What have I done right?
I have an old AthlonXP-2600+ running at 2.078 Ghz which I was assured would do 2.6Ghz, but actually doesn't and 2.178 Ghz is what we settled on. I have 266mhz ram on a board with 333mhz FSB and we sorted that out also, and specified the ram speed.
That sounds like under 10% of a speed increase, if any. But the previous kernel I built here took 55 minutes (time make -j4). The one I built after the mods (nearly same config) took 33 minutes. (make clean; time make -j4) 40%? I've clearly improved or fixed something. But what? I'm a hardware techie and it's made me dead curious. Code:
bash-4.1$ sudo grep BogoMIPS /var/log/messages |tail Code:
May 16 17:15:06 genius kernel: Calibrating delay loop (skipped), value calculated using timer frequency.. 4361.15 BogoMIPS (lpj=2180575) |
As a hardware techie, surely you are aware that clock speed has absolutely zero to do with processor performance; it is a very relative concept and totally worthless for benchmarking.
As for what kernel modifications you made specifically to increase performance, we are unable to say as you have not told us what was changed in this new kernel configuration. There are many possible optimizations that can be made in the kernel config, and if your previous kernel was the stock one that came with your distribution, just compiling the new kernel for your specific processor type rather than the x86 generic it originally was could be responsible for most of the performance increase. |
I changed nothing in the new kernel! I think I made 2 small changes. I didn't remove anything meaningful or large.That's why my jaw sagged. It was a test of the box, nothing else.
|
Athlon 2600+ (T-bred) @ 2.6GHz? Maybe on water or other exotic cooling, but not on air in my experience. About the most I recall seeing from an air cooled 2600+ T-bred is in the order of 2400MHz, with 2300MHz or so being far more common.
2.078GHz is just wrong for the 2600+ T-bred. It should have been at 2133MHz 'offically', though its normally more like 2128MHz (133x16). The 2133MHz number is 133.33x16. Were you previously running 130x16 (well, 129.875x16) to get 2078MHz? I'd guess that you are using a FSB overclock to get 2178MHz (136.125x16). Thats only a few % faster than stock, and a few more % compared to 130x16 but the extra FSB should have more of an impact than the numbers might make you guess....I doubt that you would see anything like the increase in performance you got from playing with the FSB and core MHz like that. IMO your performance increase would be in part due to moving from underclocking slightly to overclocking, but probably more due to the kernel mods. Without more info though I'm guessing.... BTW, I have run a LOT of Athlon XPs. Mostly a T-bred 2200+, offically 133x13.5 (1795MHz, 133.33x13.5 is 1800MHz). I found it ran noticeably faster at 166x11 (1826MHz). If you have an 'unlocked' T-bred, or have an nForce2 board that unlocks automagically, and can find some DDR-333 it would be worth trying 166x13 (2158MHz). I found that running asymmetric FSB/RAM with the RAM slower (333FSB with 266MHz RAM, etc) works out to be slower overall than running symmetric (266FSB with 266MHz RAM, etc..). I also found that 266FSB with 333MHz RAM, etc, gave you virtually no increase in performance. Quote:
Aside from that, I semi-disagree. You cant just bung up the core MHz 10% and expect 10% more preformance, which is what a lot of people expect. Still, more MHz normally helps CPU performance, up to a point. Try pushing any CPU to the 'absolute max' and you tend to see it will run slower than it would at lower MHz values, for various reasons. |
Quote:
Yeah, the salesman was a cocky noob, and full of crap. I didn't care really, as long as it was cheap :-/. They had issues with motherboards not reporting speeds correctly, and were changing bios versions wholesale. Then when I blew that old m/b I kept cpu & ram, and just bought whatever would run them that wasn't a via chipset. Quote:
Quote:
Code:
bash-4.1$ cat /proc/cpuinfo Quote:
My only guess was that my ram was being clocked at 166Mhz, which could have meant extra wait states, but it was faster in wine (with a WINEDEBUG trace) than one core of a twin core turion I have here running at 2.0Ghz, so I can't have been losing much. So I presumed human error. MY human error. I ran that kernel build again - make clean; time make (without the j4) and it built in _30_minutes 18.301 seconds :-O. Faster again. Maybe that's not hugely surprising, as we are challenged for ram (1 gig). But it's high time I followed rule number one as a techie 1. If it works, don't fix it! This problem also qualifies for exclusion under rule no. 2 2. If it's not worth it, don't fix it. Thanks for all thoughts, guesses facts, and factoids. |
Quote:
BTW, one of the old 'rules of thumb' I've seen for overclocking is "push the CPU as far as it will go and still boot to windows, then drop back a speed step or two". Its not infalible, but generally it makes sense. Quote:
Another non-fan of VIA chipsets? I'd probably take one over a SiS chipset for the Athlons, if I didnt know what I do now. All the KT266a/KT333/KT400 boards I had, or other people I know had, are dead now. I've still got a running nForce1 (asus A7N266) and a SiS 735 (a ECS K7S5A of all things). The A7N266 I pulled out of a plastic bag from under a bench at a place I used to work. A music shop, not a computer shop. LOL Quote:
IIRC the 166/333MHz FSB t-bred and thortons did have a lot of issues with some chipsets and speeds not being reported right, no matter what you did. You can change BIOS versions till the cows come home (or if you have a bodyclock like mine, change BIOS versions till the cows leave the shed) and still not have things work just the way they should. BTW, the thortons are just a barton with 256K cache disabled. Or thats what I heard anyway, its pretty hard to pick the difference between thorton and t-bred, you have to check the model numbers on the actual chip. If its a thorton, that would explain how you managed to get 2333 MHz, I'd doubt that you would see 2333MHz with t-bred on a SiS chipset board, they tended to be fairly lame overclockers. 2180MHz, your board is running a higher than standard FSB- 167.7MHz, or close enough to that it doesnt matter. But then you should have been getting 2096MHz @ 167.7x12.5..... I dont know. If it is running a slightly higher than standard FSB clock, its not that common with Athlon XP boards, it was more common in the original Athlon and P3 boards. I'm cynical enough to think that the manufacturers bumped the speed slightly to make the boards benchmark better. If you dont mind me asking, what board is it? Quote:
If it was me, I'd be poking that system.....a lot....to try to figure out whats is happening. But I've got Socket A motherboards hanging around which makes testing easier, and I'm a stubborn bastard when it comes to hardware I consider myself to know fairly well. For any normal, sane person rule #1 and #2 should come into play- the system is working, its not worth trying to 'fix' it. I'd probably just chalk it up to old hardware and (sorry) cheap chipsets being quite variable at times. The ECS K7S5A I have runs well, but every now and again it will for no apparent reason run really slowly. Like 'P3-733' slowly when running a 2200+. Rebooting normally fixes that, and if it doesnt turning it off for a while does. |
Quote:
Now that 166 Mhz shown in the bios is something I can't set. I haver (from memory) CPU SPEED - By jumper ACTUAL SPEED - - 166Mhz(Greyed out) DRAM SPEED - - 133Mhz What that 166 Mhz drives is anyone's guess.There's all sorts of caveats about overclocking, but on every page there's mention of it, so they presume you will anyhow. |
Yeah, remember that for AMD that 2600+ means equivalent to Intel 2.6 Ghz, or something like that. Either way the CPU freq is not very important nowadays, other things are more important, like cache, FSB, etc.
|
Quote:
The 'premium' chipsets and motherboards were capable of 200/400MHz FSB, the 'consumer' level chipsets (which were really just the same chip with different markings) and motherboards were limited to 166/333FSB. Quote:
As for the overclockign everywhere in the manual, thats fairly typical again of late model socket A boards. By that time a lot of users were moving to Athon XPs for overclocking, and the manufacturers figure that techies and 'normal' system builders would just ignore the overclocking options. But if they didnt have some way of overclockign it could hurt sales in various ways. Quote:
Sempron XXXX+ = equivalent to a P4 celeron of XXXX MHz. That is why a tbred Athlon XP 2200+ is 1800MHz, a Sempron 2200+ is 1500MHz (mind you, the Sempy is 333FSB, the Athlon is 266FSB, but lets not go into that now LOL) Quote:
|
Quote:
|
Quote:
Who knows what bulldozer will do though, it could well pass the i7s. The leaked (haha) benchies look pretty good. |
I'm amazed to see this one go so far.
@H_TeXMeX_H: I'll note that you would run an i7 for performance; we are in the bicycle territory here, not sports car. I bought one of the few motherboards that were 1) Socket A 2) Agp for my video card - a throwaway from a gaming pc & 3) did not have a via chipset. I think that makes clear where I stood on performance. And temporarily _I_have_ the cash to buy a stupidly 'better' box. @cascade9: You're dead right about Asrock making a jumperless board. The Design shows signs of a tug of war. 166 is actually fsb (found it) and set from jumpers. Multiplier is jumperless and automatic, but if you set the bios to "By Jumper" you can use the jumpers to set things. They don't supply the jumpers - I had to. |
Blame the self-confessed 'borderline athlon XP anorak' for how long this thread has run for. :D
|
Well it pushed me to have a look at Intel's i7. For me the telling spec is 32nm against AMD's 45nm.
All the major logic families were designed for 2.0um fab. We have come a long way. |
Interesting, the very last thing I look at is process sizing. I guess everybody checks different specs.
Intel always gets the newer fab processes out the door faster. Mostly because they get the needed equiment before AMD. Bulldozer will be 32nm from everything I've seen, but until they are released its hard to know for sure. |
Of course, everything new will be 32nm or smaller. There's no way back, because the performance would drop.
I deplore this renaming and remarketing to excess. There are 4 basic states in any cpu cycle. Address instruction, Read instruction, decode, & execute. There is a limit to what read ahead and cache can do. The fab size (32nm vs 45nm) feeds back into reduced capacitance, which greatly lessens current consumed, and also increases speed. 2 formulas that matter are: 0.5 x C x Ve2 for the energy stored in a capacitor (C in farads, answer in joules) V x C x Fe2 (or something) for the energy expenditure clocking a capacitor. Gates for your standard mosfet taking any power can be 15pF (15e-12F). It's not hard to see massive energy consumption in a cpu this way. |
Quote:
Unless you are count the power consumption allowing more transitors, or higher clocks, but even they can have little effect. See 130nm P4 'northwood' vs 90nm P4 'Prescott', even with a dieshrink there was very little performance gains (but to be totally honest, it wasnt just a dieshink, there were various changes) Quote:
I got used to marketoids renaming chips years ago. Pentium Pro, opps, that flopped, lets repackage and rebadge it as Pentium II.... |
Again, I look for FSB freq and cache size, which I think are most important nowadays. The CPU freq is less important, because CPUs are so powerful now that it is more important to get data to them fast, so they can process it. This is where cache size and FSB come in. When I buy a processor I look first at the highest FSB the mobo supports, then choose the one with the highest cache size, as long as it is affordable. Then, of course, I check the mobo manual or online info to make sure it actually supports that CPU.
|
Quote:
Quote:
The Pentium 4 had a "532" mhz hypertransport but it's notional width was 8 bits, so when you boiled it down, it was switching 32 bits at 532/4 = 133 Mhz. My laptop has a "667" mhz bus. We're 64 bits wide, and I believe a divide by 4 is applicable again, which is in fact 166.75 Mhz bus speed. I see 1333 Mhz stuff out there now. I refuse to believe that speeds have doubled, until someone who sees through all the b.s. tells me so. Marketing hype has doubled. There are several good reasons for this skepticism. Without getting overly technical, take track length. The address/read cycles are where the ram is read. On one clock cycle, the address will be good by the end of it; on the next, ideally, the data will be good, and can be read back. At 1333Mhz, electricity can only travel a maximum of 22.5cm (3e8 / 1.333e9), so distance across the board becomes a significant delay factor. Back in the 80s, I worked temporarily in Amdahl repairing machinery that made their cpu units for the huge mainframes they built. One of my pains in the neck was a clock wire machine, which gave out wire of a specific length. The clocks (~125Mhz) all had to be wired with wires of specified length (+/- 0.1 in. or 25mm). The next model (after they fired me) was water cooled. These had 6-9 boards of 11 x 11 SMT ECL logic chips with huge cooling towers, and ran from a 5V 400A(!) power supply. Fab does make a difference! |
Ok, it has been a while since you checked all the marketoid junk hasnt it? (no offence intended).
AFAIK, Intel has never used hypertransport, QPI is the current Intel equivilent to hypertransport. You are right about the P4 FSB BTW. Though there were also 400MHz, 800MHz and 1066MHz P4 FSB models. I cant recall what the (real) memory speed is on the 1066MHz FSB models (they might be using DDR2, see below), but for the 400, 533 and 800MHz models the 'real' memory speed is 1/4 of the FSB (due to quad-pumping). Since hypertransport and QPI, FSB doesnt really exist. When you see '1333MHz' its the memory speed. BTW, 1333MHz is almost low now, 1600MHz is more common, and 2400MHz is now on the shelves. But that number, like all DDR numbers, is semi-marketing. The real core speeds of memory hasnt changed much in years now.... DDR1 200-400MHz divide by 2. So 'DDR200' has a real core speed of 100MHz, and the most common DDR400 is 200MHz real core speed. BTW, DDR1 got at least as high as 550 (275MHz) but its rare to find DDR RAM with higher speeds than DDR-400. DDR2 400-1066MHz divide by 4 (same as DDR1, but with a x2 bus clock multiplier). So DDR2-400 is back to a real core speed of 100MHz, DDR2-800 is 200MHZ, DDR2-1066 is 266MHz. DDR2 went at least as high as 1200, but again, thats rare. DDR3 800-2400MHz divide by 8 (same as DDR1 but with a x4 bus clock multiplier). DDR3-800 is again 100MHZ real core speed, the most common DDR3-1600 is 200MHz and the fastest you can get now is DDR3-2400, 300MHZ. Speeds....difficult to say that they have doubled. Bandwidth sure has, but bandwidth does not always equal speed.... As for the technical data sheets, yeah, they did start dumbing them down years ago. Yes, I agree, it was in part because the technical data sheets would have made the marketoids look a bit silly. Using the same 130nm P4 'Northwood' vs 90nm P4 'Prescott example I used above, you can see why they didnt release a full data sheet- Northwood- 8KB L1, 512K L2 cache. Prescott- 16KB L1, 1024KB L2 cache. (FSB varies, but you can get the same FSB with Prescott as with Northwood). So if you just check the 'basic' specs, Presscott looks like a clear winner. But when you start looking closer, its not so clear- Northwood- L1 cache latency 2 clocks, L2 cache latency 1 clock, 256bit L2 bus. Prescott- L1 cache latency 4 clocks, L2 cache latency 2 clocks, 128bit L2 bus (probably). If intel had of put all that into a technical data sheet, it would have been much more clear exactly what differences there is between the Northwood and Prescott CPUs. Which is nowhere near what you would expect if you just checked the basic specs, which is what the marketoids were pushing. |
I'm pretty sure that with DDR2, anything beyond 800 Mhz requires overclocking. At least that's what I've heard and experienced.
|
Quote:
Quote:
Motherboard brand and chipset was a random choice. DDR2-1066 support is from the intergrated memory controller (though it needs a comptible chipset to work of course) and is AM2+ only. Well, it might be true with AM3 as well, I've never checked. Quote:
I'm not 100% sure on intel, though I think that to use DDR2-1066 and higher is technically overclocking with the intel chipsets. |
I'm putting a "Solved" on this one, because we are pretty much in agreement. We're actually a long way from where we started, but who cares? It had been an interesting and indeed educational discussion of cpus in general, and I thank you guys for condensing your hours wasted assessing Intel and AMD's Panglossian marketoid blurb so astutely. You have reinforced my resolve not to waste similar time myself, and verified my decision to desist in the first place. (Pangloss, btw, was a hopeless optimist in Candide- "This is the best of all possible worlds").
I ran an i586 for years, doing everything I do today and more, and although I felt the need for an upgrade, I didn't benefit much from it when it came. I remember what people did on CP/M (Z80@4Mhz, 64k of ram). These days I see people doing pretty much everything with little arm based netbooks and a tiny ssd. The OLPC is hugely impressive for what it is.I'll read the marketing factoids when I am specifying a server, or a gaming machine. |
Quote:
Glad is wasnt totally boring, or you're polite enough to overlook the boredom factor. Quote:
Quote:
For gaming machines, its easier just to check you budget, then get a few different sourced benhmarks. Aside from 'dodgy' hardware a few different benchmarks and soem thinking time is all you need for gaming setups. (dodgy hardware, intel and AMD have both done it, though intels done it worse than AMD. Nobody is going to put "we were pushed to get this hardware out the door and it could have issues" in the marketing blurb, so its besides the point anyway.... |
Quote:
Quote:
EDIT: Now that I think about it, everyone rewrote that or bits of it, and did it better than the original guy. Voltaire certainly did, and it was spun into 'the optimist.' |
Quote:
I'll take the boring pintless and stupid warning, I'll take it under advisment. Voltaire is a satirical polemicist, so I can se where an overview of one of his plays could sound pretty awful. If its as bad as advertised, I'll drop it and read something else. Thats if I can find a copy of Candide in the library, I dont like reading non-technical documents from a scree. n Quote:
|
I apologize if this seems like "hijacking", but as I'm reading through this thread, the impression I get is along the lines of "don't believe any specs that you read online", because chances are that they're always leaving something out to deceive you. Great, so when I (maybe) do my own desktop build, it's gonna be even harder to look up specs for a decent machine because I'll have to look up specs far beyond what is listed on the site or whatever (and WTF am I supposed to find this extra info??). :(
I'm clearly a dumbass for having not taken into account things like "double/quad-pumping" (taking twice/four times the amount of clock cycles to get data across the FSB), etc. :cry: |
MrCode, dont sweat it!
I dont totally agree with business_kid, I'm a bit less harsh. You can belive the specs, noone who wants to stay in business tells outright lies about them. When Intel say 'P4 Prescott 1MB of L2 cache' it is the truth. They just left out the (probable) fact that the L2 cache on Prescott is in some ways slower than it was with Northwood. You can put that down to 'marketing' or if you want to be charitable you can say that nobody really cares so intel just omitted to tell anybody... Where it gets nasiter is when you start looking at the marketing factoids. I'l give you an example- Quote:
Fudzilla (and most other sites) have added this bit- Quote:
(unless there is some 'in 3D mark' disclaimer in the fine print down the bottom of the graphics intel made, I cant find a better copy where the text is readable). It actually is about 3x faster in 3Dmark- http://www.pcstats.com/articleview.c...id=2343&page=8 But that doesnt equate to 3x faster performance in games, not even close. Scroll down on the pcstats link and you will see that '3x faster' becomes 'less than 2x faster' in performance for at least one game, and I'd suspect that would be similar over many games. So if you just check the marketing junk, you could well think that the G45 will give you framerates 3x faster than G33, which just isnt true. There is also the way that people mislead themselves- I've heard many people talking about upgrading from a SATAI (1.5Gbit/s) to a motherboard/controller with SATAII (3Gbit/s) or SATAIII (6Gbit/s) support 'because my HDD will be 2-4times faster". Nope, thats not true, mechanical HDDs are JUST starting to saturate SATAI bandwidth. No mechanical SATA HDD will get close to filling SATAII bandwidth, let alone SATAIII. About the only time that you extra bandwidth will make a difference is when you are copying a file over that will fit into the HDD cache. Worring about 'hidden' specs isnt really worth it. Using a similar example to the P4 L2 cache above but something that is current, really not that many people even care that there are minor differences in latency between versions of the i7 (and i3/i5 as well for that matter). For the engineers working on CPUs, it matter, for an end user its at best a point of intrest. There is info around on minor, mainly esoteric differences like the cache latency, but even finding it can be a major quest. Then you have to decide if the info is even trustable... Dont worry about it, if you ever do go to biuld your own computer avoid the Intel, AMD, ATI, nVidia, WD, Samsung, Seagate, Hitachi, etc. marketing, and trust decent benchmarking, user reviews, even semi-random forum posts over marketing blurbs. I wouldnt call you a 'dumbass' for not paying attention to pumping. I know people who spec and build computer who dont know anything about it, and they %&#$%&$ should know. As for "hijacking", I wouldnt worry about it. From the posts I've sen by business_kid, they are more than able to tell both of us to stop mucking up the thread, and do it a polite way as well. Its marked 'solved' anyway, even though I'm still at a loss at to why there was that huge jump in performance. ;) |
@cascade9
Thanks for the helpful advice. :) Honestly, if I'm looking at computer hardware at all, I try to look up the specifics. For example, if I'm looking at a video card/GPU, I'll look for stuff like the clock speed(s, if applicable, e.g. shader vs. core clock, etc.), memory type/amount, etc.…I wouldn't trust something like "3x faster than the previous model" right away without doing a bit of technical research first. :p Most of what I was concerned about was whether I do enough "looking" to really try to find a decent piece of hardware. WRT the whole SATA thing: that actually makes a lot of sense; a mechanical HDD's speed is mostly dependent on both its platter RPM and the actual transfer rate (AFAIK). If the transfer rate is blazing fast, but the platters can only spin so fast (and hence data can only be put into the HDD cache so fast), then you're not going to be able to use all that bandwidth. Quote:
Well, I suppose I can rest relatively easy now. I guess it just goes to show that I've still probably got plenty to learn about the specifics. :p |
Quote:
There is always an element of luck in hardware IMO. Back when Socket A was current, I would never have guessed that 'low end' motherboards with SiS chipsets would outlast many of teh more 'top end' socket A motherboards with VIA chipsets. But from my experience, that it seems to be the case that (for socket a at least) cheap-assed SiS chipset can last a long time, longer than the average VIA chipset. The long term lifespan, which I consider to be _VERY_ important, is at best an educated guess though. :( Aside from lifespan, you really wont notice much difference between chipsets/motherboards/CPUs/RAM of the same type. Boards using chipset 'XXXX' with CPU 'YYYY' and RAM 'ZZZZ' are normally in an performance envolope of 1-2.5% in general (normally all parts of that type, no matter the manufacturer, will be within 1-2.5%). Aside from things like 'factory OC' models, the same is true of video cards. You've also got to consider 'what do you need it for'. The number of times I've seen people buy expensive 'gaming' video cards because of flawed logic like 'it costs twice as much its got to be twice as good'. They would have been a lot better off getting a 'normal' dekstop card and spending the extra money on a SSD. Mismatched tasks and hardware is not only a waste of money, it can affect lifespan of the system, due to things like 'gaming' cards having higher power requirements and more heat output. That sort of mismatching is easy to avoid as long as you dont fall into the marketing trap, and think about what you would be using the computer for. Quote:
|
Quote:
Quote:
|
Quote:
Quote:
|
All times are GMT -5. The time now is 02:55 PM. |