#29 Is it exhaustive? Does it check for the 'bad' values for sin, cos, tan, atan, div, sqrt, etc? Or does it just check against the operations and data required for Prime95 to do its thing?
#28, prime95 explicitly tests for rounding errors of the nature you described. It ensures that all 80 bits of the floating point value that are returned are equal to the precalculated value in the program's database.
#20 and #26, yes, those programs can give you *some* sense of security but neither are exhaustive tests. As #26 says, even parts running at their rated/spec'd speed can have problems that just weren't detected by the manufacturor (but this is really rare). You don't necessarily need registered modules, btw... just ECC ones. Registered modules deal with other problems (having enough drive on the bus to operate the modules properly, for example).
As far as returning bad results, some errors can be purely data related. An oversimplified example is that the CPU adds 2+2 and gets 5 (not that this particular example will happen, but there are circuit timings inside the CPU that are data related). Odds are, if you are playing a game, the screen gets a pixel the wrong color or some geometry isn't quite right for a frame but both are too fast to notice. Just remember that 'distance' is the operating parameter of the CPU clock speed. The longest path through the CPU (in a clock driven circuit - which most CPUs are) determines the maximum clock speed. Only one path through one pipeline stage in the whole CPU has to be too long to run at your overclocked speed for the thing to be unstable when that one data+execution occurs.
Anyway, to each his own. Overclock if it gives you pleasure, just don't recommend it as something for someone else to do without giving plenty of disclaimers about it. As I said before, I used to overclock everything but then I decided it really wasn't worth it. Bragging rights just became a non-issue for me and if I needed a faster CPU that bad, I could just buy it and not have to worry about it (nearly as much).
#17- if stability is paramount to you, then you should be using a system with registered parity memory modules as they pretty much guarantee you won't get any errors from them. That's why they are almost invariably used by businesses in mission-critical servers. Anyone who uses unbuffered non-parity modules runs the risk of data corruption very occasionally even if they don't overclock.
Myself, I've used unbuffered non-parity modules for many years because they're cheaper and faster and as far as I know they have never caused me any problems, apart from an incident last year when a memory stick went bad and corrupted lots of important data before the system crashed with a by then all but unrecoverable hard-drive. I hope it won't happen again as it was a nightmare at the time, and I wasn't even overvolting the module (a stick of Crucial/Micron) which went bad.
The only way to be truly safe is with registered parity modules.
zebo's comments are VERY pertinent, also consider
ocz value vx (OCZ4001024WV3DC-K) can still be tempting, considering almost half the price -- the dfi is also a must as most people haven't tested limits of it using max voltage levels of other boards (2.8v?)
also are ballistix cheaper than any available TCCD's ?? -- are ballistix > TCCD or TCCD > Ballistix ??
In all fairness Barkuti, he is testing the memory max capabilites at the highend which is impossible to do w/o some processor variance due to memory variations. i.e. all memory clocks to different levels.
But in general I agree it paints a picture of highend ram as a "have to have" to realize these performance increases when in fact processor speed is playing a more signifigant part.
Nice memory review Wesley. However, there's something "flawed" on it, like in all past memory reviews.
Your measurements for highest CPU/memory performance aren't done right, because you should try to minimize CPU/LDT clockspeed differences between the tested memory platforms - I mean, use the damn memory dividers. There's still a lot of misinformed people about the issue, but you should all know, THERE'S NO PERFORMANCE PENALTY FOR USING MEMORY DIVIDERS ON Athlon 64.
For example, on your past "OCZ VX Revisited: DDR Updates on DFI nForce4" memory review, you settled for 318 MHz on OCZ PC3200 Platinum Rev. 2 modules. At 9x multiplier ratio (1:1 LDT/MEM), that translates into approximately 2862 MHz CPU clockspeed. That was compared to 10x267 MHz for the 4000 VX Gold, which translates into a much lower value of 2670 MHz CPU clockspeed. Despite the incredible disadvantage the VX memory still got a superb result.
But if you had used some dividers to equalize CPU clockspeed, you could have set, assuming 2862 MHz as the absolute top clockspeed for the CPU, the same LDT frecuency and CPU multiplier for the VX modules, and a RAM divider of 5/6; that would have translated into 265 MHz RAM clockspeed, close enough to the max.
The combination of increases in CPU and LDT clockspeeds would have rendered a noticeable increase in top performance for the VX platform, leaving TCCD memory in the dust.
A retest for the not that high clockspeed modules would be nice.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
41 Comments
View All Comments
Joepublic2 - Tuesday, May 17, 2005 - link
#30, I don't know. You could probably get an anwser if you asked at http://mersenneforum.org/.fitten - Tuesday, May 17, 2005 - link
#29 Is it exhaustive? Does it check for the 'bad' values for sin, cos, tan, atan, div, sqrt, etc? Or does it just check against the operations and data required for Prime95 to do its thing?Joepublic2 - Monday, May 16, 2005 - link
#28, prime95 explicitly tests for rounding errors of the nature you described. It ensures that all 80 bits of the floating point value that are returned are equal to the precalculated value in the program's database.fitten - Monday, May 16, 2005 - link
#20 and #26, yes, those programs can give you *some* sense of security but neither are exhaustive tests. As #26 says, even parts running at their rated/spec'd speed can have problems that just weren't detected by the manufacturor (but this is really rare). You don't necessarily need registered modules, btw... just ECC ones. Registered modules deal with other problems (having enough drive on the bus to operate the modules properly, for example).As far as returning bad results, some errors can be purely data related. An oversimplified example is that the CPU adds 2+2 and gets 5 (not that this particular example will happen, but there are circuit timings inside the CPU that are data related). Odds are, if you are playing a game, the screen gets a pixel the wrong color or some geometry isn't quite right for a frame but both are too fast to notice. Just remember that 'distance' is the operating parameter of the CPU clock speed. The longest path through the CPU (in a clock driven circuit - which most CPUs are) determines the maximum clock speed. Only one path through one pipeline stage in the whole CPU has to be too long to run at your overclocked speed for the thing to be unstable when that one data+execution occurs.
Anyway, to each his own. Overclock if it gives you pleasure, just don't recommend it as something for someone else to do without giving plenty of disclaimers about it. As I said before, I used to overclock everything but then I decided it really wasn't worth it. Bragging rights just became a non-issue for me and if I needed a faster CPU that bad, I could just buy it and not have to worry about it (nearly as much).
Zebo - Monday, May 16, 2005 - link
I wonder if they'll sued by redline? I used to have redline bikes as a kid.http://www.redlinebicycles.com/
PrinceGaz - Monday, May 16, 2005 - link
#17- if stability is paramount to you, then you should be using a system with registered parity memory modules as they pretty much guarantee you won't get any errors from them. That's why they are almost invariably used by businesses in mission-critical servers. Anyone who uses unbuffered non-parity modules runs the risk of data corruption very occasionally even if they don't overclock.Myself, I've used unbuffered non-parity modules for many years because they're cheaper and faster and as far as I know they have never caused me any problems, apart from an incident last year when a memory stick went bad and corrupted lots of important data before the system crashed with a by then all but unrecoverable hard-drive. I hope it won't happen again as it was a nightmare at the time, and I wasn't even overvolting the module (a stick of Crucial/Micron) which went bad.
The only way to be truly safe is with registered parity modules.
Zebo - Monday, May 16, 2005 - link
Ballistix is better than TCCD under 255Mhz.. TCCD above that. IMO both are more desirable in that they run low volts any mobo can push.If you want to talk about discount UTT this is where its at: http://shop2.outpost.com/product/4292564
Only $60 a stick. See here for performance.
http://www.xtremesystems.org/forums/showthread.php...
xsilver - Monday, May 16, 2005 - link
zebo's comments are VERY pertinent, also considerocz value vx (OCZ4001024WV3DC-K) can still be tempting, considering almost half the price -- the dfi is also a must as most people haven't tested limits of it using max voltage levels of other boards (2.8v?)
also are ballistix cheaper than any available TCCD's ?? -- are ballistix > TCCD or TCCD > Ballistix ??
Zebo - Monday, May 16, 2005 - link
In all fairness Barkuti, he is testing the memory max capabilites at the highend which is impossible to do w/o some processor variance due to memory variations. i.e. all memory clocks to different levels.But in general I agree it paints a picture of highend ram as a "have to have" to realize these performance increases when in fact processor speed is playing a more signifigant part.
Barkuti - Monday, May 16, 2005 - link
Nice memory review Wesley. However, there's something "flawed" on it, like in all past memory reviews.Your measurements for highest CPU/memory performance aren't done right, because you should try to minimize CPU/LDT clockspeed differences between the tested memory platforms - I mean, use the damn memory dividers. There's still a lot of misinformed people about the issue, but you should all know, THERE'S NO PERFORMANCE PENALTY FOR USING MEMORY DIVIDERS ON Athlon 64.
For example, on your past "OCZ VX Revisited: DDR Updates on DFI nForce4" memory review, you settled for 318 MHz on OCZ PC3200 Platinum Rev. 2 modules. At 9x multiplier ratio (1:1 LDT/MEM), that translates into approximately 2862 MHz CPU clockspeed. That was compared to 10x267 MHz for the 4000 VX Gold, which translates into a much lower value of 2670 MHz CPU clockspeed. Despite the incredible disadvantage the VX memory still got a superb result.
But if you had used some dividers to equalize CPU clockspeed, you could have set, assuming 2862 MHz as the absolute top clockspeed for the CPU, the same LDT frecuency and CPU multiplier for the VX modules, and a RAM divider of 5/6; that would have translated into 265 MHz RAM clockspeed, close enough to the max.
The combination of increases in CPU and LDT clockspeeds would have rendered a noticeable increase in top performance for the VX platform, leaving TCCD memory in the dust.
A retest for the not that high clockspeed modules would be nice.
Cheers