Custom Hardware Making a Comeback

In IT circles, development tends to follow a steady path of more power, faster throughput and, lately, less energy consumption. Odd, then, that technology architectures behave like a pendulum: Yesterday’s old ideas become tomorrow’s innovative new approach.

We’ve seen this in the distributed vs. consolidated paradigm, clients vs. desktops, and the public-private-hybrid cloud debate. Now, it seems there is another shift in the works: the growing movement away from commodity hardware and back to customized systems.

To be sure, commodity still rules the roost, particularly when it’s of the x86 variety. And yet, even within the Intel platform, there is plenty of specialization going on. Intel has admitted that it occasionally provides customized silicon for its top-tier customers, primarily as a means to enable unique offerings in their product lines. But even when the system is not destined for the channel, Intel is not above a little custom work, as long as the buyer is interested in Google-level volume purchases.

In many ways, this latest round of custom hardware is driven by Google, which is known for deploying proprietary hardware throughout its data centers. Much has been made of the secretive Google server, but the company has its own networking designs as well. A purported “Pluto Switch” turned up in Shelby, Iowa, of all places, confusing a couple of techies on networking-forum.com who tried to power it up before returning it to Google. The company offered a reward for its return, but wouldn’t shed any light on its inner workings.

Meanwhile, Cisco is using a customized ASIC to power the new Algorithm Boost feature in the Nexus 3548 switch, rather than the standard Broadcom silicon it usually deploys. Company reps say this is the best way to drive new ultra-low latency networks, and it isn’t out of the realm of possibility for Cisco to adapt this philosophy to its network interface lines and even the Unified Computing server. The Algorithm Boost system features a warp mode that reduces the switch’s address table eight-fold, to about 8,000 hosts, enabling latency to drop to 190 nanoseconds.

Customization is also the way many smaller firms carve out market share. Infineta Systems, for instance, went with a customized appliance rather than a commodity server for its Data Mobility Switch, designed to improve WAN performance by a factor of 10. At the same time, Nimbus Data went with its own SSD design for the new Gemini storage system, lowering costs with a 2Xnm MLC Flash system rather than eMLC, and then adding beefed up DSP and ECC technology in the controller. This allows the company to fine-tune data handling between NVRAM and Flash to better suit targeted applications.

For a while, it seemed that hardware was largely irrelevent with the advent of virtualization and the cloud. For the enterprise, performance became a function of what was stipulated in service-level agreements. But CIOs are still under constant pressure to do more with less, so many are taking a renewed look at the physical layer again.

In many cases, commodity systems will continue to prop up the routine functions of the data center. But as specialty applications start calling for that extra push, nothing suits like a piece of custom hardware.