Posted by John Keller

THE MIL & AERO BLOG, 19 Feb. 2013. I wonder who remembers the days when military high-performance embedded computing , which today we call HPEC , involved PowerPC microprocessors with AltiVec technology, or before that the Intel i860 digital signal processor.

Anyone? (crickets) ... anyone at all?

Well all right then, but my point is that we've come a long way since the i860 and AltiVec days, driven by some of the latest central processing unit (CPU) technologies like the latest-generation Intel Core i7, the latest field-programmable gate arrays (FPGAs) from manufacturers like Xilinx, Altera, and Microsemi, and general-purpose graphics processing unit (GPGPU ) technologies from NVIDIA and AMD.

I'm a little sheepish to admit that I remember when one of the fondest computer science goals of the Defense Advanced Research Projects Agency (DARPA) was to create "a gigaflop in a soup can."

Today we're approaching accessible embedded computing technology that can provide a gigaflop on a chip, and there doesn't seem to be any end in sight.

GPGPU technology -- which is a massively parallel processing architecture that particularly lends itself well to floating-point digital signal processing -- continues to evolve on a trajectory similar to what we've known as Moore's Law, and GPGPU software programming tools are making this technology broadly available. FPGA devices and software programming tools are improving apace.

DARPA -- never to be counted out when it comes to high-performance military computing -- lately has launched the ICECool program, short for Intrachip/Interchip Enhanced Cooling, which promises eventually to fabricate in-device liquid cooling as part of the semiconductor manufacturing process.

The ICECool program, now aimed primarily at embedded computing and RF monolithic microwave integrated circuit (MMIC) power amplifiers, seeks to remove waste heat from electronic components at the rate of one kilowatt per square centimeter heat flux, and one kilowatt per cubic centimeter heat density.

DARPA's involvement in HPEC expends much farther than the ICECool program. Last fall the agency awarded research contracts to SRI International in Princeton, N.J., and to Reservoir Labs Inc. in New York for the Power Efficiency Revolution For Embedded Computing Technologies (PERFECT) program, which seeks to overcome power efficiency barriers that limit the capabilities of military embedded systems.

The DARPA PERFECT program aims to increase the power efficiency of HPEC systems to from today's 1 billion floating point operations per second per Watt (GFLOPS/w), to 75 GFLOPS/w.

Imagine the kind of embedded computing capability these technologies could yield if they ever come to pass.

Today military intelligence, surveillance, and reconnaissance experts complain that today's data processing capability and thirst for intelligence, surveillance, and reconnaissance (ISR) information leaves us "swimming in sensors, and drowning in data."

Well, the HPEC technology available today, and promises for the future, could go a long way to solve that problem.

The issue with swimming in sensors and drowning in data isn't so much a problem of too much data, as it is too much SUPERFLUOUS data. Sensors are everywhere churning out oceans of data. There aren't enough processors or human analysts to keep up -- today that is.

But just take a look at GPGPU technology. This processing approach began life as a high-performance graphics engine aimed at high-end gaming. It didn't take long for the embedded computing community to catch on that GPGPUs were high-performance parallel processors applicable to DSP applications like radar and sonar processing, electronic warfare, and signals intelligence.

One interesting thing about GPGPU technology, however, is not so much its intrinsic parallel processing capability, but is closely related to its graphics processing roots.

One expert explains that graphics processors are particularly good at putting together lots of data into a coherent picture. What embedded computing designers also are finding is that the graphics capability of GPGPUs also are extremely good at TAKING APART complex pictures boiling them down only to the information that's needed.

Think finding a needle in a haystack, really fast.

Now think about persistent-surveillance applications that can stare at a wide area like a city for weeks and months at a time. From that universe of data, might be quickly extracted information about specific terrorist or other suspicious behavior ...

... or finding other kinds of needles in different haystacks.

It's going to be very interesting.

There's one thing that bugs me, though, and it involves GPGPU technology -- more to the point, its crummy name.

In the aerospace and defense community, HPEC designers primarily are involved with complex signal processing, not graphics, so I think we ought to lose the GPGPU name and its reference to graphics.

So I'm proposing a new name for what until now we have known as GPGPU technology, and here it is: high-performance embedded parallel processing, or HPEPP for short.

Give it some thought; heck, it must might catch on.