While the annual SC conference is always a worthy experience, SC16 may have been the most rewarding ever – filled, as it was, with some fantastic insights from fellow HPC professionals, news on the latest technologies that will advance our industry, and the chance to catch up with colleagues and friends old and new.
At RedLine, we’re still talking about the show, and thinking about how we’ll deploy what we learned to our clients’ benefit. I asked our team to jot down a few of their own impressions of SC16 while the event was still fresh in their minds. Here’s what they had to say:
Mark Potts: Overall, SC16 was quite the boisterous, energetic bustle of over 10,000 people meeting, reacquainting, and exchanging ideas for future projects and business. This year, I found that there was considerable discussion of GPU acceleration and deep learning (using GPU acceleration), which coincided with NVidia’s humongous booth at the center of the trade show floor. I also felt that there was a bit more emphasis on community-based development efforts, and was pleased to see that The Portland Group Inc. (PGI) has released a “community” version of its compiler, which can be downloaded by individuals for free. I am already working with it on my laptop!
Chris Young: Mellanox’s new offerings in HDR is big news. Mellanox InfiniBand started over 10 years ago at SDR rates running 4 lanes at 2.5 Gbps (10 Gbps) and doubled that with DDR (double-data rate) at 20 Gbps a few years later. Then they jumped to QDR (quad-data rate) at 40 Gbps … to 56 Gbps with … and most recently to EDR at 100 Gbps.
With HDR, they have truly doubled-down on performance. This is a massive jump. They announced a 1U 40 port switch (an odd increase over the previous 36 ports they’ve offered for years). HDR offers full backwards compatibility with previous versions of InfiniBand.
An interesting feature Mellanox offers is a splitter cable which takes a 200 Gbps port and splits it into two 100 Gbps connections. This offers an interesting value proposition. Potentially “future-proofing” your system allowing you to start with lesser expensive EDR HCAs in your servers while decreasing your initial capital outlay while investing in a future technology you can deploy in your following cluster.
We’re still waiting for Mellanox to release its large footprint “core” switches in HDR before we’ll be able to see the full solution.
Mellanox had a very glitzy presentation at its reception with smartly dressed hosts and hostesses – but the stars of the show were comedians including Gabriel “Fluffy” Iglesias. The entire presentation is available to watch here.
Kit Menlove and Terry MGuinness: Primarily due to energy and thermal limitations, the push to exascale and beyond involves a movement into increasingly varied (heterogeneous) and reconfigurable hardware and runtimes. Many researchers are now pushing the idea that the days of “write once, run everywhere” in the supercomputing space are winding down – that more effort needs to be put into making code easier to port and optimize than making it more portable out of the box.
One significant contributor to this is the CU2CL project which can now translate over 90% of CUDA code into OpenCL (and mark up the rest for manual intervention). While other similar projects are also underway, more attention needs to be given to making code easier to refactor, easier to port to a different runtime, and easier to optimize – we will be doing a lot more of it in the days ahead!
Adam Dorsey: There was a strong showing from ARM at SC16, with ARM itself having a booth, as well as Cavium, Colfax, SGI, and other ARM partners showing off ARM-based systems. Intel had its Phi architecture on display, with SuperMicro, Tyan, and Colfax (again) also displaying systems with either Knights Landing CPU-based models (Colfax) or the more traditional accelerator card setup (Tyan, SuperMicro, others). Also, nVidia was promoting its newly-released Pascal architecture, and AMD showed off ROCm, the Radeon Open Compute platform. The popularity of these alternative architectures continues to grow as gains in power efficiency and speed are realized by clusters using them. At least one speaker session at SC16 addressed this movement away from the traditional Von Neumann architecture, detailing the ongoing paradigm shift, as well as possible future architectures and paradigms.
Get In Touch
As you may have guessed, we were certainly busy at this year’s show. We learned a lot, and we were happy to share our own expertise as well. If we didn’t get a chance to meet with you at SC16 last month, let’s rectify that. Just reach out and we’ll set up a time to talk about your HPC needs.
Keep connected—subscribe to our blog by email.