Youve seen the TV auto ads that warn, “Professional driver, closed course. Do not attempt.” They remind me of microprocessor companies dueling over which has the fastest chip. The benchmarks they cite recall the empty city streets, deserted mountain roads and other venues beloved of ad directors but rarely encountered in day-to-day driving.
Perhaps “do not attempt” was the warning that should have been given to Hewlett-Packard recently before it tried to corner on two wheels by announcing its HP Workstation xw4100 with an Intel 3GHz Pentium 4 CPU and an 875P chip set. Later on the day of the announcement, Intel admitted that this combination had shown “anomalies” in behavior (that have since been resolved).
One had to wonder which users had been impatiently awaiting this desktop dragster; even on the most intensive applications, HP predicts less than a one-fourth reduction in task times, with only “up to 5 percent” improvement in mainstream applications. To put these figures in automotive terms, the 0-to-60 time could fall from 7 seconds to 5.5, but the 15-minute drive to the store would lose only 45 seconds.
In the world that we see over our steering wheels, that sounds like the kind of improvement that we could get by learning to avoid the slow lanes and to time the traffic lights. Its less than I would want from putting a turbocharger on my minivan and is less than you should get from putting new workstations on the desks of our engineers or portfolio analysts.
In contrast, one processor chip has achieved a noteworthy track record for real-world performance: Im talking about IBMs Power4 processor. Its design reflects not mere pursuit of benchmark numbers but throughput for next-generation tasks.
The two CPU cores on a single Power4 chip share more than 100GB per second of bandwidth to their shared L2 cache and more than 55GB per second to memory. Personally, I overflow on numbers that big; perhaps, like me, youll find it easier to think of 55GB as 26 hours of DVD video—a stupefying amount of data. And the IBM eServer p655 series, shipped late last year, can hold four- and eight-processor blocks that stack up to 128 Power4 processors per frame.
Page 2
But even this impressive capacity and scalability are not the most remarkable parts of the Power4 story. When Microprocessor Report honored the Power4 chip design last year, as its choice for “Best Workstation/Server Processor,” it went beyond the usual plaudits for throughput and scalability to note that IBM “actually met shipment schedules publicized two years ago.”
We all get a kick, Im sure, out of the “actually” in that quotation; its as if no one ever imagined that such a thing could happen. Indeed, the time its taken to fulfill the promises of Intels IA-64 Itanium—if “fulfill” is yet the word—and the delays of AMDs x86-64 Opteron, which finally shipped last week, must make many of us wonder who can haul our real loads on real roads.
For those of us engaged in high-performance computing projects, with schedules on the walls that cover everything from laying the foundation for the building to making sure that the software will work, it would be nice to feel confident that the processor chips will be there to plug in when we need them.
When you look closely at your enterprise IT architecture, what do you perceive as the speed bumps on the road to productivity? Do you see CPUs in need of a power boost? What I see is a need for higher system throughput: for encryption accelerators for VPN performance, XML accelerators for Web services transactions, solid-state hard disks and in-memory databases for interactive data analysis tasks, and graphics accelerators for image processing and 3-D visualization.
We have excellent platforms now for tasks of every size, and their opportunities for improvement are greater at the edge than at the core. Lets challenge system builders to give us better-balanced designs and reward chip makers for meeting real needs today—not for telling us what we cant attempt until tomorrow. Doing anything else encourages chip makers to achieve their goals, not ours.