PALO ALTO, Calif. In a renewed bid to build a market for Java-based chips, Sun Microsystems Inc. is poised to launch a powerful family of devices called MAJC (for Microprocessor Architecture for Java Computing; pronounced "Magic") this month at the elite Hot Chips conference. The architecture reflects a fresh approach from Sun toward Java silicon for a broad embedded arena spanning handheld devices, set-top boxes and back-end communications gear.
The move comes three years after Sun tipped plans to field a trio of Java-specific microprocessors: picoJava, microJava and UltraJava. In the interim, picoJava has sagged in the marketplace while microJava silicon appears unlikely to move into wide-scale production.
However, Sun is looking for MAJC to change its fortunes. The device is built around a powerful UltraJava core that executes 128-bit-wide very-long-instruction-word (VLIW) instructions. MAJC also incorporates digital signal processing capability and is adept at handling gigabytes per second of streaming data, according to the project's lead architect, Marc Tremblay, a distinguished engineer at Sun.
Equally important, MAJC is scalable, so that implementations can range from single- or dual-processor embedded versions to full-blown multiprocessing engines that contain all the cores on a single die.
In many ways, MAJC is aimed at the new age of streaming data, where gigabytes of audio travel over the Internet to appliances such as Web phones and handheld computers. Indeed, Sun earlier this year introduced Jini, a distributed-computing software technology, to enable such devices to communicate with one another. MAJC could work in concert with Jini, Sun executives said. However, they are loath to call it a "media chip," likely because the phrase conjures up images of industry failures such as Chromatic Research's Mpact.
"It's not a media processor," said Jeff O'Neal, group marketing manager for MAJC at Sun. "It's a general-purpose [device] for a new class of applications where you see the convergence of various data types. We think of it as a media-rich environment."
Lead architect Tremblay points to electronic commerce as the sweet spot for the architecture. "The type of computing people are going to be doing [moving forward] will be quite different. It will involve e-commerce as well as carrying video information for e-commerce. So we're betting that high-data-rate broadband in a streaming fashion is going to happen."
The first actual MAJC devices are expected sometime next year: a low-end embedded offering selling for around $50, according to industry sources. A dual-core MAJC processor would run at 500 MHz and sell for about $250.
The latter implementation could deliver as much as 6 gigaflops of performance, according to Will Strauss, president of market-research firm Forward Concepts (Tempe, Ariz.). "This is very, very powerful," he said. "On paper, it looks really good."
Nathan Brookwood, principal analyst for Insight64 (Saratoga, Calif.), agreed that MAJC is "very innovative." With die shrinks and improved designs continuing to increase the number of transistors that can be squeezed onto a chip, Brookwood added that MAJC's emphasis on parallel processing is a good way to use effectively all the real estate that is becoming available on a single piece of silicon.
MAJC is most different from Sun's previous Java chip ventures in that it will execute compiled Java. This boosts performance over the traditional interpreted-byte-code approach. Hot-Spot optimization technology in the compiler will further improve run-time performance, Sun said. And MAJC can run applications compiled into its native instruction set from other languages, such as C or C++.
In building the device, Tremblay and his team took a big gamble. "We got lucky to a certain extent," Tremblay said. "At the time we started, four years ago, Java was a month old. We made a bet that it was going to take off." Today, Tremblay said that "the size of the MAJC design team is three digits."
In terms of inside technical details, Tremblay and the company played it close to the vest. Sun will disclose details and firm product plans for the chip in October at the Microprocessor Forum.
However, it is thought that the 128-bit-wide VLIW instruction word can hold up to four op codes. Sun sources also hinted that MAJC will implement multithreading, perhaps in a way that alleviates the current interrupt-handling concerns caused by Java's garbage-collection (that is, memory cleanup) purposes.
On the marketing front, O'Neal said MAJC designs are not compatible with Sun's Sparc/Solaris architecture, which indicates where the company sees the new designs being used. O'Neal divides the world of embedded, networked applications into three tiers: the end-user applications, which could range from a set-top box to a PDA or a third-generation wireless phone; the routers and hubs that link these devices to the network; and the powerhouse servers that pump data to and from these portals and across the network. With Solaris-based designs entrenched in the final category, O'Neal sees the first two categories as the logical target market for MAJC.
Both Brookwood and Strauss agreed that the design could see use in a wide variety of applications, but that its DSP-like qualities make high-end routers and hubs a likely early use.
"This is intended to be a real-time RISC engine, and its prime function is at the gateway to a network, for routing and switching," said Strauss. "This is something that the Ciscos and Nortels and Bay Networks of the world are going to look at very closely."
However, Sun is not currently a player in that market. While Strauss noted that the chip offers some significant advantages in datacom switching, there is no infrastructure for a brand-new MAJC architecture and it will take some effort to convince design teams to consider it. "Sun is moving now into markets where they haven't been before," Strauss said. "This chip will have a lot of horsepower, but I've seen powerful chips fail before."
MAJC is Sun's boldest attempt at Java silicon since it launched its picoJava architecture which has been quietly withering on the vine in 1996. While South Korea's LG Semicon disclosed last month that it has working silicon samples using the picoJava II core, LG is one of just a handful of companies to actually implement the design in silicon.
Earlier this year, Sun shifted its marketing strategy to promote wider use of the design by offering it to potential licensees with no up-front charges. Sun claims more than 1,000 users have downloaded the specifications, but so far has not confirmed whether it has received any royalties from completed products.
Tom Halfhill, embedded-processor analyst with Cahners Microdesign Resources (Sunnyvale, Calif.), said the picoJava design was aimed at the desktop and back-end processing segments, and used too much power to go into handhelds. However, "there's not much need for a Java chip in the back end," he said. "The picoJava just wasn't designed for the kind of products where people really wanted to use Java. It would make more sense to have an architecture designed for low-power, battery-powered devices. That's where I think they should have taken their Java chips."