A couple of weeks ago, we took a look at the overclocking possibilities provided to us by Intel’s FC-PGA Pentium III 500E and 550E in Part 1 of our Overclocking the FC-PGA series.  In that article we found that the extremely high overclockability of the new FC-PGA CPUs was due to the following reasons:

• The FC-PGA 500E/550E are virtually identical to the 733MHz Slot-1 counterparts, and thus hitting 733MHz with a 500E/550E shouldn’t be out of the question.
  
• The 500E/550E are only running at 1.60v, 0.05v down from the default 1.65v of the Slot-1 CPUs.  This allows for a bit of breathing room with overclocking since increasing the core voltage to 1.65v is still within the recommended range for the Pentium III’s Coppermine core.
• Since they have such a low clock-multiplier (relative to the Slot-1 Pentium III Es), increasing the FSB setting to 124MHz and beyond provides very realistic overclocked settings unlike the newer Celerons which carry extremely high clock multipliers (i.e. 7.0x/7.5x/8.0x), consequently making the higher FSB settings useless for overclocking purposes.

We also pointed out that through the use of a newer Socket-370 to Slot-1 converter such as Iwill’s Slocket-II or MSI’s 6905Master, the FC-PGA Pentium IIIs would work perfectly fine in most BX boards, offering an excellent upgrade path to users that have faithfully stuck with their BX setups.  Unfortunately, staying with your BX setup limits the success you would have at higher FSB settings because of the strain it puts on your AGP graphics card, due to the lack of a ½ AGP clock divider. 

This limitation is obviously non-existent with motherboards based on chipsets with official support for the 133MHz FSB as they feature a ½ AGP clock divider (133MHz / 2 = 66MHz = AGP clock specification).  The beauty of the ½ AGP clock divider is that it allows FSB settings, such as the 140MHz and 150MHz frequencies, to be taken advantage of, because, even at 150MHz, the AGP bus is only running at 75MHz, which is still within the range of toleration for most AGP cards including the tons of GeForce and TNT2 based cards we’ve tried at the speed. 

At 150MHz the only limitations are really the chipset and your memory as the ¼ PCI clock divider keeps the PCI bus at 37.5MHz, which is close enough to the 33MHz specification that most peripherals don’t have a problem with it.  This makes the combination of an official 133MHz chipset and the FC-PGA the perfect solution for overclockers that aren’t looking to spend a lot of money on a new Pentium III setup.  And thus we have the basis for Part 2 of our ongoing series of Overclocking the FC-PGA CPUs - overclocking FC-PGA using the 133MHz FSB and beyond. 

The first question we asked ourselves was, what 133MHz chipsets are out there that would be perfect for this little experiment?  Out of the currently available chipsets, we have the 810E and 820, both from Intel, and the two VIA solutions, the Apollo Pro 133 and Apollo Pro 133A. 

The i810E is far from a reasonable solution as its integrated i752 AGP graphics accelerator is far from the best for gaming or professional application performance; if you’re not into gaming/professional apps then the i810E may be a valid solution but then again that leaves business applications as the only programs you run, in which case a cheaper Celeron would be a better path (definitely from a price perspective) to go down. 

The high cost associated with RDRAM will keep the i820 chipset out of the price range addressed by this comparison, but if your budget allows you to purchase 128MB of RDRAM at $900+ then you can afford to pick up a true Pentium III 733.  The i820 + SDRAM solution is, however, a viable option.  The performance drop provided by the i820 + SDRAM combination as a result of the Memory Translator Hub required for SDRAM support on the i820 isn’t worth the added cost, and you’re better off sticking with an older BX setup.  Regardless, this is a solution we’ll be taking a look at in the future, just not within the scope of this article.

The situation with VIA’s two chipsets is interesting.  Released virtually back to back, the Apollo Pro 133 and Apollo Pro 133A are practically identical chipsets, except that the 133A supports AGP 4X and the 133 offers only AGP 2X support.  It should be noted that the 694X North Bridge of the 133A is a 510-pin solution whereas the 693A North Bridge of the 133 is a 492-pin chip. 

Unfortunately, this keeps the 694X from being a pin-compatible replacement to the 693A thus making the 133A an undesirable solution to manufacturers that already have a design created and in production with the 492-pin 693A North Bridge of the Apollo Pro 133.  This is why companies like ABIT are releasing boards that are still based on the older Apollo Pro 133 chipset and thus don’t support AGP 4X which is a feature of the 694X North Bridge of its 133A successor. 

Then again, from our tests, the advantage of AGP 4X over AGP 2X is far from noticeable and shouldn’t be a major concern for most users.  This makes both the 133 and 133A viable options and ideal platforms for use with the FC-PGA CPUs.  Since the two perform identically with the exception of the AGP 4X advantage of the 133A, we chose to go with the 133A for our overclocking tests.

Apollo Pro 133A Motherboards

We are finally seeing motherboards based on the Apollo Pro 133A chipset emerge in the market.  The trend was started by Tyan with their Trinity 400 which was actually used in the first 133MHz FSB Micron PCs that shipped with the Pentium III 733 when the i820 platform was unavailable.

At last year’s Comdex we gave a brief preview of what you could expect from motherboards based on the 133A chipset as we moved into the year 2000.  And although we still haven’t seen the wall of motherboards that we were shown at Comdex appear in the retail channels, we are quickly getting there.  Manufacturers are realizing that, for most users, the i820 platform isn’t a viable solution and that there needs to be an alternative with 133MHz FSB support, making the 133A an ideal candidate. 

We have been playing around with three specific Apollo Pro 133A based motherboards, the Tyan Trinity 400, AOpen MX64 and the FIC KA11.  While you’ll see individual reviews of all three of these boards shortly, if you’re looking to make a purchase soon we’re going to offer some quick buying tips on these three motherboards.

Tyan Trinity 400 S1854

Click to Enlarge

The Tyan Trinity 400 has been out for the longest, and although it has been available for sale for quite some time, until January 5, 2000 users had been forced to run their Trinity 400 with a beta revision of the BIOS.  The final revision 1.00 BIOS just came out on the 5^th and has solved quite a few problems that were present with the previous 0.93 and older BIOS revisions. 

The Trinity 400 itself features both Socket-370 and Slot-1 interfaces, with only one of the two capable of being occupied at any given time.  A problem that does exist with the Trinity 400 is that when using the FC-PGA Pentium III, there is no way to force the board to detect the CPU as a 133MHz FSB CPU (since there is no way to change the pin state on the CPU itself, i.e. no jumpers on a socketed CPU) so the AGP divider is always set to 2/3. 

Unfortunately, because of this problem, we were unable to boot the Trinity 400 at anything above 124MHz with the FC-PGA in the socket.  This renders the Socket-370 virtually useless for the purpose at hand, but since a Socket-370 to Slot-1 adapter can be used in the Slot-1 interface on the Trinity 400, this problem can be avoided.  It would have been nice if Tyan had included jumper settings to force the detection of any CPU used as a 66/100/133MHz FSB CPU, but we’ll just have to make do with the FC-PGA + Socket-370 to Slot-1 adapter combination. 

Click to Enlarge

The latest revision of the Trinity 400’s BIOS (1.00) does add some interesting options to the Chipset configuration such as AGP Driving Control  and CPU IOQ Size.  The AGP Driving Control setting allows for a number of possible settings to be inputted in as hex values, but, from our experience, it’s best to just leave this setting on Auto.  The next setting is even more obscure than the AGP Driving Control, which is CPU IOQ Size.  The setting has two options, 1 level and 4 level, and although we originally thought this setting would have no affect on performance, it turns out that the default setting of 1 level results in up to a 15% drop in performance when compared to running with the value set to 4 level.  We honestly can’t say what this setting controls, but in the end, setting it to 4 level resulted in faster performance with no noticeable sacrifices, so be sure to keep it at level 4 if you can. 

The last problem we had with the Trinity 400 was that, upon enabling AGP 4X on the board, it would cause 3D applications and games to lock up with a TNT2 and it would result in visual artifacts when used with a GeForce.  However, due to the lack of a performance improvement, we had no problem disabling AGP 4X because, in the end, as a result of that mysterious CPU IOQ Size setting, the performance of the Trinity 400 was noticeably greater (even while running in AGP 2X mode) than either of the other two boards which ran in AGP 4X mode. 

From a features standpoint, the Trinity 400 makes use of the Mobile South Bridge which is the 596B chip from VIA.  The other option would be the 686A which adds integrated hardware monitoring, support for an AMR slot (which is useful for system integrators but useless to the rest of us), and support for four USB ports.  Because the Trinity 400 uses the cheaper 596B chip, we only see two USB ports on the board. 

AOpen MX64

Click to Enlarge

The AOpen MX64 is a MicroATX form factor Apollo Pro 133A board with a Slot-1 interface.  Out of the three boards mentioned here, the AOpen was by far the most stable and the difference was definitely noticeable.  The board is very limited in terms of expansion (3 PCI slots), but it should be a very low-cost 133A board that boasts incredible stability as its primary feature. 

The MX64’s BIOS does not feature any of the odd settings we noticed with the Trinity 400; unfortunately, this also means that it doesn’t allow us to adjust the state of the CPU IOQ Size setting which resulted in the performance of the MX64 being a good 5 – 10% slower than the Trinity 400.  Since the CPU IOQ Size setting was previously not present in the Trinity 400 BIOS setup, it shouldn’t take all that much for AOpen to include a similar feature in a future BIOS update. 

Of the three boards mentioned here, the MX64 was the only one to feature the Super South Bridge (686A) which boasts support for integrated hardware monitoring, AMR, and four USB ports, and, while AOpen didn’t take advantage of the AMR support, they did make use of the four USB ports which can be a pretty useful feature.  It may not be the fastest or the best for expansion, but the MX64 is definitely the most solid Apollo Pro 133A out of the three and we expect it to be one of the most solid out of the entire 133A crowd once the rest are released, as has been AOpen’s tradition as the second largest motherboard manufacturer world-wide. 

FIC KA-11

Click to Enlarge

Of these three motherboards, the KA-11 gave us the most problems.  The board itself seems to be a quickly constructed solution where the 694X North Bridge replaced the 693A used in one of their older designs, specifically the KA-6110.  While this could be the explanation of the problems we encountered, they may be attributed to other unknown factors.  Regardless, the conclusion remains that the KA-11 performed very poorly in the stability department.

The KA-11 would randomly lock up during normal operation, sometimes even completely shutting down the system while not even undertaking any heavy tasks.  It’s obvious that the KA-11 has some issues, but, hopefully, they are not problems that FIC can’t solve.  They shouldn’t be, because two other manufacturers have already proven to us that they are capable of producing 133A boards without these issues. 

While the board’s Award BIOS setup doesn’t have the same CPU IOQ Size setting as the Trinity 400, it is noticeably faster than the AOpen, but still slower than the Tyan with CPU IOQ Size set to 4.  Once again, it seems like there is quite a bit of room for improvement in performance through BIOS enhancements/tweaks and, as more of these boards are released, expect performance to approach a more uniform level among competitors.  Until then, it seems like Tyan is on top. 

One unique feature of the KA-11 is the 6-pin power connector it features in addition to the 20-pin ATX PSU connector.  The 6-pin connector can be used for fan speed control of the power supply’s fan.  While this may not seem like the most useful feature, it is unique nonetheless.  The KA-11 was also the only board of the three to feature four DIMM slots.

Overclocking the FC-PGA

Now to the really important part, how far did these chips go?  We already knew that the four CPUs that we used in the last comparison made it up to 733MHz and 667MHz for the 550Es and 500Es respectively, but what about with the increased FSB settings offered by an Apollo Pro 133A board?  In order to try our luck once again, we used the same CPUs while adding two more graciously supplied to us by Azzo computers. 

The highest FSB setting available on all of the three Apollo Pro 133A based boards we looked at was the 150MHz setting, so it made sense for us to aim straight for that mark.  At the 150MHz FSB a 500E would theoretically end up running at 750MHz and a 550E would end up running at a comfy 825MHz.  And the results of our overclocking attempt?

At the default core voltage of 1.60v, all of the 500E and 550E CPUs we tried worked perfectly at 750MHz and 825MHz respectively.  And since these tests were run on a platform with official support for the 133MHz FSB, the ½ AGP clock divider allowed the 150MHz FSB to be used without any noticeable sacrifice of stability.  At 825MHz, the 550E instantly becomes the fastest performer from Intel since the Slot-1 Pentium III line currently “only” goes up to 800MHz. 

Even if you’re set on the i820 platform you can expect a similar overclocking potential out of the 500E/550E CPUs.  And if you happen to get a poor yield and it won’t work at 750/825MHz, then the 140MHz and 133MHz FSB settings are always there, and once again, courtesy of the ½ AGP divider, they’ll both be stable regardless of what applications/games you’re running.

The Test

Historically, VIA has been known for having inferior memory performance when compared to their Intel counterparts. And as the Content Creation Winstone 2000 test shows us here, history does have a tendency to repeat itself.

The first thing you have to understand is that Content Creation Winstone 2000 mimics real world application usage to the point that it's scores should definitely be paid attention to if you have the habit of multitasking which most hard core users do (otherwise, what's the point of all this hardware?). When you are running multiple applications at once, you begin to eat up a considerable amount of memory and being able to access your memory in an efficient manner is directly related to your system's performance in situations like this, which is exactly what CC Winstone 2K reproduces.

As we just mentioned, VIA has never been the best at producing chipsets that could outperform their Intel counterparts in memory performance. For this reason we see the FC-PGA on the BX platform outperform, on a clock for clock basis, the same CPU when used on the Apollo Pro 133A (VIA 694X). Even at 825MHz using the 150MHz FSB setting (thus putting the memory at 150MHz), the FC-PGA on the 694X platform can't beat the 550E running at 733MHz on the good ol' BX platform.

SYSMark 2000 approaches benchmarking in a slightly different way than CC Winstone 2K in that it only runs a single application at a time. This results in a slight shift of the performance dependency from memory performance to the rest of the system as a whole (although memory performance is definitely still a factor).

This allows the 825MHz overclock to be pushed to the top of the list with a very impressive SYSMark 2000 score of 163. However on a clock for clock basis the 694X platform is still outperformed by the BX. This would be a definite downside to going with the 694X (Apollo Pro 133A) if it weren't for the fact that the BX chipset, with it's lack of a 1/2 AGP clock divider, isn't as good of a solution for overclocking using 133MHz+ FSB settings.

Realistically speaking, on a BX board with an AGP card not capable of running at 89MHz (133MHz FSB * 2/3 AGP divider = 89MHz) would most likely limit FC-PGA overclockers to speed below 700MHz. When you can hit 733 - 825MHz relatively easily on an official 133MHz FSB platform, the Apollo Pro 133A is definitely worth it, in spite of the fact that it is slower than the BX.

As the test scene moves to 3D gaming, the 133MHz FSB scores from the BX platform are nonexistent due to the inability of our test TNT2 cards to properly work at the increased AGP bus frequency. But on the bright side, the 1/2 AGP divider provided for by the Apollo Pro 133A allows for scores at the 133MHz FSB but also the 150MHz FSB to be presented.

The 550E at 825MHz is loving the attention it's getting because of the apparent lack of competition from the BX chipset. There is virtually no way you're going to be able to run an AGP card reliably at 100MHz on a BX board (150MHz FSB * 2/3 AGP divider = 100MHz), thus making the 825MHz setting we achieved impossible unless you are using a board with official support for the 133MHz FSB.

The 750MHz clock achieved with the 500E using the 150MHz is also very impressive. While our PC133 SDRAM had no problems working at the 150MHz FSB setting, the beauty of the Apollo Pro 133A chipset is that you can run your memory at 33MHz lower than the FSB, meaning at 150MHz your memory could be running as low as 117MHz. This opens up quite a few new overclocking avenues for those with memory that just can't make it at the higher FSB settings.

And if it ends up being that your chip won't handle the 750/825MHz settings, the 140MHz FSB setting leaves the 700/770MHz settings open for experimentation.

As the TNT2 is fill rate limited at the 1024 x 768 x 32 setting, all of the scores end up coming out the same.

We have a similar situation with the 550E at 825MHz running Unreal Tournament. A very impressive score of 41.3 in the UTBench test, one of the highest scores we've seen from an Intel CPU (don't be fooled by the relatively low frame rate, UTBench is a very intensive test). The 500E at 750MHz isn't far behind at 39.8 fps and even the 133MHz FSB settings are faster than the fastest BX settings we were able to run.

Unreal Tournament loves the higher FSB settings offered by our Trinity 400 test bed (Apollo Pro 133A), the 550E at 825MHz is pulling in an easy first place and the 500E at 750MHz, once again, isn't far behind.

Unfortunately if you look at a clock for clock comparison, the Apollo Pro 133A (694X) platform continues to fall behind the old BX.

The scene doesn't change much as we shift to Expendable as the game of choice. Expendable as a test is a very CPU intensive benchmark and thus appreciates the overclocked settings we achieved with the 500E/550E CPUs.

At 1024 x 768 x 32 we are fill rate limited by the TNT2 which means that all of the scores should come out to be the same. This isn't exactly the case, the Apollo Pro 133A test bed was a few tenths off that of the BX, while it's nothing to get worked up about it is very interesting. Perhaps the effects of a poor AGP implementation on VIA's part? or just the memory performance issues we encountered earlier? Or both...?

Conclusion

In the first article we investigated just a small portion of the overclocking potential of Intel's FC-PGA Pentium III 500E/550E based on the 0.18-micron Coppermine core. Now, armed with a true 133MHz FSB chipset, we are beginning to see exactly how far these things can be realistically pushed.

Mind you that our 550MHz to 825MHz overclock wasn't accomplished with supercooling, water cooling, or using an insane amount of fans, we didn't even do so much as use an "overclocker's" heatsink, instead we just stuck to the retail heatsink/fan from Intel (which isn't a bad combo at all) and were blessed with some very impressive results.

As with all overclocking attempts, you are risking the life of your CPU as well as sacrificing stability for the benefit of the increased speed at no added cost. The degree that these two downsides are felt depends entirely on your situation in particular and could vary from severe to virtually nonexistent, as was the case in our tests. It's a disclaimer that we must remind everyone of, but honestly speaking, the FC-PGA 500E/550E are two very impressive processors from an overclocker's standpoint and should have no problem reaching speeds between 700 - 800MHz. But what about the future of FC-PGA Pentium IIIs?

The next FC-PGA part due out will probably be a 600MHz part which, unfortunately, won't boast much higher of an overclocking potential than the 500E/550Es that we've been playing around with. The 6.0x multiplier is still low enough that a 6.0 x 133MHz setting would result in a reasonable 800MHz overclock, but it's just a little too high to be used with something like the 150MHz FSB, at least with current yields on the 0.18-micron process.

One thing we'd like to see is the 160MHz FSB setting implemented on some Apollo Pro 133A boards, when combined with the 5.0x multiplier of the 500E the setting could make an 800MHz overclock (160MHz x 5.0) a reality for 500E owners. As the yields on Intel's 0.18-micron process continue to improve, we can expect to find the new Pentium IIIs making it to even higher levels, but it is highly unlikely that the same 50% overclocking potential (i.e. 550E to 825MHz) will be seen for at least a little while longer.

Keep your eyes open for boards emerging with the Apollo Pro 133A chipset, we'll be taking a look at a few options in the near future and we will make it a point to continue the coverage as more 133A boards are released.