Russian supercomputer architect Dr Boris Babaian, often referred to as the Seymour Cray of Russia is planning to challenge Intel Corp with technology that uses similar techniques to Intel’s Merced chip. In recent interviews with The Microprocessor Report, and with our sister monthly magazine Computer Business Review, Babaian revealed that he has his sights keenly on the privatized future of his company Elbrus International. The Elbrus-1 and Elbrus-2 supercomputers the firm developed were a mainstay of the Soviet military and parts of the public sector. But Elbrus-3, reputed to have been twice as fast as the Western competition, never made it to market because of the collapse of the Soviet state. In that machine, Babaian claims to have pioneered a number of techniques that are only now beginning to be exploited by Western firms, including the use of VLIW very long instruction word techniques that are a feature of Merced. Babaian is now proposing to utilize those techniques, among others, to make an Intel-killer microprocessor he calls the E2k that can emulate both the Intel x86 and its next generation IA-64 instruction set in silicon, while outperforming anything that Intel can offer. Using a 0.18 micron process and a clock speed of 1.2GHz, the chip will run at 135 SPECint95 and 350 SPECfp95, says Babaian, and use only 126 mm square of silicon and 35 watts of power. The Microprocessor Report estimates that, using the same process, Merced would clock at 800MHz and deliver 45 SPECint95 and 70 SPECfp95, on 300mm square of silicon and 60 watts of power. Merced, however, is thought to be a year ahead of the E2k in the development cycle. The key, says Babaian is binary compilation, a technique most visibly used to date by Digital Equipment Corp in its FZ32 emulation product, which runs Intel x86 programs on the Alpha. But the Elbrus binary compiler will be built into the silicon rather than offered as a software add-on. Binary compilation is also the technique that secretive Silicon Valley start-up Transmeta Inc is said to be working on, and intriguingly, Transmeta CEO Dave Ditzel worked with Elbrus for several years while he was at Sun Microsystems. The E2k is currently only implemented in an executable Verilog database, and to build a prototype, Babaian needs substantial investment – as much as $60m. Until now the project has been funded by Elbrus’s maintenance and systems integration division, and by work Babaian and his team have been carrying out for Western companies such as Sun Microsystems Inc and Avant! Inc. The search for capital is being managed by San Francisco-based Plantagenet Capital Management, which says that two European semiconductor giants, Germany’s Siemens AG and the Franco-Italian STMicroelectronics, are both showing an interest. Among other things, the money will buy access to Western semiconductor facilities. There are problems, however. Accepting investments from a Western partner will mean the company will have to be more open about its finances, which could provoke a visit from Russia’s notorious tax police. And remembering its only the paranoid survive motto, it’s likely that the next midnight knock on the door will come from Intel Corp’s lawyers.
