Two startups seed Java silicon

Two startups seed Java silicon
By Alexander Wolfe, EE Times
October 18, 1999 (2:48 p.m. EST)
URL: http://www.eetimes.com/story/OEG19991018S0025

OAKDALE, Iowa — Two ambitious new Java-chip efforts could give the lie to the growing perception that Java-based silicon is dead.

Instead, the separate efforts under way at Ajile Systems Inc., here, and Advancel Logic Corp. (San Jose, Calif.) seem to indicate that architectural approaches other than Sun's picoJava core will be the way silicon architects will go if Java is to succeed in the embedded world.

For its part, Sun Microsystems Inc. said its Java silicon work is proceeding apace. "We're broadening our line with MAJC [a new, high-end networking processor] and our picoJava licensees are going full-steam ahead," said Fadi Azhari, marketing manager for Java silicon at Sun. "The fact that other companies are doing Java processors only proves that it's a good idea and they're needed." For example, South Korea's LG/Hyundai is offering embedded chips based on Sun's picoJa va 2 core.

Nevertheless, both the Ajile and Advancel tacks stand in stark contrast to Sun's work. Each of the startups claims to lick the performance problems that have dogged Java in the embedded world by eschewing the traditional interpretive-execution model used in most Java silicon. Instead, each is going directly to native execution.

Ajile, which was spun out of Rockwell Collins in July 1999, bases its work on the 20 years of experience the startup's engineers have had with stack-based computing — a technique they claim fits Java to a tee. The company's AJ100 Java processor features multiple Java virtual machines (JVMs) and an embedded real-time operating system on the chip itself. That dispenses with the need for a hefty external RTOS, because everything is done in microcode.

"We have the best hard real-time, period," said Ajile president George Hwang, a former co-founder of Integrated Device Technology. "Nobody comes close. We can run at 1-microsecond real-time context switching. Any body else is talking milliseconds."

Ajile arose out of the frustrations of a group of Rockwell Collins engineers who built the GEM1 Java engine in 1997. They tried to convince what was then called Rockwell Semiconductor Systems (now Conexant) to take the part commercial. However, that division chose to focus its efforts in the communications market. But GEM1 supporters, both inside and outside the company, kept the effort alive on the back burner.

To get Ajile going, the startup's engineers have signed an agreement with their former employer giving them rights to commercialize the technology for the merchant market. Rockwell also has an unspecified equity stake in Ajile. Hwang is also seeking venture-capital funding.

The second company, Advancel Logic, is pursuing a system-on-chip approach. The aim is to enable OEMs to pursue new possibilities for Java-based apps in the embedded world and beyond. For example, Advancel is already working on a device that will place its Java core alongside a digita l signal processor on a single die.

Advancel's unique twist is a Java engine that is actually a dual-language processor: It runs both the C programming language as well as Java natively. "This is a synthesizable design. We do it in Verilog," said Advancel president Anand Sheel. "There is no reason it cannot go in the 160-MHz range.

"What we have done is a less than 80,000-gate device, which does Java and C, and it has performance numbers that are in theory as good or better than some of the ARM and MIPS cores," said Sheel. "We are not sacrificing any performance and you are getting the advantage of Java."

Burrowing beneath the surface, each company highlights different advantages in their respective designs.

At Ajile, the bet is that the stack-based architecture the company is offering will be best suited to running Java. "The JVM spec is defined around a stack-based architecture, in the sense that the instruction set defines 'push' and 'pop' operations — like in an HP-style calculator, " said Nick Mykris, Ajile's vice president of engineering.

According to Mykris, "A stack architecture keeps code density very small, because of the use of implicit addressing." (In practice, implicit addressing means that an 'add' instruction will know that it should pop its two operands off the stack, rather than look them up in memory.)

Ever-longer programs are the bane of Java programmers, so an engine that reigns in software size requires less memory and is therefore cheaper to build into a system.

"We have a two-to-three-times code improvement over register-based machines," Mykris said.

Multithreading angle

Two other design twists should stand the AJ100 in good real-time stead. The first involves garbage collection, which Java uses to reclaim leftover memory space. That is useful for large programs but wreaks havoc with real-time apps. That's because a garbage collector can interrupt a program at will, taking as much time as it needs. Consequently, a programmer can never b e sure whether his code will exhibit deterministic performance by responding to real-time events on an as-needed basis.

Many Java designers are at work on elegant garbage-collection algorithms that minimize that problem. Ajile's creators took perhaps the quickest approach of all. "There is no garbage collector," said Mike Masters, CPU architect at Ajile. "We do all our memory allocation during initialization." Since the chip will be running a tightly bounded embedded app, as opposed to a lengthy enterprise program, memory overrun shouldn't crop up as a problem.

The second twist is the AJ100's implementation of multiple JVMs on one processor. That is, three or four separate, virtual JVMs can run on one device. Each JVM can handle one aspect of operation — the first, CRT control, the second, applet downloading — without interfering with one another.

That makes for a safer design. As Mykris put it, "If one JVM goes crazy, it won't affect the other ones. It bounds the limit of what each c an do."

But most important, according to Mykris and his colleagues, is the ability of the AJ100 to run without the need for a bulky real-time operating system on top of it. "We've implemented threading inside the microcode," explained Ajile's Masters. "So there's no traditional RTOS layer to worry about, and there's no JVM."

While the lack of an RTOS, or even a downsized real-time executive to perform mundane scheduling tasks, seems mighty unusual, Masters claimed that was part of the plan. "You don't need them," he said. "The threading primitives [low-level commands] — as well as instructions such as 'Yield,' and 'Notify' — are implemented directly in microcode. So the [AJ100] architecture knows how to work with its timers to set up delay queues."

That enables the processor to time-slice among the different JVMs as well as the multiple threads within a given JVM. The result, the company says, is deterministic operation that fits the bill for real-time requirements. It also enables OE Ms to write tightly wound applications that can be burned into ROM, to make a cheap, memory-constrained embedded board.

Ajile officials said the AJ100 is intended to compete as an ARM7-class machine and will be aimed at deeply embedded applications. Although Internet appliances are the holy grail, the first target market is factory-floor automation. To that end, Ajile is currently putting together a PC104 form-factor evaluation board and kit, outfitted with the AJ100.

Just as Ajile is updating the old concept of stack-based computing for the brave new Java world, Advancel Logic is similarly cognizant of the past.

In its case, Advancel is spinning a processor that taps emerging Java technology while it utilizes the tried-and-true C programming language. "I think people have underestimated the legacy of C," said Advancel president Sheel. "The world in our mind is a mixture of C and Java. The newer stuff is definitely being done in Java. But there is quite a bit of code and products in C."

Advan cel's core, therefore, is fitted with a lengthy complement of microcode so that it executes both Java and C natively. "It is a rather innovative dual-language processor," Sheel said. "In the non-native machines, the problem with Java in interpreted mode is that the system [memory] footprint requirement is too large. We execute Java byte codes natively."

According to Sheel, a JVM is not required, unless the application requires some dynamic loading. For such situations, a small JVM — the JVLite from Japan's Access — is available.

Advancel's offering is aimed at deeply embedded apps and dissipates less than 1 milliwatt per MHz. The company will roll out its first OEM design win next month in France, when STMicroelectronics unveils a smart card that uses the smallest member of Advancel's TinyJ Java family. That particular implementation runs Sun's JavaCard spec.

Aimed at the sweet spot of the embedded market is their TinyJ engine, which supports Sun's embedded and personal Java strains.

Sheel said he has just received samples of TinyJ silicon back from a fab in Japan, which he would not name. "The 0.35-micron device has been tested at 66 MHz. Now in the works is a 0.25-micron implementation which is expected to run at 100 MHz."

Next up is a TinyJ line extension that will include on-chip DSP functions. Sheel said Advancel is pursuing this internally as well as with an outside company, but declined to name that firm. Silicon is expected in the second quarter of 2000.

Perhaps the most aggressive aspect of Advancel's business plan is its strategy to offer the TinyJ engine as a system-on-chip core. "Most of the customers we have are adding stuff like USB," said Sheel. "Depending on what you put around the core, it's going to cost from $10 to $20 in quantity. It all depends on how much memory, whether you put flash around it. We're even thinking of putting an FPGA block in there, so there are some programmable gates available and also depending on memory."

However, both Ajile and Advancel see industrywide prices falling into the single digits (i.e., well below $10) once quantities get firmly ramped up.

Sheel vows that Advancel's engine will be in many products by mid-2000. "It's just that a lot of these companies don't want us to talk about it yet."