Graphics chips used for general-purpose computing (aka GPGPUs) are also something of a new curiosity in embedded. GE’s Pepper calls them “an enabler, a pull through” to help win bigger contracts. They are appearing in high-end radar systems and an emerging class of embedded supercomputers for full-sized drones and other military apps.
GE uses both Nvidia and AMD graphics programmed both with Nvidia’s proprietary CUDA and OpenCL. “CUDA is much richer programming environment so right now it has the technical advantage, but in the long haul OpenCL will be the ultimate play,” he said.
Early versions were racks on rollers (left). Currently versions are more densely packed units in smaller chassis (below).
At the high-end, GE is shipping boards supporting 10G Ethernet and taking its first orders for ones running double-data rate Infiniband. So far it’s been able to avoid the switch to optical links but “at about 14G things get hairy for copper,” he said.
“We’ve talked about optical for more than 30 years, and there’s a lot more discussion than deployment,” he said. “There will come a time for optical, but if we can make do with copper we will,” he added.