Billions and billions of transistors

At In-Stat/MDR's recent Microprocessor Forum 2002, Intel Fellow John Crawford described how high-end microprocessors built within the next five years will contain more than a billion transistors. This figure is a major milestone in our industry. Today's most complex CPUs have 200 million transistors or more, and take years to design. How will we design chips with far greater complexity? What will these gigachips look like, and when will their capabilities trickle down into mainstream applications?

Crawford said one way to simplify the process of designing monster microprocessors is by replicating elements of previous designs. In his presentation, Crawford showed a hypothetical descendent of Intel's Itanium processor family consisting of four CPU cores similar to the one in today's Itanium 2, plus a large shared cache for temporary storage of important data. This approach would create a chip with some 1.5 billion transistors, without requiring a dramatic redesign of the processor core.

I'm sure Intel will improve the Itanium core greatly between now and then, and we'll eventually see individual CPU cores with a billion transistors. However, simple replication can deliver big performance gains. Built in a 65-nanometer manufacturing process, Crawford's proposed chip-scale multiprocessor would run at 6GHz vs. today's 1GHz Itanium 2. Four cores running at this speed would be some 24 times faster than Itanium 2 on typical server applications.

Power consumption also would increase over current chips, but not nearly so much. Companies such as IBM already have developed the technology needed to remove hundreds of watts of heat from a microprocessor, so this issue won't impede the industry's progress. However, we should be prepared for complex, expensive thermal solutions such as liquid cooling.

Perhaps the more interesting question for most of us is this—when will gigachips appear in personal computers and consumer electronics we can buy? Manufacturers have been showing billion-transistor DRAMs for many years now, so the question isn't entirely speculative. By the end of this decade, we can expect to see affordable PCs with gigachip processors, multi-gigabyte memory arrays and multi-terabyte hard disks. Amazingly, similarly configured PDAs and smart TVs will arrive by the end of the next decade.

What benefits will such systems bring to customers? Though Moore's Law permits predictable numeric prognostication, it's harder to see how this technology will be used. Only multimedia data has a chance of consuming so much storage space, but simple recording and playback can't use up the processing performance offered by a chip such as the one Crawford described at MPF 2002.

By the end of this decade, we can expect to see affordable PCs with gigachip processors, multi-gigabyte memory arrays and multi-terabyte hard disks

To me, the most likely use for so much power is multimedia content analysis—not just speech recognition, although this will be important enough, but recognizing faces, places and text in video streams. Academic and commercial research labs already are working to develop software for these tasks; within 10 years, it'll be ready for the rest of us.

Think how much more useful your digital camera will be once your PC is able to sort through your picture collection to find all the photos of your kids. Think how much more useful your TV will be when it can produce show transcripts in real time and find every documentary and movie filmed at your next vacation destination.

Professional users will use future computers to cure diseases, design spacecraft and create multimedia works of art—but it's the more mundane applications that will create billion-customer markets for billion-transistor processors.

Peter N. Glaskowsky is Editor in Chief of In-Stat/MDR's Microprocessor Report. For more information on topics covered in this column, visithttp://www.chipadvisor.com. Send Peter e-mail atchipadvisor@ideaphile.com