-
Welcome to Overclockers Forums! Join us to reply in threads, receive reduced ads, and to customize your site experience!
Supplementary SuperPi Competition!
- Thread starter Operant
- Start date
- Thread Starter
- #24
Added the rules. I was trying to think of a way to include the ram speed within the score as well to somewhat show the tweaking efficiency.
Sorry to nitpick here Oper.
Lol this nitpicking is good! It's my first competition so any notes about it are good. Better to learn now so any future competitions will be that much better!
- Joined
- Aug 11, 2008
If you just want to add RAM speed to the efficiency rating, you should be able to just multiply by RAM speed too, making it E = Tsec * Fcpu * Fram. However, this will make the numbers quite large, so you may want to divide by something like one million just to make the numbers smaller and easier to read, so E = (Tsec * Fcpu * Fram) / 1,000,000
The only thing I don't like about just using multiplication in efficiency calculation is that it makes the lower result mean more efficient. Not that it really matters as long as people pay attention, but many people like larger numbers meaning more. So, I'd rather see an inverse thrown in there after the initial calculation to prevent possible confusion, making it something like E = C / (Tsec * Fcpu * Fram). C can be 1 for a straight inverse, but the numbers will be quite small. So, I would use C = 1,000,000,000 or something to make them more manageable.
- Joined
- Aug 11, 2008
Right now I'm working with variations of Tsec * Fcpu * Fram * 1/Cas * 1/Trcd * 1/Trp * 1/Tras
I inversed the timings before multiplication since, unlike with frequency where lower time + lower freq is more efficient, with timings it's lower time + higher timings = more efficient.
The numbers look to be resulting correctly in simple examples when only varying the timings. When varying both RAM freq and timings, that's where a gray area could come in. To me, the numbers seem reasonable when varying both freq and timings as well. It would be best to apply scaling coefficients to each variable accounting for how much each variable contributes to the superpi time. I'm not a superpi guru to know how much each variable contributes...any suggestions?
Is there anything that jumps out as needing a major fix from this sheet?
NOTE: first efficiency column is "lower = better", second efficiency column is "higher = better", and the third is exactly like the first except scaled down.
I inversed the timings before multiplication since, unlike with frequency where lower time + lower freq is more efficient, with timings it's lower time + higher timings = more efficient.
The numbers look to be resulting correctly in simple examples when only varying the timings. When varying both RAM freq and timings, that's where a gray area could come in. To me, the numbers seem reasonable when varying both freq and timings as well. It would be best to apply scaling coefficients to each variable accounting for how much each variable contributes to the superpi time. I'm not a superpi guru to know how much each variable contributes...any suggestions?
Is there anything that jumps out as needing a major fix from this sheet?
NOTE: first efficiency column is "lower = better", second efficiency column is "higher = better", and the third is exactly like the first except scaled down.
- Joined
- May 10, 2009
The issue is that each variable increases or decreases an amount determined by all the other variables, plus all the tweaks.
The beauty of the classic method is that if you take a time done at 100bclk (or 200fsb, or whatever) and increase the base clock by 10%, the score can be predicted with high accuracy working backwards from the 100bclk efficiency number. Note that it does not work if you change variables other than baseclock of course.
That's what makes it a viable efficiency number, it can be used to predict scores. It was more useful when the base clock could vary more (FSB), but it still works now.
Almost regardless of the formula you could look to have extremely good efficiency in a RAM timing based model by running one timing way high. The SPI32m calculation time won't change much, but the efficiency number will go way up.
I realize I'm coming off as a naysayer, but IMO there is simply no way to include the RAM in spi32m efficiency.
Dropping TRCD from 11 to 9 will do different things at 4GHz than it will at 5GHz, it'll do different things at 2133 than it will at 2200, it'll do different things on an IB than than a SB, it'll do different things on a 3570k than a 3770k even (L3 cache size, more L3 cache = memory less important).
What you're going to need to do is test every combination of ram multipliers and timings at a given CPU speed and tweak level, and try to come up with a predictive model.
Then you'll need to do it at other CPU speeds as well as with other CPUs.
AMD and 1156/1366 will be even worse, as the L3 cache and IMC speed isn't locked to core speed like it is on 1155 or to NB strap like it is on 775.
The beauty of the classic method is that if you take a time done at 100bclk (or 200fsb, or whatever) and increase the base clock by 10%, the score can be predicted with high accuracy working backwards from the 100bclk efficiency number. Note that it does not work if you change variables other than baseclock of course.
That's what makes it a viable efficiency number, it can be used to predict scores. It was more useful when the base clock could vary more (FSB), but it still works now.
Almost regardless of the formula you could look to have extremely good efficiency in a RAM timing based model by running one timing way high. The SPI32m calculation time won't change much, but the efficiency number will go way up.
I realize I'm coming off as a naysayer, but IMO there is simply no way to include the RAM in spi32m efficiency.
Dropping TRCD from 11 to 9 will do different things at 4GHz than it will at 5GHz, it'll do different things at 2133 than it will at 2200, it'll do different things on an IB than than a SB, it'll do different things on a 3570k than a 3770k even (L3 cache size, more L3 cache = memory less important).
What you're going to need to do is test every combination of ram multipliers and timings at a given CPU speed and tweak level, and try to come up with a predictive model.
Then you'll need to do it at other CPU speeds as well as with other CPUs.
AMD and 1156/1366 will be even worse, as the L3 cache and IMC speed isn't locked to core speed like it is on 1155 or to NB strap like it is on 775.
- Joined
- Feb 1, 2009
- Location
- Illinois
I like the rule changes. I kind of opened a can of worms when I said Xs used a 3mhz +/- 
But the way it was clarified works nicely. Any velocity up to 5002.00mhz is acceptable, any over is not. The 2mhz is alloted due to the floating bclk (my bclk is set to 100 in the bios, but boots at 100.01 in Windows). I like these rules. When it comes to efficiency, it would be nice to have the ram and such in there, but it becomes really complicated.
But the way it was clarified works nicely. Any velocity up to 5002.00mhz is acceptable, any over is not. The 2mhz is alloted due to the floating bclk (my bclk is set to 100 in the bios, but boots at 100.01 in Windows). I like these rules. When it comes to efficiency, it would be nice to have the ram and such in there, but it becomes really complicated.
- Joined
- Feb 1, 2009
- Location
- Illinois
I guess you could include them, but when it comes down to it, the most important aspect is usually speed...with relation to the timings. And that is the difficult relationship to pin down, as Ed explained.
- Joined
- Aug 11, 2008
Yeah, the classic method is the very good and simple prediction formula since CPU speed contributes the most to the time. Adding other things will always make it more accurate, but may not be worth the effort. Especially when considering the results from testing across platforms would have to be averaged before implementing into a single formula (which could increase error range), or there would need to be a different formula for every platform at the very least.
I think we're getting way out of the scope of what I was trying to do. I wasn't trying to make a predictive model, I was just trying to come up with something fairly simple that can provide a general idea of who is being more efficient in this competition while taking RAM settings into account since I saw people thought it would be nice.
So, I gave it a shot witout going extremely math intensive. All I had in my mind was that the resulting numbers just needed to differentiate between submissions with varying RAM speed and timings (basically, given a specific time, the submissions with the worst settings reaching that time would be most "efficient" or get a better score), not that it had to be a highly accurate predictive model...
I know the numbers aren't as accurate as they could be since no meaningful scalars were used for the variables, and the time intensive testing Bobnova mentioned would be needed to get an idea of meaningful scalars.
I think we're getting way out of the scope of what I was trying to do. I wasn't trying to make a predictive model, I was just trying to come up with something fairly simple that can provide a general idea of who is being more efficient in this competition while taking RAM settings into account since I saw people thought it would be nice.
So, I gave it a shot witout going extremely math intensive. All I had in my mind was that the resulting numbers just needed to differentiate between submissions with varying RAM speed and timings (basically, given a specific time, the submissions with the worst settings reaching that time would be most "efficient" or get a better score), not that it had to be a highly accurate predictive model...
I know the numbers aren't as accurate as they could be since no meaningful scalars were used for the variables, and the time intensive testing Bobnova mentioned would be needed to get an idea of meaningful scalars.
Last edited:
- Joined
- Apr 3, 2012
- Location
- Barberton, OH
What if you make separate leaderboards since this is a sort of Low Clock Challenge?
Intel Ivy
Intel Sandy
AMD Phenom II
AMD FX/Trinity
I don't see people submitting with anything else except maybe Llano, but 3.6 will be able all Llano will do without LN2 as well. Older Bloomfield and Lynnfield i7s won't get near 5 GHz either.
The best Ivy will do at 5 GHz with RAM on air is around 6:20, Sandy 6:30, if anyone is looking for a target.
Intel Ivy
Intel Sandy
AMD Phenom II
AMD FX/Trinity
I don't see people submitting with anything else except maybe Llano, but 3.6 will be able all Llano will do without LN2 as well. Older Bloomfield and Lynnfield i7s won't get near 5 GHz either.
The best Ivy will do at 5 GHz with RAM on air is around 6:20, Sandy 6:30, if anyone is looking for a target.
- Thread Starter
- #38
If you feel it would fit better there, go for it.
This is a general competition through so you don't HAVE to compete in the overclocktagon bonanza if you compete in this.
Similar threads
