A recent performance benchmark conducted by Citihub resulted in a lot of talk about “overclocking” of CPUs, and what the driver for it is. IntelligentTradingTechnology.com asked Chad Attlesey, CTO of overclocking leader Hardcore Computer, to explain what it’s all about, how to overclock reliably, and what performance gains might be seen.
Q: Can you begin by explaining what overclocking means, how it came about, and why one would want to do it?
A: Overclocking is increasing the clock frequencies of computer components (CPU, GPU, memory, etc.) faster than their default settings. Computer components are often limited in performance by the power envelope they need to operate at. This almost always is directly tied to limits of conventional cooling systems (air or hybrid cooling) to keep the components cool at higher temperatures resulting from higher operating frequency. Running faster requires more power, which in turn creates more heat. If you don’t remove the heat adequately, the parts don’t operate properly or outright fail.
There is always a demand for higher levels of performance. Overclocking has been around practically as long as the electronic computer. In the past, overclocking often produced unreliable results. However, Hardcore Computer’s overclocking performance is tested and guaranteed stable for use at the enterprise level.
At Hardcore, we often refer to overclocking of certain components as “right-clocking” because we know there is a lot of lost potential in simply running at default speeds. Anyone that wants the fastest computer with the lowest latencies possible would want to do this.
Q: But as you say, traditionally the downsides of overclocking are heat and reliability issues. How does Hardcore address those?
A: One thing Hardcore does better than everyone else is cool all components and keep them in the temperature range they are designed to run at optimally 24/7/365. Our patented total liquid submersion cooling system allows us to keep the computer system stable, even under high overclocking speed conditions. You are only as good as your weakest link. Hardcore cools all electronic components through direct contact with “Core Coolant” dielectric oil that fills and flows through our systems. At a very low flow rate, Core Coolant very efficiently and effectively removes heat. By eliminating the thermal barrier, performance is then, for the most part, limited by what the silicon can do. We employ hardware, firmware and software solutions that help us to further push the limits of the silicon.
Q: What server and blade products have you developed? What’s the relationship with Supermicro?
A: Our current server offering is the LSS200. The LSS (Liquid Submerged Server) systems we currently ship make use of the Supermicro X8DTT line of twin server boards. They plug into a standard 19-inch-rack-mounted shelf. We can fit eight dual socket systems in a 5U height. Hardcore enjoys a great partnership with Supermicro. We have been working together with motherboard overclocking work, next generation product development, as well as coordinated marketing and sales efforts. Supermicro is a great match for us.
Q: What kind of CPU speed up is possible with your servers, and what is the typical resulting performance boost at the application level?
A: Overclocking can result in performance gains anywhere from 20% to 100% or more depending on the CPU. With current top of the line Intel Westmere CPUs, we can see as much as 40% improvement or more. An interesting thing to note is that application performance is often improved exponentially, rather than linearly by overclocking. This has to do with improvements in bus and memory operation, in addition to improved CPU clock frequency. For example, in certain cases we have seen as much as 8x application performance boosts with 30-40% overclocking.
Q: As well as the CPU, do other system components benefit from the liquid cooling?
A: When you overclock the base clock frequency of the CPU, memory speed is also increased. Bus communication can also be improved. We have been shipping desktop systems since 2008 that employ overclocking and optimisation of the overall system. We have been shipping overclocked GPU solutions to the consumer market since we started shipping product. Because of this stable overclocking capability, we often ship systems that can more than double the benchmarks of other’s using similar components.
Q: How practical is it to simply replace conventional servers with your products? Are there any special requirements? How about external heat dissipation?
A: Hardcore has developed an enterprise solution that is easy to implement in any new or existing data center, or closet installation for that matter. LSS200 systems slide into a standard 19-inch-rack-mounted shelf and liquid manifold system. As I mentioned earlier, eight dual-CPU systems can fit in a 5U space. These systems are available with a large variety of I/O and PCIE 16x daughter card options. Power and I/O cabling is done just like any other discrete server.
If your data centre application does not have the ability to plumb coolant lines outside, does not have chilled water or simply restricts access to anything besides power and I/O outside the rack, we have a cooling distribution unit (CDU200) that plugs into the rack-mounted shelf and does heat exchange into your hot isle like your air or hybrid-cooled systems. You don’t get the improvements in energy efficiency of pumping the heat outside the data centre, but you still get the improved performance.
If you can pump liquid outside the data center to one of numerous energy-efficient heat exchange methods (like in-floor heat of adjacent offices or generate electricity from a geothermal reverse heat pump) this will give you the best return on investment. If you have chiller water in your data centre, use liquid-to-liquid heat exchange on the waste water side (we can use inflow temps up to 35C) for essentially free cooling. In these cases, the need for air conditioning is virtually eliminated, reducing your cooling costs by 80% or more.
Q: What’s the status so far in terms of real deployments by financial trading firms, and by software developers looking for a performance gain?
A: We have deployed proof of concept systems or product in eight of the top banks between London and New York. Several quant trading firms and other financial trading firms are already working with our systems. As far as developers looking for performance gains, we got some good indications of our potential impact in the financial market when we recently received this outside validation:
“The test results exceeded expectations, with an average latency of 237 nanoseconds with FAST size 16byte messages scaling to 271 nanoseconds with 128byte messages. This represents a 34% improvement on previously published intra-host benchmarks by TIBCO and is 3.3 times quicker than existing standards.”
Q: What’s next in terms of product development, and impact on trading firms?
A: Our goal is to continue to develop and ship the fastest, lowest latency, over-all highest performance systems on the planet. We listen to our customers and we will continue to use what we learn to make our products even better. We plan to release LSS systems with the new Intel architecture. Our current chassis solution will be able to swap-out upgrade to Intel Sandybridge. This allows for an easy, less expensive upgrade path for our customers. We also plan to expand our product line in 2012. These new products we think will allow Hardcore to deliver a more complete set of even lower latency solutions to much more of the trading and general FSI data center market in 2012.