November 06, 1998

G3 upgrade cards unstable?

By Cameron Crotty

Installing a G3 upgrade card in an older Mac seems like a great way to get high performance for a low price, but there may be a tarnish on the silver lining: An incompatibility between the PowerPC 750 processor and the ROMs in PowerPC 604-based systems can create serious system instability, and the consequences range from nonfunctioning I/O cards (e.g. SCSI, video) to complete system crashes. In extreme cases, the instability may damage or destroy the data stored on a user's hard drive.

According to industry sources, the problem is difficult to detect and isolate, but several accelerator card manufacturers, including Newer Technology Inc., PowerLogix R&D; Inc., and Sonnet Technologies Inc., confirmed that the incompatibility exists.

Jon Fitch, formerly vice president of engineering at the now-defunct Power Computing Corp., said Power also discovered the problem while his team was developing PowerPC 750-based Mac clones. "It took us a long time to track the problem down," Fitch said. "We had 30 to 40 machines up running around the clock, and we were gettting a 100 percent failure rate."

The systems took as little as one hour to fail or as long as two days, Fitch said. Power developed a fix for the problem, but Apple shut down the clone market soon afterward, and the point became moot.

The problem occurs because of "predictive" or "speculative" processing, a feature found on most modern CPUs. Instead of simply executing commands in the order received, the processor looks ahead in the instruction queue and tries to predict which way the code will branch. As part of this operation, the processor peeks at various memory locations throughout the system, including RAM, ROM and I/O memory.

Unfortunately, while RAM and ROM are designed to be read multiple times, certain parts of I/O memory cannot be accessed without altering them. For instance, some registers automatically reset to their initial state after being read. This silicon demonstration of the Heisenberg principle -- where the act of observing an event alters it -- would have no bad side effects if the processor immediately acted on the data that it obtained. But the processor is computing predictively -- ahead of the actual code flow -- and thus frequently doesn't execute instructions in strict queue order. Thus, it's possible for the processor to change the value of an I/O memory location (simply by accessing it) without the software being aware of it. The result is a creeping corruption that can prevent cards from functioning, completely crash the system or introduce errors in the data stream-errors with side effects as benign as a video glitch or as malignant as a corrupted hard drive.

Opinions vary among the accelerator card companies that acknowledge the incompatibility. Newer Technology said it has known about the problem since it started working with the PowerPC 750 processors and has developed a hardware fix -- one that requires about $30 in silicon per upgrade card. "If you're not using a Newer Technology upgrade card, you're running a serious risk of losing data," said Darryl Hinshaw, Newer's vice president of engineering.

Sonnet Technologies agreed there is a problem but said it has a fix in the form of an extension that loads at startup and tweaks a bit in the memory management unit to prevent the processor from incorrectly accessing the I/O memory. Newer disputed Sonnet's claim, saying that fixing the problem when extensions load leaves the user open to damage during the early parts of the boot process.

A Sonnet spokesman further claimed that peripheral card manufacturers can prevent errors from occuring by including synchronization software in their cards' firmware.

PowerLogix said it also recognizes the problem, but it insisted that, outside of a few "notorious" cases, the average end user is extremely unlikely to run into difficulty. Mark Reviel, PowerLogix's vice president of engineering, said the benefits of fixing the problem would be minor compared to the cost and delay involved in engineering a solution. "These problems are nonexistent in G3 Macs," he said. "We may go back and address it if the market gets stronger, but right now we're more interested in the future."

On the other hand, Sean Mohle, chief technical officer of Mactell Corp., rejected any reports of system-level incompatibilities. He chalked up the alleged problem to longstanding reports surrounding incompatibilities with SCSI cards from Adaptec Inc. Mohle further asserted that the reason Newer Technologies' cards work with Adaptec products is that "Adaptec entered into a nondisclosure agreement with Newer" so the companies could share development information -- a relationship not shared by other upgrade card vendors.

Adaptec could not be reached for comment for this article, but industry sources said the company is preparing a white paper on incompatibility issues that will lay blame squarely on the problems with the G3 chip and 604 ROMs.

Dantz Development, another company whose product (Retrospect Remote) has been affected by upgrade card compatibilities, has gone even further. Dantz said that it set up a Power Mac 7500/100 (with 32 Mbytes of RAM and running Mac OS 8.1/8.5) and a Power Mac 9650/350 (with 192 Mbytes of RAM and running Mac OS 8.1/8.5), each equipped with two different tape drives. Dantz then tested accelerator cards from PowerLogix, XLR8 and Newer Technology. According to Dantz Vice President of Marketing Craig Isaacs, "Newer's cards worked. The others didn't." Isaacs said that while running the XLR8 and PowerLogix cards, "At between 50 Mbytes and 500 Mbytes of copying files, the computer freezes. It doesn't happen every time, but it is extremely reproducible" -- four out of five times.

Isaacs also said Dantz tested the PowerLogix and XLR8 cards with Iomega Zip and Jaz removable drives. The Macs didn't freeze up, but in roughly 10 percent of the trials Retrospect Remote reported data corruption. "In other words, either the file backed up is corrupted or the comparison is corrupted, but either way there's something wrong," Isaacs said.

Dantz said it is completing tests on cards from Sonnet Technologies and plans to publicly release all results within a week or two.

Copyright © 1998 Mac Publishing LLC. All rights reserved. Reproduction in whole or in part in any form or medium without express written permission is prohibited. MacWEEK and the MacWEEK logo are trademarks of Mac Publishing LLC.

Feedback Page