kind of application as one can do with NVIDIA’s products would likely be a real pain — you’d have to throwback to the days of casting your algorithms in imaging terms in order to get any benefit. To be fair, however, they don’t appear to be targeting this for us. It would be interesting to know if any of you have tried the product in a more traditional HPC application, though…anyone?

Comments

  1. As a note there are no restrictions in place preventing the use of native GPU programming methologies including DirectX, CUDA and OpenCL. However in order to communicate between multiple CPU boards over PCIe the MIL library will be reuqired and which point the native GPU methologies may be used with the GPUs belonging to the associated segment.

  2. John West says:

    Thanks, Dwayne, for the clarification.

Resource Links: