Until relatively recently, AMD was a definite force on the Top500 list of supercomputers, especially once its Opteron managed to eek its way into around 108 systems on the list back in 2007, giving Intel a run for its x86-64 money.
The last couple of bi-annual Top500 lists, however, show that Opteron share declining, despite what the company’s GM of the Server Business Unit, Andrew Feldman, calls superior technology, particularly on the APU, interconnect and power efficiency fronts.
To highlight the changes, we took a look at the Opteron share of the Top500 list based on each June’s numbers since after AMD announced it in 2003. As you can see below, they clearly peaked in 2008 and are on the downslide, especially when compared to Intel’s 80% slice of the Top500 pie in 2013.
The graphic below, which is based on all Opteron generations, shows a definite high point, but with Xeon Phi thrown into the mix and some new innovations on the performance front from NVIDIA, and the potential for some strong ARM-ing, it’s hard to foretell just how steep the next step down will be if we refresh this graph in November following the next list.
Remember too, that it’s only so comprehensive since it doesn’t take flops or different systems into account–and recall that there’s a small Cray component here since several systems got a round of new Opterons–but of course, that’s not going to be the case since the XC30 is Intel-only. The point is, there is downward trend.
During a conversation this week with Feldman, we asked quite frankly how invested AMD is in the HPC market–a question we’d hoped to ask in person at ISC but AMD didn’t have a direct presence for the first time in years. While he admitted that HPC is a thin slice of their overall business, it’s still an important one, and still the target for further development on their APU, especially since lucrative markets are demanding power-efficiency and performance.
Feldman pointed to a few key areas where they’re focused on offering up HPC options. While academia and research were token mentions, the hot areas for their APU he said are in both oil and gas and government–the “three-letter agencies” to be exact.
“When we look at what industrial supercomputing users are looking for, it’s the ability to deliver more compute per dollar–and then more compute per watt. If you look at the Top500 or Green500, it’s not about the horsepower of each piece, it’s about how many pieces you can tie together, how you can break up the work, and do a placement algorithm.”
“if you think about every person being recorded at an airport, all the facial recognition work that’s being done, and even what came out recently about the NSA and its operation that’s where our APUs shine. The government is looking for power efficient architectures that offer the performance they need, Feldman says.
Pattern recognition, log analysis and other similar gather and hunt operations, however, seem to fall more in the field of large-scale data analysis as it’s being done on stripped-down hardware. In other words, this is an ideal fit for Hadoop, MapReduce and the slew of other new frameworks that have sprung up to tackle such problems.
When asked whether or not what the government is doing on the data mining fron is really HPC, he said that there is some use of MPI and other techniques to extract performance. The oil and gas industry, is a standard HPC area, but a lot of what Feldman discussed fit into more of a “big data” paradigm. This makes sense, however, because their roadmap is about broadening their reach–the Opteron isn’t specifically for HPC applications, he reminds, subtly hinting at the fact that Xeon Phi is. They’ve opened a larger market with their Opterons and will continue to do so via the Seattle project and its coming push to lay the Freedom fabric (which comes from SeaMicro–the company Feldman spearheaded) on the die with ARM cores.
Feldman is confident the HPC community will take to what’s now called Seattle, a project to move the Freedom fabric onto the die with ARM cores, which is expected to happen in Q1 of 2014. He says that their differentiation is the fabric, especially with the move to smaller (but more efficient) cores. Since the matter of overhead on I/O, disk and other components doesn’t disappear, the interconnect becomes increasingly important.
Feldman said they’re not even trying to keep up with Intel, they’re trying to innovate right around them. It’s not a game of “catch-up” for anyone but Intel, says Feldman, pointing to their numerous fabric buys (QLogic, Aries) that have yet to amount to anything that’s in production.
On that note, he says that even if Intel did manage to tap those fabrics, they are relevant for the supercomputing market which again, he says is a “very modest” part of their business and roadmap. AMD is focusing on creating new chips that will appeal to the supercomputing crowd but keep them afloat because of real enterprise viability.
And, somehow, he managed to say all of this without uttering the phrase “big data” once. And thank you.
Feldman says users, at least the enterprise and big government ones they’re chasing, don’t want an “array of GPUs that are doing the narrow work they’re best at, then also having to move that data back and forth from the GPU’s DRAM into the CPU’s DRAM or by trying to create extraordinarily complex software systems to write to it via CUDA and the like.” He claims hat what users need is “a single processor that has graphics and more traditional compute and allows both to access the same DRAM” and a simpler programming model–in their case, HSA.
The problem, and Feldman agreed, is that AMD isn’t getting this message out there in terms of its HSA He says that it’s a shame because it’s far easier to write than OpenCL and CUDA, at least for those who can write code in C–although that base is probably not as big as it used to be.
Again, the aim is to appeal to a far wider set of users than NVIDIA is targeting with its GPU and Intel with Phi. But Feldman assured us that this widening of the net doesn’t mean they’re letting HPC slip through the holes.
“We believe that the combination of best in class fabrics like our Freedom fabric and our ability to build the industry’s best APU will have a meaningful and direct impact on the HPC market.”A market that, Feldman notes is a pretty thin slice versus a healthy chunk of pie.