Scott Ross, executive director of high-performance computing at UBS, and other Wall Street IT experts testified on behalf of Intel's new Nehalem (Xeon 5500) chip at an Intel event last night at Nasdaq Marketsite. (Incidentally, you wouldn't know a recession was going on at this event, which was packed with Wall Street techies and flowing with California wines and exotic appetizers.)

Interestingly, Ross strongly rejected the common notion that in order to benefit from a move to multi-core processors, firms have to redesign their applications to be multithreaded. UBS has achieved significant performance improvements by running ordinary single-threaded, C++ pricing and analytic applications on the new servers, he says. Out of the box, with no application changes, Ross says, the firm saw an average 70% performance increase for equities and a 250% gain in fixed income applications, with no increase in power consumption. The Nehalem features that provide these gains, he said, are on-chip memory controllers and hyperthreading.

"We want to continue scaling without parallelizing," Ross said, noting that parallelization requires a special skillset and is expensive. "My peers tell me, 'Scott, you're in denial, you should have a team parallelizing your analytic libraries, what are you doing?' Historically, we've scaled easily and efficiently through grid technology and process replication. We're not in denial, but we're not going to do it until the ROI is there. It's a very expensive endeavor, and with Nehalem, we've been able to continue scaling in a cost-efficient way."

Ross appeared committed to Intel's latest technology. "For the majority of our high performance computing cases, standard X86 hardware provides the best ROI and Nehalem is the platform of choice for this strategy," he said. The firm is replacing old servers with Nehalem machines, for HPC and more prosaic computing, and getting compression ratios of 10 to one.

Terry Roche, global head of information management systems at Thomson Reuters, told the group that Thomson Reuters has been reworking its Reuters Market Data System to run in an accelerated way on Nehalem servers without changing APIs. In tests, the data provider has been able to send out 22 million updates per second over 100 connections. This re-architected RMDS will be available early in the third quarter of this year.

Don Tyson, CTO of Sungard Ambit and Corporate Banking, reported that Sungard has benchmarked applications in the areas of credit and operations risk, interest rate, foreign exchange, asset liability management, and matching and reconciliation solutions on the new chip versus the last one (the Xeon 5400, Harpertown platform). The software provider has seen 30% to 64X performance improvements across a broad spectrum of applications.

Specifically, Sungard put its asset liability management software (which analyzes balance sheets and performs risk simulations), on 12 Nehalem servers in a grid alongside a 7400 database server and a one-gigabit NIC card. The result was a 30% performance improvement with no source code changes. With the help of Intel engineers, it reengineered its Ambit matching and reconciliation software to be more efficient and installed it on a 5500-powered app server alongside a database on a 7400 server, a one-gigabit NIC and Intel solid state drives and was able to perform seven million transactions in 1.5 minutes, down from 96 minutes on the Harpertown server, a 64X improvement.