Hi Folks!
Don't recall seeing a thread specifically on this topic so apologies if this is redundant.
Before overclocking and benching new hardware, I've learned that doing a little research up front pays some nice dividends. Leveraging what's already been done by others provides good target states and improves the efficiency of the dialing in hardware and running of the benches process. Having clear ideas of what we're shooting for makes things a lot easier. The process described below works for most any hardware but my focus is primarily on cpus, vgas and ram.
Process Overview:
1) Go to the relevant page at hwbot that shows the cpu, vga, ram or mobo you're planning to bench.
2) For your hardware, look at the top 5-10 results in each benchmark.
3) Make note of all relevant info for your particular hardware/setup including vcore, cpu speed, fsb, bclk, multi, ram speed, ram timings/subtimings, gpu clock, gpu memory clock, driver versions, etc, etc.
4) Identify and make note of subs/settings for those submissions that are noticeably more efficient than those around them (you'll give those extra 'weight' when trying to figure out what you're aiming for as far as settings). If you don't see a super efficient, stand out sub in the top 5-6 places, search down the rankings until you find one then make note of those settings.
5) Identify patterns to determine your target settings. On a macro level, you may decide on a single target state while on the micro level, you may want to have target states for each benchmark (or groups of benchmarks).
Notes:
- If you bought any hardware from other benchers, use that person's subs as targets for your setup. For example, if you got a used cpu then use the prior owner's frequency, vcore, fsb/bclk and multi as your targets.
- If you happen to find screenshots that show additional info (like bios settings, ram subtimings, etc), make note of those, too.
- Discard any results that, due to cooling (or other items like epower boards, etc), may be completely unattainable for your setup. For example, if you're using ambient cooling for your ram, don't bother trying to hit the timings of someone who takes their ram sub-zero.
- If you plan to do 3D, in addition to capturing all the information for your card(s) (including driver used), also check the cpus (and cpu settings) used for each of the 3D benches.
- Due to diminishing returns, I usually look only at a sampling of benches. Researching results for every benchmark takes much longer and, if you picked your sample correctly, doesn't really yield much additional dividends over the sampling.
Process Example:
The research steps and process can be used for basically anything benchable but to give you an idea, here's my research on 4930k's using ln2 on the processor. Line items in BOLD were the ones that appeared to be the most efficient result(s) for each given bench.
So what to make of all these data points? First off, it's very clear that for a 4930k, it's best to shoot for an initial bclk of 125. For vcore under ln2, expect something in the 1.6-1.8v range.
For the ram side of the equation, maximum efficiency seems to be quad channel except for pifast which may prefer dual channel and hwbotprime which looks like it may be better to go single channel (I've excluded triple channel as a consideration for now as these particular setups are known to drop a channel in certain scenarios). After you dial in your ram, the single vs dual (or quad) channel issue can be easily checked by doing two runs for each bench (pifast and hwbotprime) where the only difference is the number of channels used. Excluding the very high efficiency in the 3rd place spi32 sub (that person took their ram cold), target a speed of 2360 - 2600. The best sticks to use appear to be PSC at 8-12-8-28-96 1T. BBSE may also be an option but that would require additional testing. If you don't have any good PSC, the next best would be SAMSUNG's running around 9-12-12-21-108 1T in the 2700 - 2850 range. Since wprime is known to not be very ram dependent, other speeds/timings may provide better results. For example, perhaps HYPERS at 1000 7-7-7-20 (in possibly single or dual channel) or something could be better? This will need to be tested.
Conclusion:
Although far from perfect, doing the research up front can help get setups dialed in quicker and yeild, at least from what I've experienced, improved final results. Of course, every piece of hardware is different so ymmv but at least the process gives you a good idea of what to hope for and expect.
If anyone has additional thoughts, experience, etc...additional input and ideas are very welcomed.
Happy benching and good luck!!!!
-Stefan
Addendum: When you get to your futuremark test runs, compare results with other similar hardware. A lot can be learned from other users' submission details.
Don't recall seeing a thread specifically on this topic so apologies if this is redundant.
Before overclocking and benching new hardware, I've learned that doing a little research up front pays some nice dividends. Leveraging what's already been done by others provides good target states and improves the efficiency of the dialing in hardware and running of the benches process. Having clear ideas of what we're shooting for makes things a lot easier. The process described below works for most any hardware but my focus is primarily on cpus, vgas and ram.
Process Overview:
1) Go to the relevant page at hwbot that shows the cpu, vga, ram or mobo you're planning to bench.
2) For your hardware, look at the top 5-10 results in each benchmark.
3) Make note of all relevant info for your particular hardware/setup including vcore, cpu speed, fsb, bclk, multi, ram speed, ram timings/subtimings, gpu clock, gpu memory clock, driver versions, etc, etc.
4) Identify and make note of subs/settings for those submissions that are noticeably more efficient than those around them (you'll give those extra 'weight' when trying to figure out what you're aiming for as far as settings). If you don't see a super efficient, stand out sub in the top 5-6 places, search down the rankings until you find one then make note of those settings.
5) Identify patterns to determine your target settings. On a macro level, you may decide on a single target state while on the micro level, you may want to have target states for each benchmark (or groups of benchmarks).
Notes:
- If you bought any hardware from other benchers, use that person's subs as targets for your setup. For example, if you got a used cpu then use the prior owner's frequency, vcore, fsb/bclk and multi as your targets.
- If you happen to find screenshots that show additional info (like bios settings, ram subtimings, etc), make note of those, too.
- Discard any results that, due to cooling (or other items like epower boards, etc), may be completely unattainable for your setup. For example, if you're using ambient cooling for your ram, don't bother trying to hit the timings of someone who takes their ram sub-zero.
- If you plan to do 3D, in addition to capturing all the information for your card(s) (including driver used), also check the cpus (and cpu settings) used for each of the 3D benches.
- Due to diminishing returns, I usually look only at a sampling of benches. Researching results for every benchmark takes much longer and, if you picked your sample correctly, doesn't really yield much additional dividends over the sampling.
Process Example:
The research steps and process can be used for basically anything benchable but to give you an idea, here's my research on 4930k's using ln2 on the processor. Line items in BOLD were the ones that appeared to be the most efficient result(s) for each given bench.
So what to make of all these data points? First off, it's very clear that for a 4930k, it's best to shoot for an initial bclk of 125. For vcore under ln2, expect something in the 1.6-1.8v range.
For the ram side of the equation, maximum efficiency seems to be quad channel except for pifast which may prefer dual channel and hwbotprime which looks like it may be better to go single channel (I've excluded triple channel as a consideration for now as these particular setups are known to drop a channel in certain scenarios). After you dial in your ram, the single vs dual (or quad) channel issue can be easily checked by doing two runs for each bench (pifast and hwbotprime) where the only difference is the number of channels used. Excluding the very high efficiency in the 3rd place spi32 sub (that person took their ram cold), target a speed of 2360 - 2600. The best sticks to use appear to be PSC at 8-12-8-28-96 1T. BBSE may also be an option but that would require additional testing. If you don't have any good PSC, the next best would be SAMSUNG's running around 9-12-12-21-108 1T in the 2700 - 2850 range. Since wprime is known to not be very ram dependent, other speeds/timings may provide better results. For example, perhaps HYPERS at 1000 7-7-7-20 (in possibly single or dual channel) or something could be better? This will need to be tested.
Conclusion:
Although far from perfect, doing the research up front can help get setups dialed in quicker and yeild, at least from what I've experienced, improved final results. Of course, every piece of hardware is different so ymmv but at least the process gives you a good idea of what to hope for and expect.
If anyone has additional thoughts, experience, etc...additional input and ideas are very welcomed.
Happy benching and good luck!!!!
-Stefan
Addendum: When you get to your futuremark test runs, compare results with other similar hardware. A lot can be learned from other users' submission details.
Last edited: