I've finished AMD's up to the K6-166 Mhz by now and will work on the 200 sometime today. Unfortunately, I am waiting for a K5-133 MHz to be sent to me for testing from a fellow Canadian. My 117 Mhz K5 wouldn't run at 133 MHz, and I don't know any which will. I got my K5 to run almost all the tests at 125 MHz (1.5x83), with the exception of 2 benchmarks, 3dmark99max and Final Reality. Those will have to wait for the K5-133 Mhz as well. I don't have any of the new values on the spreadsheet yet, but I suspect the overall performance of a K5-125 Mhz will match that of the K5-133 MHz due to the 83 Mhz vs. 66 Mhz enhancement.

Now to respond to your comments. Comparing the Cyrix 6x86-120 with the AMD K5-120, I see that for ByteMark's 32-bit DOS FPU calc, they both got 75% of a P90, but in Quake 1 (640x480), the Cyrix scored only 8.2 fps whereas the AMD was awarded 10.1 fps. Looking at WinBench99 - FPUWinmark, the Cyrix scored 247 and AMD scored 388.

For integer operations for ByteMark, the Cyrix got 160% of a P90 and the AMD got 183%. Looking at interger ops using CPUMark99, the Cyrix got 9.74 and the AMD K5 got 9.43. Both chips were run at 2x60.

From the results noted above, the Cyrix 6x86 and K5 seemd fairly equal integer-wise, but the Cyrix fell very behind the K5 for FPU operations. The main weakness of the K5 was in the area of division, ALU and FPU. For division, Cyrix was twice as fast.

Some examples from PassMark at 120 MHz
Cyrix ALU Division: 5.2
AMD ALU Division: 2.5

Cyrix FPU Division: 2.9
AMD FPU Division: 1.5

The Cyrix fell way far behind in all other math tests. For example, Cyrix integer addition: 18.9, AMD K5 was at 28.2.

For the K6 at 133 Mhz, ALU divide 4.1, FPU divide 3.1. That is certainly more than just 13 MHz of improvement. I've been pretty impressed by the K6 numbers so far, however I generally do not like to start making conclusions until I've finished the project.

The major benefit of running these tests with one computer and one person is that all the results are directly comparable. Uncertainties with different motherboard performance, operator errors, etc. have been eliminated.

I have not gotten to the non-MMX Pentium's yet. I only started the AMD K6 tests at 133 Mhz, so we'll have to wait for the 133 Mhz data from the K5's for a 133 Mhz show down. I tried to source my own K5-PR200 for some time but was unsuccessful.

My biggest mistake with this work has been by way of using a Matrox Millennium G200. This particular graphics card didn't have optimised OpenGL drivers for driver version 4.33 and I could only use up to version 4.33 for the non-MMX Cyrix 6x86 chips. For whatever reason, version 5 and 6 of the driver just wouldn't boot into Windows using a non-MMX Cyrix 6x86. Driver version 6.28, however, worked fine on the MMX Cyrix's. What this means is that for 486's and the non-MMX Cyrix chips, the 3DMark99 and Quake2-OPenGL scores are not following the trend. Luckily, I was not going to include the OpenGL-based Quake2 scores, only the software ones for the final charts. I will need to account for the 3DMark99 scores somehow though. Luckily, Direct3D stuff worked fine with Matrox driver 4.33.

It would have been more interesting for me if Cyrix would have created a 200 MHz (4x50) 5x86 for the socket 3. People were still buying 486 boards new in 1997 - I did and so did my friend. I wonder how this would have altered the Cyrix/AMD landscape? Perhaps in an alternate universe, this is what happend?

EDIT: Nice Pentium killer article you found! I've saved it into a PDF.

Interesting notes on the K5, K6, and 6x86 performance.
Division does seem to be something of a Cyrix specialty, both integer than floating-point on the 5x86 and 6x86, even compared to the Pentium. (though the P6 FPU boosted division performance, the MII ended up still in the ballpark, I believe)

feipoa wrote:

It would have been more interesting for me if Cyrix would have created a 200 MHz (4x50) 5x86 for the socket 3. People were still buying 486 boards new in 1997 - I did and so did my friend. I wonder how this would have altered the Cyrix/AMD landscape? Perhaps in an alternate universe, this is what happend?

Yes, supporting socket 3 longer might have made a difference too, and might have given AMD a hard time there rather than letting them take that niche as their own (with the AM5x86).
Short of 200 MHz, they could have pushed 150 (3x50), 160 (4x60), 166 (5x33), and 180 (60x3) parts . . . and made the 133 much more common. The 160 and 166 parts would probably have been the most flexible for older board not working well above 40 MHz.

My impression is that Cyrix was hoping to move out of the budget sector and closer to Intel's high-end market with the 6x86, and priced correspondingly high initially too. They ended up dropping back to "value" level prices to compete with the Pentium though. (only Intel, with their brand name and marketing could manage those high prices/margins at the time)
So, in that sense, it may have made even more sense to keep pushing the 5x86 . . . a smaller, cheaper chip with better yields suited well to the lower-end and mid-range market sectors.

Plus, it may have been practical to produce 5x86 cores for use in both Socket 3 and Socket 5 packages, so they could shift packaging as demand shifted as well. The full 6x86 wouldn't be practical for pushing on both platforms due to the manufacturing cost and power consumption, the latter which was already problematic enough on Socket 5.
Actually, the lower power consumption also should have been a boon for socket 5 acceptance, being more compatible with boards with weaker voltage regulators, some power supplies, and requiring simpler cooling hardware.

I'm also not entirely just how much bigger the 6x86 die was compared to the 5x86. x86guide lists .65 micron 5x86 as 144 mm2 (slightly smaller than the .6 micron P54C) and the .65 micron 6x86 at 394, but the latter seems a it excessive (larger than the .8 micron P5). I've seen this figure elsewhere though, so perhaps it's accurate, but it would mean the 6x86 was more than 2x the size of the 5x86.

Oddly, they also list the 6x86-80 as 210 mm2 (unlike all other .65 micron entries), and list both the .5 and .35 micron 6x86Ls as 169 mm2, while making no mention of the .44 micron version.
Assuming that all 3 of those die sizes are legitimate Cyrix figures, but are somewhat mis-labeled/mismatched, it might be the case that the .65 models were 394 mm2, .5 was 210 mm2, .44 was 169 mm2, and the .35 micron models were smaller still but aren't accounted for. (perhaps around 100 mm2)
Those values would all fit reasonably well for die shrinks of those sizes, but I find that 394 mm2 figure a bit strange still, especially with the 5x86 as a workable alternative.

Hahaha! Are those yours?

feipoa wrote:

Hahaha! Are those yours?

If those were all 433's you'd be soiling yourself. 😁

sliderider wrote:

If those were all 433's you'd be soiling yourself. :happyhappy:

I'd have probably choked on my own drool by now!

I wouldn't mind a NIB MII-366 either, especially if it was the rarer 2.2V variant.

Are any of those the 2.2V version, or just the 2.9V?

The poster of that photo has remained pretty tight lipped, which serves to fuel my curiosity.

Those aren't mine, though I do have 2.2 M2's, 366 and 400, though I bet those boxes were collecting dust there for months.

I clearly remember when these were available to buy, but after the disappointment of not being able to software decode DIVx I gave up on Cyrix and didn't bother on buying the PR433

by the way, this is to show that not everywhere the PR433 existed, it did exist a processor capable of 300Mhz. In fact my mobo (Soyo 5EHM) recognized the 300Mhz as PR350.

Do you know where I can get a boxed Cyrix MII? Who's the owner of those units in the photos?

I find it interesting that the MII-350 is listed as 100x3; this is usually marked as MII-433. It is also curious how MII-333 is 100x2.5; this is usually labaled as MII-366. I've run the 2.2V MII at 333 MHz without any issue.

Pictures are from a shop in Japan, if you look closely you'll find that the Cyrix parts were more expensive than the Celeron. National Semiconductor who manufactured the CPUs at the time sold them nicely packaged through their Japanese subsidiary NSJL, of course with some corrections to their PR as it always should have.

Odd that both 83x3 and 100x2.5 variants were granted the same PR rating of 333. I'd have expected 100x2.5 to be slightly better, but I guess there wasn't another standard Pentium frequency between 300 and 350, other than 333.

feipoa wrote:

Odd that both 83x3 and 100x2.5 variants were granted the same PR rating of 333. I'd have expected 100x2.5 to be slightly better, but I guess there wasn't another standard Pentium frequency between 300 and 350, other than 333.

Also strange that the 233 MHz part is still listed as 300.

By Red Hill's comments on general windows desktop/office application performance, the PR266 (2.5x83) was the last accurately rated part, PR300 parts really should have been 266, the PR333s really should have been 300, and the PR366 (which they never sold) was guessed to be around 333 or maybe 350. (talking K6-2/Celeron/PII performance)

For games/multimedia performance, you'd need an entirely different ratings scale, of course, and the Cyrix chips usually weren't a great value for those tasks compared to Celeron or K6-2, though they fared better against the overpriced PII parts iirc. (price/performance wise) If you take overclocking ability into account, the K6 and Celeron gained a bit more value too.

double post

kool kitty89 wrote:

feipoa wrote:

Odd that both 83x3 and 100x2.5 variants were granted the same PR rating of 333. I'd have expected 100x2.5 to be slightly better, but I guess there wasn't another standard Pentium frequency between 300 and 350, other than 333.

Also strange that the 233 MHz part is still listed as 300.

and the PR366 (which they never sold)

They sold at least two

http://www.ebay.com/itm/Cyrix-MII-366GP-250MH … =item19d2375d2e

I think he was refering to the Redhill Tech. shop as not having sold them.

Wow, $42! Weren't they only like $99 new?

feipoa wrote:

I think he was refering to the Redhill Tech. shop as not having sold them.

Wow, $42! Weren't they only like $99 new?

There's probably that much in gold on that chip. 😳

sliderider wrote:

feipoa wrote:

I think he was refering to the Redhill Tech. shop as not having sold them.

Wow, $42! Weren't they only like $99 new?

There's probably that much in gold on that chip. 😳

From what I recall, the gold-top AMD, Cyrix, and P5/P54 chips have around $15 of gold in them at ideal yields, but taking real-world yields and processing costs (and risks) into account, you'd definitely want to pay significantly less than that to get any sort of profit out of it. Gold-top PPros have about $50 of gold in them, but you're not going to get 100% of that, and then there's the added costs too.

Ceramic/plastic/aluminum top S5/7 chips of the same era have roughly $6-8 gold content iirc. (similar for S370 and Socket A depending on vintage -later chips started using thinner plating, though the higher pin-count counters that somewhat for socket A)

I was looking at the PCSTATS MII-433 review again, and noticed another problem with their testing set-up.
http://www.pcstats.com/articleview.cfm?articleid=1187&page=2

Not only is the P5S-B motherboard used not a particularly good match for a performance comparison, but the test is set up with 96 MB of SDRAM when the SiS 530 chipset in that board only supports 64 MB cacheable with the 512 kB cache used on that board, so those tests were effectively being run without an L2 cache . . . or minimalistic use of the L2. (with the base 32 MB going uncached and only software using RAM beyond that getting cached)

This is in addition to the generally lesser performance of the SiS 530 compared to the top of the line ALI and VIA based SS7 boards, and questionable performance in general of the SiS 530 at SS3 bus speeds (at least Red Hill's site makes a comment about not trusting it beyond 66 MHz). Performance would be worse still if onboard video was in use.

It's also true that the VIA Apollo Pro 133 based Soltek 67KV used with the Celeron wouldn't quite be up to the performance standards of contemporary BX or ZX based systems, but then again it should still have been favorable next to even the better SS7 boards or LX or older VIA Celeron/PII boards. For a sensible comparison of such value/lower-end market hardware, it would make sense to use comparable boards for both CPUs. Although lower-end S370 boards still tended to cost more than the best SS7 boards on the market.
That's omething Red Hill specifically mentioned in the context of the VIA Cyrix III/C3 being a relatively poor value due to motherboard costs compared to a K6-2/III/+ system, in spite of the lower CPU price. (interestingly, this also seems true for current AMD platforms, with a much wider range of low priced motherboards than contemporary Intel socket platforms offer, in spite of the CPU price/performance not hugely favoring AMD over Intel in the low-end)