View Single Post
Old 27 February 2012, 13:54   #2373
Join Date: Jun 2010
Location: out in the wild
Posts: 1,251
Originally Posted by Doc Mindie View Post
But anyways: If someone was adventurous anough to patch EXEC to give the resource tracking and stuff... how ahrd would that be? would all tools (like Avail) also need to be updated to be able to utilise the tracking system?
Isn't that a nice question to be discussed in the Coder's forum? Doesn't really have a relation to a specific accelerator.

Anyway, we investigated the reported crashes from some ACA1231 customers in the past few days and followed each hint systematically.

1. power supplies
Some power supplies deliver a slightly higher voltage than the 5.0V that the Amiga and accelerators are made for. We found that we could make the board crash faster if the voltage is higher. We have therefore continued our tests with a little more than 5% over-voltage of 5.3V.

2. ACAtune
The ACAtune tool had a bug that added memory to the system that does not exist. If this non-existent memory is accessed, the computer crashes immediately. This can take a few seconds, but it may also take half an hour, so the reason is not obvious. I have posted a new version here last week, and that's also available on the Vesalia site - please only use this latest version.

3. variation of boards
Vesalia has conducted a large test of all cards where they found that some cards work without any problems, and others crash pretty quickly. They have separated the cards into stability classes, where about two thirds of the cards worked without any problems, and about one third showed frequent crashes after warming up. We have continued our tests with these cards.

4. Power supplies and instable cards
The boards that were prone to crash were tested again with a lower voltage of 5.0V. The surprising result was that these cards remained unstable, only the time until they crash was longer. This shows that power supplies are not the root cause of the crashes.

5. cooling
We have now created an especially critical setup with a power supply that was set to 5.4V. One of the "quickly crashing" cards was equipped with a tiny 30mm by 30mm aluminium sheet, thickness was 4mm. This was a plain aluminium sheet with no cooling fins. To further reduce heat flow, the metal was placed on styrofoam (which is known to be insulating). This setup has run an endless LHA-loop over the whole weekend without any crash.

This test has shown that even with a minimal cooling sheet metal, one of the most unstable cards is taken to a reliability level that we had already reached during devlopment. Vesalia and I have therefore decided to add a cooler that's considerably larger than the one that we have successfully tested: With an area of 22.7 square centimeters it's more than twice as big as the one we've tested. In addition to that, we will use a high-tech heat transfer pad, which has much better properties than the glue I have used in the test setup. The cooler is mounted with two screws that may only be tightened by hand - no tools required, and even "no tools recommended", because these screws should not be tightened too hard.

All customers who already have an ACA1231 will get the cooler, the heat transfer pad and mounting material free of charge in a separate shipment from Vesalia. There is no need to request it - you will get the cooler automatically. New cards will be shipped with the cooler pre-installed.

We apologize to everyone who had temperature-related problems. Nobody of us has expected such a huge variation in CPU temperature sensitivity, especially as all these CPUs are from the same production batch. The extremely fast memory interface of the ACA1231 is obviously taking the processor to the edge of it's ratings: With "little to no waitstates", the CPU has a lot less time to cool down and thus produces more heat than any other 68030 accelerator did before. With older memory interfaces inserting up to five times as many wait states, temperatures on older cards are considerably lower.

Maybe this is the reason why Freescale is not listing the 68030FE40 any more - the fastest processor in CQuad-Pack that they're listing is 33MHz. However, a little added cooling (yes, very little) helps bringing the processors to the specifications that Motorola published before they became Frescale.

Schoenfeld is offline  
Page generated in 0.03893 seconds with 10 queries