The Cape Cod Disaster

Here’s a motherboard Intel very quickly wanted to forget about:

Intel Desktop Board CC820 with CPU and SDRAM, early 2000
Intel CC820 Desktop Board (early 2000)

It’s the Intel CC820—or Cape Cod—desktop board, a product that was late to market (not unusual) and within a few months, the subject of a recall (quite unusual). As the CC820 designation suggests, the board was built on the ill-fated Intel 820 ‘Camino’ chipset.

The Camino chipset was supposed to be released roughly in mid-1999 as a replacement of the workhorse 440BX chipset for mainstream desktops. At the same time, Intel changed how it segmented the market. The 440BX supported up to two processors, while the 820 didn’t (though the 820DP variant did); the 820 chipset was targeted for typical desktops, while the Intel 840 chipset was meant for high-end workstations with two processors. Note that the Intel 810 chipset was meant for “value” PCs, and played that role quite successfully.

The 440BX chipset was limited to 100 MHz FSB; the 820 supported 100 and 133 MHz FSB for the then-new Coppermine Pentium III processors. The 440BX chipset was limited to ATA-33, while the 820 supported ATA-66. The 440BX only supported AGP 2x, while the 820 provided AGP 4x capability.

But that wasn’t all. The 820 chipset also supported RDRAM, or Rambus DRAM. And therein lay the problem.

On paper, RDRAM was great, providing much higher bandwidth than SDRAM (synchronous DRAM), the most common memory type then in use. For boards based on the Intel 820 chipset, there were just two problems: In 1999, RDRAM was vastly more expensive than SDRAM, and Intel realized late in the game that while two RDRAM modules (or RIMMs) worked reliably with the 820 chipset, three or four didn’t (in retrospect one has to wonder if the high price and scarcity of RDRAM hid the problems for a time, because getting more than two RDRAM modules was so hard).

This was embarrassing for Intel because the (already postponed) 820 chipset launch was planned for September 27, 1999, but had to be canceled.

Intel “solved” the reliability problem by only providing two RIMM slots on its own 820-based board, the VC820 (VC for Vancouver). But that still did nothing for the price problem.

Intel’s solution was the 82805AA Memory Translator Hub (MTH) which could service two SDRAM DIMMs and made them look like Rambus DRAM to the 82820 MCH (Memory Controller Hub).

Intel FW82805AA Memory Translator Hub (MTH) found on an Intel CC820 board.
Intel FW82805AA MTH (Memory Translator Hub)

If it sounds like a hack, that’s because it was. And because it was a hack conceived late in the game, it turned out not to work terribly well. Naturally the added translation step did not add anything to performance, and the CC820 board was slower than the 440BX chipset it was supposed to replace when using the exact same PC100 SDRAM (note that with RDRAM, the 820 chipset was in fact slightly faster than 440BX, at least on some benchmarks).

So the choice was to get a board with the 440BX and give up on modern niceties like AGP 4x, or buy a board with the Intel 820 chipset and pay a vast premium for RDRAM, or buy a board with the 820 chipset and SDRAM and pay a performance penalty instead.

Adding insult to injury, it turned out that the 820-based boards with SDRAM slots (i.e. boards with the MTH) were not only underperforming but also unreliable. That led to a recall of the SDRAM boards with the Intel 820 Camino chipset, or Caminogate, in May 2000—about six months after the boards were released.

Buyers of the Intel CC820 boards had two options: Either a refund, or a replacement VC820 board with 128MB RDRAM (again, VC820 was the “purebred” RDRAM-based sibling of the CC820). For owners of non-Intel boards with the 820 chipset and MTH, the options were roughly the same but getting refunds or replacements was possibly much more complicated.

The MTH and the CC820 were put on a shipment hold, never to return again. Intel of course claimed that only a “small number” of boards were affected, but if their explanation is to be believed, the MTH was sensitive to electromagnetic noise and it’s quite possible that most if not all boards were affected, but not all were used in a noisy enough environment for that to matter.

The type designation on an Intel CC820 desktop board.
Intel CC820, not something you see very often

The unreliable MTH often led to system instability (crashes/reboots), but could also cause silent data corruption, which is something that’s extremely undesirable in a core computer component. Intel probably felt that an unreliable chipset/board was worse than none, and decided to weather the storm and await the arrival of Willamette Pentium 4, while doubling down on RDRAM (which is a different tale of hubris and failure).

What makes the Cape Cod episode especially embarrassing for Intel is that RDRAM for Pentium III was very much a solution in search of a problem. The Pentium III (especially a single CPU) was not a particularly memory-hungry CPU and expensive RDRAM didn’t really pay off. The dual-CPU 840 chipset perhaps benefited from RDRAM, but the 820 Camino chipset with RDRAM was roughly as fast as the older SDRAM-based 440BX chipset, while 820 with MTH+SDRAM was clearly slower than 440BX with the same memory. It’s no wonder that the 820 chipset went nowhere.

The final BIOS for the Intel CC820 board is dated April 28, 2000, just before the recall. It appears that the recall was the end of the Cape Cod boards and also the end of the MTH idea. Intel left the recall notice on their site for quite a while, and existing CC820 documentation, BIOS, and drivers were available for download, but that was it.

The VC820 board lived on noticeably longer, with the final BIOS update dated May 10, 2001. But overall the Intel 820 chipset did not do well at all, with the 440BX holding on much longer than Intel expected, the 810 chipset taking over the low-end market, third-party chipsets (e.g. VIA) doing well with Socket 370/SDRAM Pentium III boards, and the Intel 840 working well enough with Coppermine Pentium III until those were replaced by Willamette Pentium 4 on the high end in November 2000.

The CC820 board now in the OS/2 Museum’s possession clearly survived the recall, either because its previous owner was happy enough with it or because the board wasn’t really used. The board is functional and runs fine with a 733 MHz Coppermine Pentium III, although it wasn’t subjected to a thorough stress test which might uncover some instability.

There is no way to tell how many CC820 boards survived to this day. It won’t be a lot since the board was only on the market for a few months and then recalled, and the board was never all that great to begin with. But that’s exactly what makes it a true museum piece.

This entry was posted in Bugs, Intel, PC hardware, PC history, Pentium III, RDRAM. Bookmark the permalink.

40 Responses to The Cape Cod Disaster

  1. Pavel Zagrebin says:

    820 supports SMP, and such dual cpu boards are relatively common.

  2. Michal Necasek says:

    Can you please point me to one? I actually have a hard time believing that any Camino boards are common, but I guess that depends on the definition of “common”.

  3. Richard Wells says:

    Have a review of an Iwill dual processor i820 motherboard: https://www.anandtech.com/show/461/15

    Earlier in the review, it talks about the dual processor variant of the memory controller. The review is also interesting in that it rates some SDRAM translator solutions as more stable than RDRAM equipped motherboards.

  4. Michal Necasek says:

    Thanks — so the regular 820 was indeed single processor only, but there was a dual variant. I wonder what was supposed to be the difference between the 840 and 820DP chipsets, one worked and one didn’t?

  5. Richard Wells says:

    The working aspect was the most important part but the i840 had a few advantages over even a potentially working i820. The i840 supported 2 channels of memory so one could have 4 RDIMMs with a reasonable memory total. It was also planned to be paired with memory repeaters to have even more memory slots though I don’t know of any shipping system that included the memory repeater. The other optional advantage was 64 bit PCI which would have been nice in the SCSI workstation market.

    Note: i840 had its own set of problems early on. ECC wouldn’t work. Still a lot better than i820. It was bad that the problems with both i820 and i840 were only noticed by motherboard manufacturers not Intel.

  6. Michal Necasek says:

    Ah right, the 840 had 4 slots and at least in theory double the memory bandwidth of the 820. I’m not sure if the 820 was initially supposed to support 2 or 3 slots but 2 was all that worked reliably.

    I have not come across an 840 board with the memory repeater, only a couple of 860 boards with one. It’s possible that in the 840 era, RDRAM was just too damn expensive to populate 8 memory slots.

  7. Chris M. says:

    During my computer repair days, I only came across one i820 machine on the bench for repair. They were pretty damned rare. Everyone stuck with the 440BX until the P4 came out or had a lower end i810 machine. Intel’s stupid move into RDRAM cost them dearly. It also stopped me from upgrading my 440BX machine at the time…. it eventually got replaced in 2008 as there were too many other problems in the industry at the time (I really wanted to avoid Netburst and Athlon chipsets sucked).

  8. Michal Necasek says:

    That CC820 is the first i820 board I’ve ever seen. My recollection is that i810 and later 815 boards were very common, and I think the later 850 chipset for P4 was somewhat widespread, but 820/840/860 boards were really rare. The 840 and 860 were never meant to sell in large numbers I guess, but the 820 was. The post-RDRAM i845 Brookdale chipset was where things took off again.

  9. Pavel Zagrebin says:

    Asus P3C-D is “common” i820 dual slot 1 board.
    Alongside with retail version it was rebranded and used in HP Kayak XM600.

    I found three boards within half a year in ~2014.
    Ironically one cpu i820 board was much harder to find and I found only CC820 (with PIII ES plugged in it).

  10. Michal Necasek says:

    Thanks. Now I know why I never came across that board. A few years ago I did quite a bit of research into dual-CPU PII/PIII boards, but my requirement was at least one ISA slot. So this one didn’t come up, nearly all those boards had the 440BX chipset.

  11. Michal Necasek says:

    And after a bit of searching it seems the typical reaction to i820 boards is not “oh yeah I had one of those” (like with 440BX) but something more along the lines of, OMG WTF is this!? And what’s RDRAM, anyway?

  12. Chris M. says:

    i820 and RDRAM was widely panned by contemporary media and by potential buyers. It was stupid expensive for little performance gain. Problem is the alternatives weren’t any better. You had VIA’s Apollo Pro 133 with its fair share of flakiness with that 686B south bridge……. it wasn’t a pretty time that’s for sure (the Apollo Pro 133A fixed most of the problems except the southbridge). Intel refused to give the market a real replacement for the 440BX, which was baffling. The i815 was too little too late. The most bizarre thing is that Intel landed up building Socket 370 server/workstation boards with a ServerWorks chipset instead of their own products!

    The original i850 Pentium 4 systems weren’t much better. Just about every Socket 423 system that came on the service bench at the place I worked at died. The best built one seemed to be the Dell Dimension 8100, the only Socket 423 machine you’ll likely find still working today.

    The only smart thing we did was strip all the RIMMs off the dead boards and resell them to the people who were holding onto circa 2002-04 i850 Socket 478 systems for no good reason, but wanted more RAM. We even made money selling a stockpile of those stupid terminators that were required to be in unpopulated RIMM slots.

  13. Richard Wells says:

    RDRAM should have been very fast when paired with the large cache variants of the Pentium 4. I thought that was a lot of the reason why Intel went towards RDRAM. The latency problems of both RDRAM and the Pentium 4 would be concealed by the cache. AMD which couldn’t build chips with the same amount of cache would be unable to match the performance.

    Agreeing to standardize on RDRAM before Rambus had arranged enough production to meet the requirements was unwise, even if Intel got 10% of Rambus. That production short fall was also was prices were so excessive. Rambus should have been about 20% more expensive because of the more complex design but the throughput might have been able to provide enough value. Well, until multi-channel DDR came out, which offered high throughput with low latency and obsoleted RDRAM.

    Intel’s run in the early 2000s was an impressive litany of mistakes that few companies have managed to equal.

  14. Michal Necasek says:

    The impression I get is that especially after the 820 Camino chipset flamed out, Intel just did not want to do it again and did their best to ignore the Pentium III. Even though at the same time they kept making the CPUs and released the Tualatin and scaled the frequency all the way to 1.4 GHz for the PIII-S.

    The Intel Socket 370 server boards with ServerWorks chipsets are quite nice, but it is kind of funny.

    And yeah, Socket 423 is about as exotic as RDRAM, if not more so. Intel went up to 2.0 GHz but the faster grades are almost impossible to find in the S423 format, whereas the Socket 478 models are common.

  15. Michal Necasek says:

    Certainly few companies made so many mistakes and lived to tell the tale (or lived to talk about it as little as possible, as it may be).

    I completely agree that RDRAM would have been able to justify a price premium, only with the prices RDRAM was initially available at, it just didn’t make any sense. In retrospect, it is an interesting question if Rambus was a victim of poor timing or if RDRAM never really had a chance because the improvement over SDRAM wasn’t big enough and DDR was too good.

  16. Zir Blazer says:

    Contemporary commentaries about the i820: https://redhill.net.au/c/c-g.html#i820

    Note that on Coppermine, RDRAM itself was a stupid idea to begin with. Because Coppermine had a 64 Bits wide 133 MHz clocked FSB, the maximum theorical bandwidth between the Processor and the Northbridge was 1066 MB p/s, which is the same Memory Bandwidth achievable with 133 MHz SDRAM. A 16 Bits wide 400 MHz (800 MHz effective, as it was Double Pumped, a la DDR) RIMM provided 1600 MB p/s. This means that there was a major bottleneck, as the Processor itself couldn’t use the extra bandwidth yet had to pay the price for the increased RDRAM latency.
    The extra Memory Bandwidth would have been very useful for Devices with major independent access to the RAM, like an IGP (Which was only found in value Chipsets, not high end RDRAM equipped ones. And it would be stupid to buy ultra expensive RAM for an IGP instead of a real, discrete Video Card), or a DMA heavy card that could completely bypass the Processor, like a SCSI Controller or NIC for Server duties. I suppose that at the bare minimum there is enough non-Processor RAM accesses that allowed the i820 with 400 MHz RIMMs to perform better in general than the 440BX, but the difference in minor scenarios was not even close to justify it.

  17. Yuhong Bao says:

    Thinking about it, I wonder why Almador was cancelled on the desktop. I think Intel even did a Intel 810e “Universal Socket 370” chipset that hopefully wasn’t popular because it lacked support for 256Mbit SDRAM.

  18. Richard Wells says:

    Almador should have fallen into a no-mans land on the desktop. While Almador gave marginally better performance than i815, Almador did not give as good performance as a $50 video card. The attempt to repurpose RDRAM into MRIMM was okay in a mid-range laptop but blew past the price points on the desktop. Even more expensive RDRAM in a special low profile stick means marketing difficulties.

    i810e was fairly popular. Low end boxes with just enough graphics to handle DVDs and sold with 64 MB of RAM. Perfect for the college student on a tight budget that does not game. The advantage of small capacity SDRAM and PCI slots was that adding memory and a video card was quite cheap since those would be surplus from higher end systems that got upgraded.

  19. Michal Necasek says:

    I wonder where the “900,000 motherboards” figure in the Red Hill article came from (not that I doubt it, I just have no idea how accurate it is).

    Shoving RDRAM into PIII systems does seem inexplicable in retrospect. Sure, the 840 chipset was a little faster than 440BX, but was that really because of RDRAM? The Pentium 4 design on the other hand was quite likely influenced by RDRAM and may have relied on high memory bandwidth because that was something RDRAM could provide. And the Willamette FSB did have the bandwidth to match RDRAM.

  20. Michal Necasek says:

    The 810 chipset (maybe especially 810E and 810E2) was also popular in low-end business desktops. Yeah it could not do 3D, but that didn’t matter, and 2D performance was actually not bad (due to using system RAM, the performance characteristics were noticeably different from discrete PCI or AGP cards). Everything was integrated so the machines were cheap, and one could still plug in a decent Coppermine or Tualatin processor.

  21. Yuhong Bao says:

    AFAIK even on laptops Almador’s MRIMM option was unpopular enough that they eventually removed the option from the datasheet.

  22. Richard Wells says:

    https://www.cnet.com/news/rambus-at-the-root-of-intels-memory-troubles/ quotes an analyst as saying nearly a million i820 motherboards with MTH would be replaced. One could check the Intel financials and see what their charge was for it though I am not sure Intel was spending the $100 per motherboard that the analyst was using for estimating.

    PC 600 RDRAM was a lot cheaper and would have sufficed for the Pentium III but, IIRC, slower RDRAM wasn’t available in quantity until the Pentium 4 was shipping. See the rage as budget Pentium 4 systems got PC 600 while Pentium IIIs were paired with PC 800. https://www.geek.com/chips/p4-often-ships-with-slow-rdram-542798/

  23. Michal Necasek says:

    On the one hand it seems odd to blame Rambus for broken SDRAM support… on the other hand it’s entirely appropriate because if it weren’t for all the Rambus mess, the MTH would not have been needed at all.

    The cnet article makes the good point that replacing the boards was one thing, but replacing SDRAM with RDRAM must have been quite expensive.

  24. Yuhong Bao says:

    Even the i815 did not support more than 512MB of RAM despite the 256Mbit SDRAM support, which Almador was going to fix. It is still important because this crossover occurred in 2002, when Tualatin Celerons was likely still being sold.

  25. Yuhong Bao says:

    “PC 600 RDRAM was a lot cheaper and would have sufficed for the Pentium III but, IIRC, slower RDRAM wasn’t available in quantity until the Pentium 4 was shipping.”
    Not true. In fact there was PC700, designed for the i820 chipset (with all others it was the same as PC600).

  26. Richard Wells says:

    There may have been slower RDRAM announced but it wasn’t shipping in quantity. I think nearly all chips that passed inspection were capable of PC 800 and unless manufacturing capacity exceeded demand no low price RDRAM would be sold. Looking at a Computer Shopper from 2000, I can find multiple vendors all offering PC 800 memory but only one listing PC 700 and none list PC 600. Prices were only for PC 800. In comparison, nearly every speed and capacity of 168-pin and 72-pin memory was listed with prices for most models in most ads.

    The 1 GB using 2 sticks of 512 MB SDRAM on budget motherboards was unlikely to be an issue in practice. The number of users planning on buying $1000 of upgrade RAM for a $500 computer tends to be modest. Anyone with the knowledge and the budget would instead be likely to pick up a motherboard with modern RAM instead of the then soon to be discontinued SDRAM.

    I actually had a surplus i810e system with scavenged extra RAM and a retired PCI video card. While I had briefly used it to test OS/2 on faster hardware (Workplace Shell flies at 800 MHz), it became a low end Windows XP test box. If the software ran acceptably on that, any user with a new system would be ecstatic over the performance. It was difficult to use the full 512 MB for a combination of Office and the software I was writing. It did limit the number of web browser tabs that could be open but a lot of web pages needed a much more powerful CPU just to be minimally responsive.

  27. Yuhong Bao says:

    “The 1 GB using 2 sticks of 512 MB SDRAM on budget motherboards was unlikely to be an issue in practice.”
    In 2000, 256Mbit SDRAM support was not important. By the time of the Tualatin in 2002 it was a different story.

  28. Yuhong Bao says:

    Interestingly the D815EEA used Socket 370 and the CC820 used Slot 1. I wonder why Intel did not produce Slot 1 Intel 815 chipset motherboards for the recall.

  29. Richard Wells says:

    Why replace i820 motherboards with i815 boards when Intel had a substantial stock pile of i820 boards equipped for Rambus? Scrapping working unsold boards would have made the whole i820 problem just that much more expensive.

  30. MiaM says:

    Generally re crappy chipsets in this era: I wounder how much the dotcom crash did impact on this? The 440 series chipsets came out well before the crash, in an environment where IT seemed to grow steadily (although not as crazy as the last year or so before the crash). Maybe both Intel and VIA kind of tried to cut costs and thus didn’t really test their products as well as they should had?

    It’s really sad that this did happen. Up to the 440 series the safe and reasonable choice for a PC motherboard were a board with an Intel chip set. After the 820 debacle and the lack of any really good chipsets for Athlon for a while, we unfortunitely went back to the situation where each buyer had to investigate the market even if they really didn’t have any interest in anything else than getting a good working PC.

  31. Yuhong Bao says:

    AFAIK the Intel deal with Rambus dates back to 1996.

  32. Yuhong Bao says:

    Though this reminds me of the Intel CPU shortages in 2000.

  33. Michal Necasek says:

    The Intel 820 Camino chipset was released in late 1999, months before the dot com bubble burst. The CC820 board recall happened at about the same time as the bubble was bursting, but it was clearly not caused by it (if anything it may have contributed to the crash). The 820 chipset does seem to have been insufficiently tested, but I don’t know if that was a result of cost cutting or if Intel’s testing procedures at the time were not sufficient to catch the problems.

    The capacitor plague of the early 2000s may have been perhaps related to the dot com crash.

  34. Richard Wells says:

    DN (August 5, 1999) has an article on Rambus signal integrity which looks like it explains how the motherboard prototype could work in the lab but not in the field. Rambus terminates into the motherboard voltage plane. I think with other loads and a cheaper power supply the RIMM might not have the correct resistance. Okay, it has been about 40 years since my college courses in electrical design but that just seems like a very poor method for ensuring reliable operation.

    Capacitor plague was going happen no matter what. It was a promising cheap manufacturing method that was always about 6 months from working correctly. Someone was going to try to put that into production and hope they could solve the problems of longevity before failures became excessive.

  35. Richard Wells says:

    That should be EDN (Electronic Design News).

  36. Yuhong Bao says:

    BTW, yes, I am well aware of the Pentium 4 launch in November 2000 that shipped with two 64MB RDRAM sticks. It is probably a good thing that 128Mbit SDRAM was not common in early 2000, so only 256MB of RAM maximum was common.

  37. Yuhong Bao says:

    Interestingly, it seems that the MRH-S (Intel 840) was already showing problems months before the MTH recall. I wonder why.

  38. Yuhong Bao says:

    Interestingly https://www.theregister.com/2000/04/03/ami_ships_840_mobo/ says that “AMI claims that any 840-based board with fewer than eight layers is likely to experience MRH errors”

  39. Michal Necasek says:

    Very interesting. And all of those problems were because of Rambus. In hindsight that must have been one of Intel’s bigger blunders, and probably cost them a couple billion dollars.

  40. Yuhong Bao says:

    https://www.samsung.com/semiconductor/newsroom/news-events/samsung-passes-milestone-by-shipping-10-millionth-rdram-to-pc-oems/
    10 million RDRAM devices would 2.5 million 64MB RIMMs modules. I wonder how many of these modules got used for the MTH recall.

Leave a Reply

Your email address will not be published.

This site uses Akismet to reduce spam. Learn how your comment data is processed.