Standard Def Steve wrote:I was never a fan of the K6 line. I recently pulled a K6-2 300MHz (512K L2) system out of the trash and found that the performance was well below that of my PII-300.
One of the first things I noticed about the K6-2 system was that the Windows GUI got a little choppy when I tried playing MP3 and AAC music in the background (streamed over LAN from my file server). The PII-300 had absolutely no problem multitasking with music playing. The PII could also handle DVD playback just fine with GPU assistance; the K6 had trouble even with a Radeon VE.
Definitely sounds like a badly configured system. Super7 systems can be tricky, but with a decent board and CPU and the right driver sets, a K6-2/300 should handle DVD playback just fine. (back in 1998 or '99 Dad had to dig around for Beta drivers to get DVD acceleration working on our PCI RagePro -only AGP versions had official support- but it worked perfectly fine for nearly all DVD playback -very very rarely did you see stuff pushing the max bandwidth of the DVD video spec, then you'd see a bit of stutter in framerate, but that was a PCI bandwidth issue, not CPU performance) The system did get upgraded to a K6-2/500 (actually a 550 but ended up running as a 500) but stuck with that old PCI Rage Pro card. (with FIC 503A motherboard, though we may have had a 503+ at some point too)
Not that AGP on Socket 7 wasn't tricky in its own right ... way worse on ALI chipsets though. Finicky plug n' play support too. (and trying to find good drivers THESE days for these old systems is a huge pain ... ) I do recall getting radeon AGP cards working in Socket 7 systems being a pain back then too, dad has a story when he had usergroup friends balking at him being able to get drivers to work at all for a Radeon 8000 series AllInWonder card. (which we ended up returning anyway due to it running way too hot)
That machine is still running, though now as a Win2000 automotive diagnostic box in the garage.
Then again, none of that is a huge deal performance-wise given our Celeron 366 system (Shuttle Spacewalker ZX board) did perfectly fine for multimedia stuff around the same time.
Getting a good Socket 7 system going needed a lot of experience and either lots of DIY competence and access to information and updates (and usergroup feedback) or a local PC dealer or friend to supply such support. (getting good, reliable, consistent performance out of some VIA based S370 boards could have some of those issues too ... hardware might be rock solid, but without the right drivers and configuration it can easily be a mess -they'll still be SLOWER than BX/ZX or i815 systems of similar quality at similar bus speeds, but they were way cheaper and sometimes more flexible with overclocking -or underclocking and undervolting if you wanted a low power system)
sliderider wrote:
Which is why I don't use benchmarks. How a piece of hardware performs in actual games is the only thing that matters. Benchmarks can be rigged or drivers can be tweaked to give false results or the benchmark itself could be heavily weighted towards functions that you almost never see utilized in actual games.
Hardware+drivers+game in question. Performance could vary vastly depending on the system hardware (including big differences between video accelerator and how they react to a given CPU and motherboard and game ... and drivers and patches). There were (and are) some magic combinations and tweaks that make things work well when the majority of others just tanks, and that goes for quality too, not just raw speed.
So, ideally, you'd build and configure a system around the most bottlenecked application (or combination of applications) you're planning on running and learn the ins and outs for what works well for its particular quirks. 3DNow! and SSE ended up useful in direct3D based games (at least ones targeting DirectX 6 and 7) and a few games had patches that used one or both as well, but it was mostly direct3d stuff that saw vast improvements. (openGL too, unless my memory is failing here)
That's also the reason the K6-2/500 was better for a fair amount of games than a K6-3/400 ... a few ere really cache-sensitive (and in the right range to benefit from that 256k associative cache ... or even combination of board-level cache), but a lot of it was raw computation bound. (be it ALU, FPU, or SIMD execution unit bottlenecked, clock speed will win over cache performance when execution time is the heavy bottleneck) Then there's I/O bound stuff, but if something was really memory latency/bandwidth critical (to the point a 1 or 2 MB board level cache won't mask it) you're going to be at a further disavantage with the poorer chipset performance as well. (comparing 100 MHz systems at least, no overclocking ... even VIA S370/slot1 systems had better throughput/bandwith at 100 MHz, let alone 133 compared to MVP3 SS7 systems -worse for ALI Aladdin, considerably worse for SIS stuff- ... though a 66 MHz celeron or PII could fare worse, especially using an LX or VIA chipset -again, ignoring overclocking here)
And Quake 1 is just an odd bird ... both in accelerated and software render configurations. Performance is all over the map for CPUs of all manufacturers and types with the sole exception of P5. (P5/P54/P55 all have dramatic clock for clock advantages to anything else around, though for some reason the Pentium II Overdrive did even more unusually well in the 686 benchmark test trials)
The Ultimate 686 Benchmark Comparison
download/file.php?id=11658
The vanilla K6-2 manages to match or beat similarly clocked celerons, PIIs, and PIIIs in Quake I but falls way behind the Pentium classic and MMX. (and the Athlon does extremely poorly) The K6 classic does a bit worse in spite of 3DNow! being unused. (the K6-2 seemed to manage some floating point and other improvements that played nice with Quake's quirks -the Winchip 2 seemed to do so as well)
QBiN wrote:
On business/integer benchmarks, everything I see has K6-III's meeting if not besting all intel offerings clock for clock until the Coppermine core offerings. The tables turn when we see FPU benchmarks and 3D gaming benchmarks. K6-III stays competitive if 3DNow is used. However, P-III jumps ahead if SSE optimized is used. I think that's where one sees the Achilles heal of the K6-III architecture. It's FPU performance is behind that of the P-II/P-III architecture and definitely behind the Athlon/Duron processors.
On that note, 'integer performance' is a bit of a misnomer too. It's usually used in a manner of 'general purpose performance' but NOT specific to integer execution and computation. (The Cyrix 6x86 family was actually a bit average in clock for clock integer execution, rather similar to the P6 and K6 line on average -though a decent bit better than the P5 family at least -it was mainly the 6x86's cache and I/O performance and quirks that favored certain applications and user environments that led to its high PR ratings -opposed to the K5 that achieved it though genuinely exceptional integer execution)
Hell, the K6's floating point performance is a lot better on paper than it is in general real-world use too, and quite a few benchmarks reflect that as well. (very competitive whetstone or even more specific floating point execution scores that rarely reflect in real world results for whatever reason -I've seen lack of pipelining as a reason there, not just lack of dual-issue superscalar FPU operation, but single-pipeline performance as well ... presumably tied to the K6's FPU prefetch being too short to keep its execution units fed and working efficiently in parallel with the dual 6-stage integer pipelines -the K6 FPU has much lower execution latency than the PII/III one which does show up in some benchmarks but again, rarely in real-world performance)
Games were ... and continue to be among the most quirky and hardware-configuration-tweak dependent applications around and rarely conform to standardized benchmarks either ... and often have contradictory performance on top of that. Reviews that cite comprehensive hardware tweaks and configurations AND entire suites of games tested are the only really useful benchmarks to go by there.
swaaye wrote:
The chatter is about how K6-3+ can supposedly match a Duron. DVDs included. I spent some time with a Aladdin V ASUS P5A and a K6-3+ 600 a few months ago trying to play DVDs and it wasn't very pleasant. It took some AGP GART tweaking to get it to happen without hardware acceleration and that only worked with a Voodoo3 because other AGP cards became unstable.
So my point is K6 has some problems with DVD playback. I'm challenging nostalgia memory used as evidence.
For computationally bound tasks, a Duron should beat a K6-2/3/whatever at everything, BUT ... for basic single-task DVD playback, ANY duron is overkill as long as the right drivers are there. In raw integer, floating point, and SIMD execution, the K7 core just demolishes the K6. It's cases where large, multi-level, caches and pipeline stalls come into play that the K6-III has an edge. (practically speaking -ignoring oddballs like Quake)
Video and sound driver issues are much more often the cause of stuttering and lag in systems like that, especially onboard audio. (VIA's AC97 compatible audio on MVP4 and Apollo chipsets were ... a bit ugly if you couldn't find the right patches and drivers to make it work right) And lots of fast AGP cards had problems on Socket 7 in general. (Nvidia TNT, Geforce, Radeon, Matrox G200 and G400 all come to mind, many won't work at all in Ali based systems) PCI video cards are usually the safer bet in Socket 7 and work a lot better with plug n' play set-up too. (Rage 128 series cards work decently well in AGP but can be a pain to get initialized -you often have to fiddle around with manually coppying .dll files and such)
I should play around with my AGP Rage Pro in my P5A-B system ... Rage 128 Pro wasn't worth the trouble, but I'd be curious to see if the older Rage Pro gives less trouble. (Win98SE should have built-in drivers to recognize it, avoiding the initialization issues)
But yeah, AGP was one of Socket7's weakest points, and definitely best avoided on anything other than an MVP3/MVP4 based system. (and even then avoided if you could help it) The only real reason to ever use an AGP card in the system was if the PCI counterpart was difficult to find or expensive.
meljor wrote:I love my k6-3+, and i really liked the redhill pages about old cpu's (i've read them all).
But i don't really find their stories usefull when it comes to performance... They PRAISED the cyrix well into the p2 era and dismissed that same p2 as a real option. Same goes for the k6-3 and saying their windows performance only got better with a 1ghz cpu 🤣 Talking about VALUE they are right, as the intel's were way to expensive, but performance wise they had no match, atleast in games. But that's the importatnt thing i guess, as they mostly talk about integer and then the cyrix/amd cpu's had great value.
Red Hill usually seemed to be on about speed of general-purpose desktop use, and had staff extremely well versed in custom-building systems (and particularly with Socket 7), so their experience with Cyrix, AMD, and those overclocked K6-III+ systems (560 MHz, 112 MHz bus, 1 MB L3 cache) seems reasonable. Given the quirks of a system like that it's not too surprising that it took an athlon XP to beat it in every single category. (their story went that a duron replacement set-up failed to manage better average performance while a Thunderbird Athlon was faster on the whole but left a few select weak points).
Multitasking-heavy office/business application environments tended to be the tour de force for Red Hill's assessments. (and indeed for the industry leading Ziff Davis Winstone benchmarks -which are exceedingly difficult to lay hands on these days ... and thus were omitted from the 686 benchmark trials a few years back) Winstone was known for testing much broader working environments than most synthetic benchmarks and most of the old PR ratings were calibrated based on those benchmarks. (the Athlon XP did not use that for its numbering, though -plus it got confusing once Intel had a breadth of processors around ... among other competition)
About 3dmark99: my k6-3+@550 or 600mhz absolutely slaughters my p2-450 (no sse) when it comes to cpu marks. But there is not a single game it can keep up and the intel is MUCH faster. So no, 3dmark99 is not a good test to compare these cpu's.
I'd imagine some games from around 2000 or slightly later would start showing a lot more improvement there, ones using API versions with drivers supporting optimized SSE and 3DNow! use. Maybe some earlier ones, but I'm not sure how many DirectX6.x games were written for it. And the number of games that both rely on CPU driven T&L computation and also support SSE/3DNow! would be a fairly narrow gap. (granted, many games with hardware T&L support can run on cards without it, so those Floating point SIMD extensions come in handy there)
