Friday, October 14, 2011

An open letter to GPGPU-based embedded computing providers


Posted by John Keller

I've noticed a very strange and perplexing thing going on lately in the embedded computing industry concerning products based on general-purpose graphics processing units (GPGPUs). Everyone wants to be the first ... well, no matter what it is -- first for this, first to provide that, first, first, first.

I know of at least three companies out there involved in nasty sniping matches over who was the first to do some sort or other with the GPGPU, and it's getting old, quite frankly.

Now the GPGPU is great technology; no argument there. It's a powerful parallel processing engine broadly applicable to digital signal processing for radar, sonar, electro-optical sensor processing and a broad range of other applications central to the aerospace and defense industry.

But I have to ask an obvious question: who cares who's first? First this, first that, biggest, best, fastest, prettiest, whatever. I care about capability and applicability to challenges that aerospace and defense systems designers have, not who's the first at anything in this market, and I'll wager your embedded computing customers feel the same way.

Let's have a competitive discussion about the capabilities of GPGPU-based devices, and how these powerful devices can help military and aerospace electronics systems designers solve their most difficult problems.

Let's restrain ourselves over who's first, and leave that argument for the bar after work.

4 comments:

  1. A point well made. I am interested in providing an elegant systems solution that works efficiently, is supportable and remains so when my design is eventually deployed. To date most vendors seem to have slapped a few GPU chips on a board, mentioned something about CUDA and likely are already in the bar when you get there. But then of course they were there first.

    Michael Drake, Chordell Systems Ltd.

    ReplyDelete
  2. Good points, and the CUDA example is quite relevant, since Nvidia processors are not currently supported by any RTOS vendors - so actually could not be deployed in any safety critical avionics unless the device manufacturer were willing to build his own RTOS, which some do...but then they'd still need a DO-178B certifiable graphics driver, which is not available as a COTS product for Nvidia GPGPUs from anyone, so they'd have to build that themselves too, and take it through certification.

    AMD, however, supports mil-aero manufacturers, have GPUs that are supported by major RTOS vendors like Wind River and Green Hills, as well as mil-aero graphics card suppliers like Aitech, Tech Source, and Lockheed Martin, and there exist field-proven DO-178B certifable drivers from third parties like ALT Software, as well as component obsolescent management programs that offer 15-20 year life cycle support from Core Avionics & Industrial.

    ReplyDelete
  3. quantum3d and nvidia were providing rtos support to the mil/aero market almost a decade ago. the simple fact was that the do178b market wasn't a good investment of time/money and development. nvidia scraps more GPUs than the market will ever deliver.

    ReplyDelete
  4. There is a large difference between supporting mil/aero with GPUs for display (which have been around for a very long time) vs. general purpose signal processing (which has only been possible since 2002, when ATI shipped the first programmable floating point GPU).

    I agree that first by itself is of no consequence. But it does speak to who has the experience in building the hardware and software infrastructure to use these devices, tightly integrated into a heterogeneous system in a meaningful and optimal, SWaP-aware manner.

    However, claiming to be first if you know you aren't speaks to something else.

    ReplyDelete