Intel OverDrive Part III: Pentium II OverDrive

Pentium II OverDrive

The Pentium II OverDrive, released in August 1998, was the swan song of the OverDrive product line. It is suitable for Socket 8 systems as an upgrade of 150-200 MHz Pentium Pro processors. Only one model was sold with a PODP66X333 designation.

Pentium II OverDrive Box

Pentium II OverDrive Box

The Pentium II OverDrive upgrades a Pentium Pro system with 66 MHz bus to a 333 MHz Pentium II, or a 60 MHz bus system to a 300 MHz processor. In other words, the processor uses a fixed 5× multiplier.

The Pentium II OverDrive is a curious beast. It’s perhaps closest to a Pentium II Xeon processor: it combines the Deschutes Pentium II core with a full-speed 512 KB L2 cache of a Pentium II Xeon. Why is that?

To understand the Pentium II OverDrive design, it is necessary to understand the Intel CPU history of the mid-1990s. The Pentium Pro (P6) was a high-end CPU which started an incredibly long line of processors and effectively killed all RISC competition. But the Pentium Pro was never a mass-market processor. It contained 256K, 512K, and later even 1M of built-in L2 cache running at core CPU speed (up to 200 MHz). It was very large and difficult to manufacture, and therefore expensive. It was also highly popular in server systems because it natively supported 4-way SMP and the fast, large L2 cache was very suitable for server and workstation workloads.

Pentium II OverDrive in blister pack

Pentium II OverDrive in blister pack

The Pentium Pro was also well known to have performance problems with 16-bit code, such as the code found in Windows 95/98, and did not support the MMX instruction set. For mass-market PCs, the Pentium MMX was replaced by the Pentium II, which was a modified Pentium Pro with MMX support, improved 16-bit code execution performance, clock speeds starting at 233 MHz, but L2 cache running at only half the CPU clock speed.

The Pentium II was not suitable for servers as a successor to the Pentium Pro. That is why Intel developed the Pentium II Xeon, a processor with Pentium II core but 512K, 1M, or even 2M built-in L2 cache.

The Pentium II OverDrive is essentially a Pentium II Xeon in a Socket 8 package, with a fixed 5× multiplier and 60 or 66 MHz bus. It would be an excellent replacement for a 200 MHz Pentium Pro with 512K L2 cache; it keeps the large and fast cache, adds MMX, and increases clock speed to 333 MHz. For slower Pentium Pro systems, the benefit would have been even greater.

So why was the Pentium II OverDrive not a hugely popular processor? Because the Pentium Pro was never sold in large quantities (so the set of systems suitable for the upgrade was never huge), and because the OverDrive was too little, too late. At the time when the 333 MHz Pentium II OverDrive was available, Pentium II processors already reached 450 MHz clock speed. A year after that, 600 MHz Pentium III processors were available. At the same time, Pentium II Xeon processors were available with larger L2 caches and higher clock speeds.

Double the OverDrive, double the fun.

Double the OverDrive, double the fun.

In other words, when it was introduced, the Pentium II OverDrive was not as fast as the standard offerings and very quickly fell far behind, with no faster OverDrives available. At the same time it wasn’t particularly cheap. As a result, those seeking top performance (and buyers of Pentium Pro systems certainly did) were better served by purchasing a more modern system with newer and faster chipset, AGP support, 100 MHz bus, SDRAM, and a faster CPU.

All these factors probably led Intel to give up on the OverDrive line entirely, especially given the non-trivial engineering effort. The promise of upgradability never seemed to have quite panned out (except for the first 486 OverDrive CPUs). In addition, now that Intel was also firmly in the chipset business, selling a new CPU together with a new board (i.e. new chipset) probably sounded like a better idea anyway. It also didn’t require any custom engineering and special marketing for relatively low-volume processors.

Details and Performance

The CPUID TFMS (Type/Family/Model/Stepping) of the PODP66X333 is 1632h. That would indicate the original Klamath Pentium II but the OverDrive has a newer Deschutes core. Why the mismatch? The explanation is very simple: When the Pentium Pro was released, Intel specified an OverDriver processor with 163x CPUID so that BIOSes could easily recognize the upgrade CPU. By the time the Pentium II OverDrive was actually released, the CPUID (family 6, model 3) no longer matched reality, but had to be kept so as to not confuse existing BIOSes.

Exploded view of Pentium II OverDrive

Exploded view of Pentium II OverDrive

Another historical curiosity is that Intel specified a VRM (Voltage Regulator Module) socket for the Socket 8 OverDrive processor, but actual units shipped with an onboard VRM.

The fast L2 cache was designed to reduce the impact of relatively slow main memory in Socket 8 systems. Intel had previously used a similar approach with the original Pentium OverDrive which sported twice the size of the L1 cache of standard Pentium processors.

In general, the Pentium II OverDrive performance is somewhere in the Pentium II 300 to 333 MHz range, but in a previous-generation platform. That means no SDRAM, no AGP, though at least the standard 440FX chipset supports up to 1GB RAM.

For number crunching, the OverDrive is great. But it won’t speed up graphics performance much (unless MMX makes a big difference in a particular application). The pure CPU performance of a Pentium II OverDrive is significantly better than that of a 200 MHz Pentium Pro, and the differential is even greater when replacing, say, a 166 MHz Pentium Pro. But it won’t speed up I/O such as slow disks and graphics cards. That’s of course true of OverDrives in general.

Pentium II OverDrive closeup

Pentium II OverDrive closeup

There was talk about the Pentium II OverDrive being a “home made Xeon“. That was not exactly wrong, but the problem was that the OverDrive didn’t work in any modern (at the time, i.e. late 1998) platform, which eliminated a good deal of the fast L2 cache advantage.

Conclusion

Since there are no other alternatives, there’s no question that a Pentium II OverDrive is the best thing for a Socket 8 system, including machines with more than two sockets (even though Intel only guaranteed dual processor operation).

Largely due to the classic too little, too late problem, the Pentium II OverDrive was not a particularly successful product. In about 1999 it concluded the line of Intel OverDrive processors which had started with a great fanfare in 1992.

This entry was posted in Intel, Pentium II, Pentium Pro. Bookmark the permalink.

19 Responses to Intel OverDrive Part III: Pentium II OverDrive

  1. jwaiv says:

    I remember these.

    15-16 years ago I built an ALR 6×6 out of spare parts scavenged from eBay, the only parts I never found were the case and PSU so I ended up modifying off the shelf items for these bits. When I originally built it I used six Pentium Pro 200MHz but I eventually upgraded it to six of these Pentium II Overdrives, at the time it was amazingly cheap but I only did it for a project. From what I can recall the 6×6 was the only system that supported more than four processors but I don’t think Intel officially supported it.

    You could even overclock a configuration with six of these by replacing a clock crystal on the 6×6 motherboard, I remember buying the crystal but never did the mod. A couple of users on a forum that is probably lost to time got systems with six of these up to around 400MHz without any trouble.

  2. jwaiv says:

    A project post from someone else seems to have survived all these years, they also used overdrives – http://arstechnica.com/civis/viewtopic.php?f=7&t=878654

  3. Velocet says:

    Please don’t use that much thermal grease. Just a little dot in the middle is enough. Using so much thermal grease will decrease cooling efficiency.

  4. Michal Necasek says:

    You’re not wrong, but that’s something you need to take up with Intel 🙂 The photo shows exactly what Intel delivered, I didn’t add or remove any thermal compound.

  5. Michal Necasek says:

    Okay, so maybe the PII OverDrive isn’t limited to dual configs either. That’s nice to know. I can only test a dual config and that seems to work just dandy.

    As far as I know, Compaq was building some 8-processor systems around that time, using custom logic. Quad Pentium Pros was the most Intel directly supported with their CPUs and chipsets.

  6. GL1zdA says:

    The Overdrive officially supported only dual processing, but I think it just means larger configurations are “unsupported” not “impossible”.

    As far as large systems are concerned, Data General did 32 CPU Pentium Pro servers (“AV 20000”). Sequent and NCR also did large Pentium Pro systems. Axil Northbridge NX801 was a 8-way able to run unmodified NT. So could NCR and smaller Data General systems http://windowsitpro.com/windows-server/8-way-scalability

  7. Michal Necasek says:

    I adjusted the article text slightly. Not really surprising that big boxes with the PII ODP… after all I ran one of those dual Celeron systems for a while myself.

  8. Morty says:

    Maybe a bit off-topic. The article mentions something I have also heard elsewhere, namely that the Pentium Pro was really slow with 16-bit code making it unsuitable for Win 3.1 and even Win 95/98 etc. which had lots of 16-bit code paths. But does anyone know anything more concrete about the cause? Were it specific instructions that were slow or transitions between 16-bit / 32-bit mode? Were the 16-bit instructions slower than their 32-bit counterparts? If so, was that because they were emulated? Also curious about the extent of the “slowness”.

  9. Andreas Kohl says:

    With usual 16-bit applications a PPro is not much slower than a Pentium MMX with same clock rate. It’s almost the same performance (or better) of a Pentium without MMX.
    Also “mass-market” PCs used Pentium Pro processors, for instance Compaq Deskpro or IBM PC365 ca. 1996/97. So it was quite common for running large networks of Windows NT workstations.

  10. Michal Necasek says:

    I will quote the relevant portion of the Collins article: With the Pentium, Intel introduced a 94-entry, two-way set associative cache of segment-descriptor cache entries. Therefore, the phrase “segment-descriptor cache” is now ambiguous, with two possible meanings. Making matters worse, the new segment-descriptor cache was removed from the Pentium Pro design, but reintroduced in the Pentium II. (The lack of the new segment-descriptor cache in the Pentium Pro largely accounted for its poor 16-bit performance.)

    The upshot is that protected-mode segment register loads were expensive on the Pentium Pro. That would hit Windows 3.1/9x pretty hard but not NT, with OS/2 taking a slight hit.

    Another problem was with partial register stalls, something that probably especially hurt Windows 9x (mixed 16-bit/32-bit code). There’s a lot about partial register stalls out there.

    The statement that “with usual 16-bit applications a PPro is not much slower than a Pentium MMX with same clock rate” is true but omits the fact that a PPro was significantly more expensive, so the same performance for a lot more money was not exactly a win.

    The Pentium Pro was an excellent CPU for Windows NT, OS/2, and UNIX, but those weren’t mass-market operating systems at the time. Perhaps Intel thought that everyone would be running 32-bit software by 1996… how wrong they were! Anyway, the Pentium II fixed the worst problems and with much higher clock speeds it was clearly superior to the Pentium MMX.

  11. Yuhong Bao says:

    http://www.angelfire.com/droid/pacman/microsoft-partners/ibm/slime.html

    I should also mention in addition to what Joe Barr said that IBM was able to port the graphics rendering engine to 32-bit in OS/2 2.1. This took more RAM than the 16-bit one in OS/2 2.0, but of course was faster on the Pentium Pro.

  12. Richard Wells says:

    The Pentium Pro design was not going to be a viable mass market design until several die shrinks had passed. Just no way to put that much high speed cache in a package and achieve a price under $500 so focusing on the high performance 32-bit market was the way to go. IIRC, the Pentium Pro was slightly faster clock for clock running 32-bit code than the Pentium II or III unless newer instructions were used.

    @Yuhong: Canopus was a bit more complicated than that.

  13. Michal Necasek says:

    The manufacturing process was definitely a problem (take two big chips, glue them together, then test), and it drove up the price. The per-clock performance is interesting, but you need to add that the Pentium Pro topped out at 200 MHz and the Pentium II started at 233 MHz. A PPro simply can’t keep up with a 300 or 400 MHz Pentium II, and even less so if the PII has 100 MHz FSB and SDRAM.

    It’s probably not coincidence that after the Pentium Pro, Intel split off the Xeon line. Although the black 200 MHz Pentium Pro with 1MB L2 cache was a Xeon in all but a name.

  14. Back in the day, we got hp workstations with Pentium pros. I got crazy Lucky with dual 150s, everyone else had single 180s if I’m remembering… anyway first thing we did was load Windows 95, as we all heard it was such crap. Granted we were all moving up from 486 and pentium 60 based machines, 95 on a ppro felt plenty fast enough. .. But faster than the older kit for sure. Then we loaded NT and it was so much better.

  15. Morty says:

    OK thanks for info – it seems that it was the PPro’s lack of segment descriptor cache that was the culprit. But I guess then that it wasn’t really the 16-bit part that was the troublesome software. It was more a case of software making frequent use of segment registeres that were exposed to it. More modern operating systems probably tended to be both fully 32-bit AND do less segment register shenanigans due to the full embracement of the flat memory model. Whereas older operating systems, like Win 9x, while in principle also being flat mode operating system and supporting 32-bit, still had 16-bit code paths and used segment registers for various purposes.

    So the PPro being slow with 16-bit code was more of a correlation thing – it wasn’t the 16-bit instructions (as in mov ax,bx etc.) that was slowing it down it just tended to correlate with other aspects of operating system behavior which DID slow it down.

  16. Andreas Kohl says:

    Ok, Windows NT or UNIX workstation or server systems were not “mass market” but an important market – especially by adding the amount of Netware and OS/2 Warp Server systems equipped with Pentium Pro processors until the end of 1998.

    I hope you also have an Intel SnapIn 386 ready by hand for some kind of prequel of this series 😉

  17. Michal Necasek says:

    No prequel currently planned… I suppose the Inboard 386 would be the first (1987 or so?), but I don’t have one. I have a RapidCAD chip (well, the two-chip set), but it doesn’t appear to be working 🙁 Maybe sometime in the future.

  18. Joshua Rodd says:

    The Intel Inboard for PC and XT would be particularly interesting since I think it’s the only time anyone tried to put a 386 on an 8-bit bus. Intel even made a set of 32-bit VxDs for Windows 3.0 386 Enhanced Mode.

Leave a Reply

Your email address will not be published. Required fields are marked *