Embedded Processor Watch

Editor: Jim Turley

In This Issue:

• Editorial: What's the Best Way to Benchmark?
• Philips Exits the Handheld-PC Chip Business
• Industry Resources: Embedded Processor Forum '99
• Industry Resources: Mobile Insights
• New Embedded IC Announcements

Editorial: What's the Best Way to Benchmark?

A benchmark is like sex. Everybody wants it, is sure they know how to do it, but can't agree on how to compare performance.

Part of the problem lies in the fact that microprocessor performance is not a one-dimensional vector. Microprocessor drag racing is all very nice, but the average embedded designer is looking to balance the often contradictory aspects of power consumption, performance, code density, price, interrupt response, and probably others. A combination that's good for one application may be unusable for another.

Benchmarking embedded chips is tough, no doubt about it. That's why we still have Dhrystones, the accepted lowest-common denominator that any microprocessor can run. Unfortunately, Dhrystone tells us very little about what a microprocessor is good at. One could argue that Dhrystone scores say more about the marketing efforts behind a chip than about its technical features.

By now you've probably heard about the embedded benchmark work underway at EEMBC. (If you're not familiar with the embedded benchmark consortium, check out http://www.eembc.org.) EEMBC's laudable goal is to eradicate the scourge of Dhrystones in our lifetime. EEMBC counts 24 of the biggest CPU makers among its members and, for such a diverse group, they've made amazing progress toward standardizing embedded benchmarks. But there may still be some crumbs between the sheets.

Realizing that no single metric can hope to capture the many varied aspects of a chip's performance envelope, the EEMBC benchmark suite consists of dozens of smaller benchmarks. Each test contains a core algorithm taken from real-world code. There are tests for automotive engine control, codecs, pixel manipulation, task switching, and lots of others. All the tests have been written in ANSI C for portability.

EEMBC is following a path somewhat similar to that taken by SPEC (www.specbench.org), which is a good thing in my opinion. Specifically, EEMBC will allow its members to report two scores for every benchmark: the "out of the box" score, and the flat-out, fully tweaked, downhill- with-a-tailwind score. The double scores allow potential users of these chips to evaluate competing processors under controlled conditions (the basic scores) and to also see what each chip is fully capable of, given some care and attention.

Nobody disputes the need for controlled, nonnegotiable, standardized testing. But I expect some controversy over how best to handle the "tweaked" scores. Exactly how much tweaking is allowed, or desirable? Should testers be allowed to alter the source code of the benchmark? Can they rewrite key algorithms? Can they take shortcuts, like hard-coding lookup tables or--dare we say it--the final results of complex calculations?

The question boils down to deciding what is important to test and what is extraneous. If your goal is to pin down a given microprocessor's abilities in real-world situations, then I believe there should be no holds barred. Any optimization, from rewriting all the C code, to creating shortcuts, to using unusual chip-specific features or instructions, is fair game in my book. This approach encourages creative and unusual solutions, which are representative of the real world of embedded programming. As long as the benchmark delivers the correct answers in a reliable and repeatable manner, the details of generating the results shouldn't matter.

It's that "reliable and repeatable" part that makes people nervous. Obviously, simply hard-coding the answers to the benchmark after a lot of NOPs isn't meaningful. And here's where EEMBC's sister organization, the EEMBC Certification Labs (ECL), comes in. ECL must first approve all EEMBC scores before they can be published. Part of ECL's role is to prove that "tweaked" scores are arrived at by sincere means. That includes pumping alternative data sets into the chip under test, to be sure that it's really executing the correct algorithm and not just regurgitating prearranged answers.

To me, a benchmark should test the abilities of a chip, not its compiler or the skill of EEMBC's programmers. Wide-open testing promotes creativity and allows vendors to exploit the unusual features of their processors. As long as the chip returns the correct result under all conditions, I don't believe it matters what's inside the black box. Forcing a particular coding convention onto dozens of different microprocessors simply discourages programmer innovation and reduces everything to the lowest common denominator. And then we'd be right back to Dhrystone.

Philips Exits the Handheld-PC Chip Business

Philips Semiconductor has quietly made a strategic exit from the business of making microprocessors for handheld organizers, PDAs, and palm-size PCs. The company's "TwoChip PIC" processor (PR31700), used in Philips's own Velo and Nino Windows CE units, is effectively orphaned, with no offspring in sight.

Philips will focus its development efforts instead on the mixed-signal companion chips for the microprocessors, such as its successful UCB1100 and UCB1200 chips. These chips provide the analog/digital interface required for touch-screen, speaker, and modem functions common on handheld electronic organizers. Philips management felt that the company's line of MIPS-based processors provided little differentiation from similar devices from Toshiba and others. Indeed, Philips's short- lived series of PDA processors was based on Toshiba core designs.

Philips will remain active with its MIPS license (the company, like most others, also has an ARM license), building MIPS-based processors for TV set-top boxes and similar systems. Philips will also develop companion devices for standalone MIPS chips in the R5000 class. Windows CE-based palm-sized PCs have not been particularly successful in the market, with sales lagging far behind the leading units from 3Com (formerly Palm) and Psion. Dividing that market share among six or more WinCE makers only makes matters worse. It appears that Philips bowed to the inevitable, shunting its development staff to more distinct product lines, leaving the CPU development in more experienced hands.

Industry Resources: Embedded Processor Forum '99

Mark your calendars, make your travel reservations, and grab your business cards. The second annual Embedded Processor Forum is returning to San Jose the first week of May. This year's conference will last for four days (two days of conference plus two full-day seminars). As before, the emphasis is on first-time announcements of new microprocessors, live debate with industry luminaries, and a comfortable atmosphere in which to mingle with your fellow embedded professionals.

There are 20 new chip announcements planned, covering system-on-chip core designs, low-power microprocessors, high-performance CPUs for media processing, and new DSPs and DSP extensions. All the major vendors-- Motorola, IBM, Intel, Sun, Hitachi, and others--will be there, as well as some surprising newcomers.

For more information, visit http://www.MDRonline.com/events/epf99, or call MicroDesign Resources (Sebastopol, Calif.) at 707.824.4004. The early-registration discount expires at the end of this month (2/28), so don't wait too long to make those reservations.

Industry Resources: Mobile Insights

Journalists and analysts are going mobile at the Mobile Insights conference in Palm Springs, February 21-23. For more information visit http://www.mi99.com or call 800.640.6497.

New Embedded IC Announcements

CS5180, CS5181 (Crystal Semiconductor) Delta-sigma A/D converters have 16-bit precision, 93-dB SNR, and input rates from 8 kHz to 400 kHz ('80) or 625 kHz ('81). Price: $12.50/1,000; Samples: Now; Production: 1Q99; Call Crystal at 512.912.3351.

AD8309 (Analog Devices) 500-MHz, 100-dB logarithmic amplifier provides both logarithmic and limiter outputs for accurate phase detection in CDMA and GSM phones. Price: $7.75/1,000; Production: Now; Call ADI at 800.262.5643.

LM4560 (National Semiconductor) AC97 PCI audio accelerator integrates 64-voice wavetable synthesis, 32 simultaneous digital audio channels; with PCI bus power management. Price: $8/1,000; Production: Now; Call National at 800.272.9959.

CPC700 (IBM) Memory controller and PCI bridge for PowerPC 600-series and 700-series microprocessors handles 83-MHz bus, 66-MHz PCI; interfaces to SDRAM. Price: $48/10,000; Samples: Now; Production: 1Q99.

YMF738 (Yamaha) Digital-audio controller chip for PCI provides integrated modem and AC97 Link functions, for audio/modem riser (AMR) cards; in 128-pin package. Price: $10/10,000; Samples: Now; Production: 1Q99; Call Yamaha at 408.467.2300.

ZPSD813F1V (WSI) Support-logic chip for microcontrollers adds flash memory, EEPROM, SRAM, 3,000 gates of CPLD, and programmable MCU interface. Price: $8.95/25,000; Production: Now; Call WSI at 510.656.5400.