Keep your Playboy mansion, Supermicro is my nerd vice palace
I'm not ashamed to admit I was drooling over those racks
Sysadmin blog With several clients facing refresh cycles, I've decided to poke my nose back into the tier-2 space to see what’s going on.
Supermicro has been on my preferred vendor list for years, so when I was in the Bay Area this summer I stopped in to see what’s new. I left impressed with what I saw, and got a swank backpack admittedly full of propaganda out of the deal. It proved to be a helpful bonus given the refresh planning underway.
Supermicro earned its name for making motherboards, workstations and servers that have become the backbone of many mid-market companies – many of my customers among them. At some point when I wasn't looking, Supermicro embraced the high-performance and high-density computing markets. Now they make things like their Microcloud server, eight Xeon E3-powered 32GB nodes in a 3U box. There's a 12-node variant on the horizon.
Supermicro became a force in graphics-processor-led computing by cramming GPUs into everything: 1U pizza boxes with 4 GPUs all the way up to a 7U SuperBlade with 20 GPUs and 20 CPUs. On the drool-worthy side, Supermicro offers pre-canned GPU clusters doing up to 42 teraflops in a single rack. Its most recent unveiling is a big-data Hadoop rack. For me, this qualifies as top-quality nerd porn.
My favourite discovery is the Fat Twin, which has two configurations: eight compute nodes (dual Xeon E5-2600 with 512GB of RAM) or four storage nodes (eight 3.5in drive bays and dual-core CPU) in a 4U box. The chassis offers redundant power supplies without the single points of failure of a traditional blade chassis; Supermicro makes its own PSUs, explaining how they managed to source power supplies with an 80 PLUS platinum energy-efficiency rating.
Supermicro’s intent with these is high-density supercomputing and heavy virtualisation applications. I took one look and saw “private cloud in a box”. In theory, one could cram four compute and two storage nodes into 4U. Sadly, we cannot mix and match compute and storage nodes in the Fat Twin chassis today; it's prevented by a small design item in the chassis.
I remain hopeful that if I poke the powers that be enough, the future will see a 4x compute and 2x storage configuration. Supermicro's roadmap for the Fat Twin line contains a 4-node 12x GPU and Hadoop-optimised configs.
Somehow, despite using Supermicro as a vendor for years, I had completely failed to notice Supermicro’s advances. I almost never ventured past their motherboard range; they simply didn't occupy mindshare anywhere else. It was a bit of a shock when I finally realised that my tier-2 motherboard vendor makes everything from high-density GPU wizardry to switches and even their own racks.
Is my completely missing the evolution of a company like this a marketing failure? Certainly I should have kept up with my vendors more than "every few years". At least some of this is marketing; Supermicro – like most tier-2 vendors – isn't a darling of the headlines. They don’t get repetitive press coverage and thus don’t get mindshare.
If a company like Supermicro can diversify its portfolio as much as they have in the last few years, I’m suddenly curious to see what everybody else is doing. How much neat stuff am I missing out on because it doesn't make it onto my radar? Now if I can just figure out how I go about expensing one of Supermicro's Hadoop racks. ®