Silicon Insider

by Jim Turley
Silicon Insider #39, June 2006

After more than 20 good, long years Intel has reached the end-of-life for a number of its x86 microprocessor chips. Within a few months Intel will no longer produce the '486, '386, or '186 processor families, ending a long line of "x86" part numbers. Henceforth the Pentium will be Intel's entry-level processor for embedded systems.

The cuts go deeper than that; the company is also discontinuing its venerable 8051 processor family, the unloved i960 family, and the MCS251 and MCS90 families. In all, Intel is stopping production of some of the most popular 8-bit, 16-bit, and 32-bit chips ever made. Several thousand embedded designers are going to be very unhappy as they scramble for alternate sources for their favorite chips.

Still, time marches on and these processors were simply too old to be profitable any more. Ironically, it was Moore's Law that led to their demise. The '486, '386, 8051 and others were designed years ago for now-obsolete manufacturing processes. Although Intel has updated these chips' production lines in the past, they'd reached the point of diminishing returns and it is no longer financially attractive to upgrade them again. Instead, Intel gave its customers one year of notice: place your last orders or forever hold your peace.

Although we, and several thousand startled customers, will be sorry to see these chips go away, the end was inevitable. The question now is how and where customers will find alternate sources. Low-end alternatives for the 8051 and MCS90 chips won't be hard to find but the i960, '386 and '486 will be very difficult to replace.

Comment on this article

by Jim Turley
Silicon Insider #39, June 2006

In a somewhat related announcement, the #2 PC processor maker has also pulled the plug on some of its embedded microprocessors. Rather than simply killing them off, however, AMD sold the product line to Raza Microelectronics. In this case, the Alchemy line of embedded processors will become Raza property.

Alchemy was originally a standalone company formed by the diaspora from Digital Semiconductor when that firm was carved up and given to Intel and Compaq (now Hewlett Packard). Many of the designers of Digital's amazing StrongARM processor left to form Alchemy, producing a "StrongMIPS" processor under the Alchemy name. That company was later acquired by AMD.

AMD's acquisition made sense (at the time) because it gave the company an alternative 32-bit architecture, much as Intel's acquisition of the Digital design team gave that company a back-up plan. Yet neither scheme has worked out as planned. Intel's XScale (its name for the Digital-derived products) has never been very popular, nor has AMD's Alchemy product line been a big hit. AMD has now spun off Alchemy and Intel is rumored to be entertaining offers for XScale as well.

Diversity is a good thing but when your company is focused on producing PC processors for the masses, nothing else can get in the way. As large and lucrative as the embedded microprocessor business can be, it's simply a distraction for these companies.

Comment on this article

by Jim Turley
Silicon Insider #31, October 2005

Silicon Valley startup Inapac has developed a DRAM design that's designed specifically for OEMs building multi-chip modules. The DRAMs are not sold in conventional packages but are distributed only as bare die. As such, the devices are only useful to fairly advanced customers that are comfortable designing and building system-in-package (SiP) products. SiP's are popular for small, space-constrained applications such as cell phones, handheld games, e-mail terminals, and the like.

Inapac's distinction is in reliability and testing. Its tag line, "Has your known good die died?" underscores the company's emphasis on dependability. Inapac claims lower defect rates than the traditional DRAM vendors (Samsung, Infineon, etc.) because of its proprietary testing regime.

Although DRAMs are typically treated as commodities, and therefore subject to wide price fluctuations, Inapac's prices can remain fairly stable. Because the company does not compete in the commodity DRAM market, it's not subject to the wild supply/demand curves that afflict other DRAM makers. Instead, its devices appeal only to SiP or multi-chip module customers, a much smaller and less volatile market.

Inapac is run as an IP licensing firm, not a fabless chip company. Customers take out a license to Inapac's DRAM design and its companion testing technology, then have the chips built elsewhere. Currently there's only one place to go: the little-known ProMOS fab in Taiwan is the only foundry licensed to produce Inapac's devices. There's no fundamental reason why Inapac could not license more foundries in the future, but ProMOS is the only one so far.

We think there are advantages to multi-chip modules in general, and the practice of bundling two or more silicon die into one package is already more widespread than most observers are aware. Almost all multi-chip, flip-chip, or SiP products include a DRAM, so Inapac's prospects would appear bright. On the flip side (so to speak), the company has only a very brief track record with little history to assure prospective clients. Moreover, because Inapac only licenses the IP, it doesn't takes responsibility for the finished product. These two factors may weigh against the company when customers are looking for comfort as much as for technology.

Comment on this article

by Jim Turley
Silicon Insider #31, October 2005

The company that practically gave its name to an industry has expanded its line of little low-cost devices even further. The PIC24F and '24H family of 16-bit microcontrollers fills in a gap that few knew existed. The PIC24 series is a step up from the existing PIC18 family, with more peripherals and a faster processor core. Apart from that, it's pure Microchip.

Both the PIC24F and '24H share the same processor core with a 16-bit ALU, 16x16 register file, 16x16 hardware multiplier, and other familiar features. The '24F and '24H differ only in speed (16 MHz versus 40 MHz), memory capacity, and peripheral mix. Both families include the usual assortment of timers, A/D converters, UARTs, and miscellaneous I/O. The '24H adds more (and better) A/D converters, a CAN interface, and a DMA to baby sit it all.

In a classic Microchip move, the new chips are pin-compatible with one another. Aggressive multiplexing allows all the same features to fit in three different package types; customers will have to balance board space against accessibility.

Naturally, Microchip claims the PIC24 chips are the fastest 16-bit MCUs available, and initial benchmarks seem to bear this out. Most instructions execute in 1 or 2 clock cycles, whereas MCUs based on older architectures may require 2-8 cycles per instruction while running at similar clock speeds. Microchip's "near RISC" architecture thus gives it an edge in performance, and the company's enormous product line gives it the breadth and depth many customers need. Prices start at $4.50 and shipments are slated to begin in 2Q06.

Comment on this article

by Jim Turley
Silicon Insider #27, June 2005

After a few false starts and several rumors, Apple Computer CEO Steve Jobs announced this week that future Macintosh computers will use processor chips from Intel, ending a ten-year relationship with IBM and Freescale for PowerPC processors.

That doesn't mean future Macs will be PC-compatible, only that they'll use the same central processor chip that Windows PCs already use. Like Chevrolet making cars with Toyota engines (which does indeed happen) it's a bit of a jolt to purists but doesn't mean the two product lines will converge.

Nor is the agreement an indictment of PowerPC's performance. Today's top-of-the-line PowerPC processors are about as fast as Intel's best chips. Macs weren't suffering due to lack of speed. Instead, it marks a tiny move toward the mainstream as Apple tries to piggyback on Intel's development work; the company becomes a customer rather than a partner, as it was with IBM.

Apple's move gives IBM a PR black eye but won't change the latter's fortunes much. Already the majority of IBM's PowerPC chips are sold into embedded applications, not Macs. Indeed, Apple was only a small part of overall PowerPC sales. With IBM's recent design-ins for PlayStation 3, XBox 360, and Nintendo Revolution the company has made a clean sweep of the next-generation video game console market, which should soften the blow considerably.

Apple has switched processors before. The company moved from Motorola's 68000 family to the PowerPC family a decade ago, and the change went quite smoothly. The key to that migration was software-translation technology that allowed old Mac programs to run on the new machines. Apple will again provide translation software, this time including technology from Transitive Systems (Silicon-Insider #19, October 2004).

Although Macintosh aficionados -- who have lost all sense of perspective -- will curse Steve Jobs for embracing the Evil Empire, the move is a good one. Apple will continue to focus on its software development and still develop its own proprietary Macintosh computers. Macs themselves will be easer to market because they'll have the same "speeds and feeds" as Windows PCs, so naive users (which is to say, all of them) can compare clock speed and memory sizes with some justification. Apple will continue to defend its 3% market share. It'll just have fewer excuses for not growing.

Comment on this article

by Jim Turley
Silicon Insider #27, June 2005

Never one to miss a trend, MIPS Technologies has added a DSP coprocessor to its line of microprocessor IP. The newly minted 24KE is based on the existing (and popular) 24K processor; the "E" stands for "enhancement."

The most remarkable feature of the new processor core is that it took so long to happen. All the other significant processor-IP vendors added DSP or DSP-like extensions to their processors years ago. MIPS has been virtually alone, defending the true RISC philosophy to the last. The company has often promoted the "more software, less hardware" approach. But 32-bit processor cores are so small now that additional features take up virtually no silicon space. MIPS still has no Java acceleration hardware, instead touting the virtues of all-software Java implementations.

The technical features of the 24KE core are unremarkable; it includes the usual features to make it competitive, including new instructions and new registers. It consumes just a few square millimeters of silicon, an insignificant amount for most ASIC designs. To its credit, MIPS provides a good software base of DSP functions for the 24KE so that customers won't have to write all their own code. It's that code, and not the details of the processor, that influences engineers today.

Comment on this article

by Jim Turley
Silicon Insider #23, February 2005

After literally years of rumor and speculation, IBM, Sony, and Toshiba finally ended their industrial strip-tease and let the world catch a glimpse of Cell. As expected, the chip family will be impressive. What's more surprising is that much of this has already been done before.

First off, "cell" is still just a code name for a family of microprocessor chips jointly developed by IBM, Sony, and Toshiba. The actual chips are as yet unnamed. Cell chips are not intended for PCs, so they're not a threat to Intel (at least, not yet and not directly). Instead, Cell chips are intended for home electronics, specifically video games, televisions, set-top boxes, and other TV-connected gadgets. That's not to say Cell is a toy -- far from it. It just demonstrates that today's toughest computing problems are in the living room, not in the missile bunker.

The initial Cell processor has an IBM Power processor at its core, surrounded by eight accelerator processors. The central processor is a 64-bitter that can execute two instructions at a time. That makes it nice, but not ground-breaking. Intel, AMD, and others also produce 64-bit dual-issue processors. Note that Cell's central processor is a Power, not PowerPC, processor. In other words, it's the original IBM architecture, not the watered-down version seen in Macintoshes and embedded systems.

The eight accelerators are all 128-bit machines, which allows them to handle either very large numbers, or several smaller numbers, at once. In graphics applications, such "vector accelerators" are quite common and mandatory for achieving peak performance. The whole bundle of nine processors is joined by several high-speed internal buses. Outside buses to memory and I/O are provided by Rambus.

The chip runs a bit faster than 4 GHz in the lab, or a notch higher than current Pentium 4 chips. Again, nice, but a bit disappointing given that Cell is still far away from production while Pentium 4s are mass-produced parts. Cell is designed to run from a nominal 1V power supply and get by with just air cooling (perhaps with a fan), avoiding costly liquid cooling.

Future Cell-family chips will certainly be more impressive, showcasing the design talents of the three companies involved. It's no threat to the PC business, but it's not meant to be. Cell is intended to invade living rooms, where it will likely debut in Sony's PlayStation 3 some time next year. From there, it will likely creep into Toshiba flat-screen TVs and DVD recorders, then into IBM equipment. Cell's details aside, it's an impressive demonstration of just how important video entertainment has become when the likes of IBM, Toshiba, and Sony expend this much effort on electronic toys.

Comment on this article

by Jim Turley
Silicon Insider #23, February 2005

Loyal readers will recall (Silicon-Insider #19, October 2004) that Xilinx's EasyPath chips offered FPGA users a cheap alternative to "hard" ASICs. Xilinx customers with fixed designs destined for mass production could essentially buy Xilinx's scrap chips at a discount. Now arch rival Altera offers a similar upgrade, though the company follows a completely different path.

Rather than peddle scrap chips, Altera developed an entirely new chip for its high-volume customers. Called HardCopy II (interstitial capitalization seems to be required in this market) the chips are smaller -- and therefore cheaper -- than Altera's conventional FPGAs. In fact, HardCopy II chips are nothing at all like their FPGA cousins.

Unlike Xilinx's approach where FPGA designs drop right onto EasyPath alternatives, Altera users have to do a bit of redesign before migrating to HardCopy II. This redesign is all automated and quite straightforward, but it does change the user's design a bit. Converted designs get faster, for instance, which is generally reckoned A Good Thing. Altera claims some designs get twice as fast, again a bonus for most customers. Those users who don't want performance to change can insert delay buffers as required to match pre-conversion speeds.

After the design and post-conversion touch-up are complete, customers send their design files to Altera for production. Unlike FPGA devices, HardCopy II chips are customized at Altera's fab, where the top three metal layers are added. In essence, Altera is now in the business of small-run structured ASICs, quite apart from its volume business of FPGAs.

Altera's "hard" alternative clearly offers better performance than Xilinx does, which most customers would find attractive. On the other hand, Xilinx's approach has no risk; the new chips operate exactly like the old chips. Both paths are single-sourced, so neither breaks the grip these two firms have on their customers.

Comment on this article