Memory Timing Benchmark (CL/tRCD/tRP/tRAS/CR/tRC/tRFC/tREF)

Hello. it's newbie on this forum.
I'm not sure that whether I can upload this kind of bench on this board. (this board seems to consist of reviews of "a certain model of" HW... am I right?)
Please announce me whenever I (unintentionally of course ;-)) forbid the rules of this forum.


...okay. then let's back to the main context.
I benched my memory with several different ram timing settings. Tests include:
1. Super PI 32M digits of Pi calculation
2. Lavalys Everest Cache & Memory Benchmark
3. Video Transcoding (1080p TP -> mkv using H.264)

and tested ram timing items are:
1. CL: 6~11
2. tRCD: 9~11
3. tRP: 6~11
4. tRAS: 24~30
5. Command Rate: 1 / 2
6. tRC: 15~40
7. tRFC: 90ns / 110ns / 160ns / 300ns 350ns
8. tREF: 3.9ms / 7.8ms

(for the test, except tRC and tREF, the lowest value is default. tRC and tREF have their highest value as default)

and other system setup is:

http://image.playwares.com/bbs/data/...e61c8cdf7d.jpg

- CPU: AMD Phenom II X6 @ 4.0GHz / NB 2.75GHz
- M/B: ASUS M4A89GTD-PRO USB3
- RAM: G.SKILL PIS PC3-19200 CL9 4GB (2GB x 2) @ 2000MHz
- Storage: Intel X25-M G2 80GB
- PSU: PC Power & Cooling Silencer 910W 80PLUS Silver


Super PI test

Note: sorry for 'magic graph'. since it takes numerous min. for 32M Pi caculation, results are differred only by a very second. so it's unavoidable to show 'meaningful' graphs.

1. CL Test
http://image.playwares.com/bbs/data/...70657a2837.jpg

2. tRCD Test
http://image.playwares.com/bbs/data/...e9feaad5ad.jpg

3. tRP Test
http://image.playwares.com/bbs/data/...a6009d9069.jpg

4. tRAS Test
http://image.playwares.com/bbs/data/...8c0458e0ff.jpg

5. Command Rate Test
http://image.playwares.com/bbs/data/...4b74967b92.jpg

6. tRC Test
http://image.playwares.com/bbs/data/...6829f36b60.jpg

7. tRFC Test
http://image.playwares.com/bbs/data/...fec4f1e594.jpg

tRFC and tREF are indexed by actual time (second) while others by clock count. It's because they directly indicate characteristics of electrical working mechanism of a memory chip. Both are related to a memory's refresh rate, tREF determines how frequently a chip shall be refreshed and tRFC determines the time interval from the moment of 'refresh' (memory access is unavailable) to when a memory access is back to available. So in case of tREF, the bigger value the better at performance, but the contrary is more likely at stability.

8. tREF Test
http://image.playwares.com/bbs/data/...703fdc0add.jpg

As of now, we covered all 8 elements. Let's see overall analysis.

Overall Analysis: Super PI
http://image.playwares.com/bbs/data/...b7154afb1c.jpg

(...sorry for Korean letter. actually I uploaded this bench to my private blog) The title of the graph is that 'how a ram timing value affects performance when we change just '1-degree' of ram timing option provided by BIOS'. As you see, tREF affects the most but not more than 1%, and followed by tRFC-Command Rate-CL. It's impressive that well-known items such as CL, tRCD affect far less than 'minor' items such as tREF and tRFC.

Then let's go to next test result - Everest Cache & Memory Benchmark


Everest Cache & Memory Benchmark
- Memory Bandwidth test -

Each Everest test consists of two parts: Bandwidth and Latency.
Let's cover the bandwidth results first.

1. CL Test
http://image.playwares.com/bbs/data/...6470662480.jpg

2. tRCD Test
http://image.playwares.com/bbs/data/...42020716dc.jpg

3. tRP Test
http://image.playwares.com/bbs/data/...125a3f6eec.jpg

4. tRAS Test
http://image.playwares.com/bbs/data/...0b6e6e0fd7.jpg

5. Command Rate Test
http://image.playwares.com/bbs/data/...651ceb5d73.jpg

6. tRC Test
http://image.playwares.com/bbs/data/...776f1b0c0c.jpg

7. tRFC Test
http://image.playwares.com/bbs/data/...c5e419fe2e.jpg

8. tREF Test
http://image.playwares.com/bbs/data/...051abf5d47.jpg

Like Super PI test and unlike other ram timings, bigger tREF value is better at performance.

Let's see overall analysis.

Overall Analysis: Read Bandwidth
http://image.playwares.com/bbs/data/...cf373423e5.jpg

Overall Analysis: Write Bandwidth
http://image.playwares.com/bbs/data/...ff4322df91.jpg

Overall Analysis: Copy Bandwidth
http://image.playwares.com/bbs/data/...9e82c1524f.jpg

In read bandwidth test, 4 most-affective ram-timing items are flatted around 2%. Command Rate-tREF-CL-tRFC are said affective items and other 4 ram-timings merely affect performance. But write bandwidth remains nearly unchanged during various ram-timing values has changed. In case of copy bandwidth, only tREF and tRFC exceed the affectiveness of 1% and others remain below 1%. In other words, only tREF and tRFC are actually affect copy bandwidth.

Then let's go to latency test.

- Latency test -

1. CL Test
http://image.playwares.com/bbs/data/...e7c97f84b8.jpg

In latency test, change of CL value derives nearly a linear trendine.

2. tRCD Test
http://image.playwares.com/bbs/data/...c4614aaa69.jpg

Seems merely changed, but it's not 'that small' compared to latter results.
...At least the trendline is undoubtedly linear.

3. tRP Test
http://image.playwares.com/bbs/data/...b529599ed7.jpg

Little fluctuation is appeared, but it seems random empirical error.

4. tRAS Test
http://image.playwares.com/bbs/data/...59fb79d252.jpg

5. Command Rate Test
http://image.playwares.com/bbs/data/...f7ae28eeb6.jpg

6. tRC Test
http://image.playwares.com/bbs/data/...60f13a9757.jpg

Guess there's only superior advantage on the value "tRC 25"? I guess not.

7. tRFC Test
http://image.playwares.com/bbs/data/...4778ca8de8.jpg

8. tREF Test
http://image.playwares.com/bbs/data/...89cb5fdfa5.jpg

Overall Analysis: Latency
http://image.playwares.com/bbs/data/...c9d56a1d9b.jpg

Latency test seems more dependant to ram-timing values than two previous tests (Super PI & Bandwidth). 3 items exceed 1% of affectiveness and among them, 2 exceed 2.5% to performance. 2.5% performance increment upon no CPU / RAM clock change is not just a small variation.

Then let's go to the last test: video transcoding.


Video Transcoding test

This test is to measure the time of converting 1920x1080 TP file to MKV, same resolution. H.264 codec is used. No GPGPU acceleration.

1. CL Test
http://image.playwares.com/bbs/data/...bd448f7a40.jpg

2. tRCD Test
http://image.playwares.com/bbs/data/...372b355a57.jpg

3. tRP Test
http://image.playwares.com/bbs/data/...95c9a9683b.jpg

4. tRAS Test
http://image.playwares.com/bbs/data/...984dbe4f7d.jpg

5. Command Rate Test
http://image.playwares.com/bbs/data/...daa8c41b67.jpg

6. tRC Test
http://image.playwares.com/bbs/data/...59a6b7d6f4.jpg

7. tRFC Test
http://image.playwares.com/bbs/data/...b2fc43c27f.jpg

8. tREF Test
http://image.playwares.com/bbs/data/...1669fa53a9.jpg

Overall Analysis: Video Transcoding time
http://image.playwares.com/bbs/data/...157c41ea24.jpg

Similar to Super PI test, no items reached 1% and tREF marks the top. below CL, affectivenesses cannot even reach one third percent. But still remember: nearly 1% reduction in transcoding time with CPU & RAM clock remain unchanged is not 'that' small.


So, by far, we covered 3 areas: Pi calculation, Synthetic bench, and Video transcoding. Let's see how a ram-timing's affectiveness differ through that benchmark areas.

Affectiveness among test areas
http://image.playwares.com/bbs/data/...34d2395133.jpg

As you see, tRP/tRAS and tRC have nearly "no" impact on performance.
Then let's calculate average affectiveness:
Average = {Super PI + (Read + Write + Copy)/3 + Latency + Encoding}/4
(name of each term indicate a ram-timing's affectiveness on that test area)

Average affectiveness by Ram-timing
http://image.playwares.com/bbs/data/...89929a6027.jpg

For average, affectiveness CL-Command Rate-tRFC are nearly flat around 1% and tREF and tRCD are around three quarter % and a half % respectively.


I think it would be the best for me to help other guys to think like below among various specs of rams:
"6-9-6 vs 7-7-7? since CL's affectiveness on Super PI is bigger than twice of tRCD's affectiveness, former model would be faster than latter on Super PI."

Thanks for reading SCROLLious (:D) post. Have a nice day~

ps. this bench had also been uploaded to my private blog: http://udteam.tistory.com/192

ps2. 24/7 running setting of my system is:

http://image.playwares.com/bbs/data/...2ec116d4b2.jpg

http://image.playwares.com/bbs/data/...83488a7fae.jpg

If other category than here seems more proper, please move it to that category :-)
I'm not yet familiar to this forum system, not familiar to English also :-(

Very nice review. Thanks for all the info.

Wow...

This is awesome! I can't believe you ran 32M so many times! I'd die lol

nice review!

pretty solid review if I ever saw one.

very nice review:up:

so 6-9-6-21 is faster than 7-7-7-21 in pi32M ? think it depends on the ram kit used, speeds etc... I got a 4sec lead with the Hypers at 7-7-7-21 vs 6-8-6-21 (Gskill Flares) at 1800Mhz... with 6-6-6-18 with the GTX 2 I get over 12 secs lead...

Nice extensive testing there... good ideas

Quote:

---

Originally Posted by Leeghoofd

so 6-9-6-21 is faster than 7-7-7-21 in pi32M ? think it depends on the ram kit used, speeds etc... I got a 4sec lead with the Hypers at 7-7-7-21 vs 6-8-6-21 (Gskill Flares) at 1800Mhz... with 6-6-6-18 with the GTX 2 I get over 12 secs lead...

---

maybe hyper's other sub ram timing items are tighter than flare's...?

0xdeadbeef, bassplayer, dopestuff, Arkuatika, bartolomius, Leeghoofd - thanks for comments :)

Quote:

---

Originally Posted by DGLee

maybe hyper's other sub ram timing items are tighter than flare's...

---

My conclusion was : PSC's aren't bad at all, though for benching Hyper is king... Though they are EOL...

For daily users PSC is perfect I tested them in games and hardly noticed any difference between the Hypers !!

For me 1800Mhz C6 is more than sufficient on AMD... At and above 2000 efficiency is less on most of the Asus AMD boards ( except the CH IV Extreme) compared to lower ram clocks...

Quote:

---

Originally Posted by Leeghoofd

so 6-9-6-21 is faster than 7-7-7-21 in pi32M ? think it depends on the ram kit used, speeds etc... I got a 4sec lead with the Hypers at 7-7-7-21 vs 6-8-6-21 (Gskill Flares) at 1800Mhz... with 6-6-6-18 with the GTX 2 I get over 12 secs lead...

---

Oh my, I checked my post again and confirmed that you're right.
For Super PI, tRCD's affectiveness is nearly same as CL's.
(my wrong assumption was that CL's contribution is twice bigger than tRCD's.)

http://image.playwares.com/bbs/data/image/spi_index.jpg

In case of comparing 7-7-7 and 6-9-6, 7-7-7 earns two times of tRCD's affectiveness and loses one CL's and one tRP's: 0.264% x 2 - 0.282% - 0.080% = 0.166% lead.
In case of 7-7-7 and 6-8-6 however, 6-8-6 seems to lead since it reduces another tRCD - if an actual test result says the contrary, the difference may be derived by other sub ram timings.

Anyway... I'd used Elpida Hyper (G.SKILL Perfect Storm) also and it was the best ram I'd ever seen - but it died so suddenly. how sad... (my 24/7 use of that ram is 1600 / 6-6-5-15 or 2000 / 7-7-8-21)

very nice test! learned me some new things. will check your test sometimes in future.

Quote:

---

Originally Posted by Tomasis

very nice test! learned me some new things. will check your test sometimes in future.

---

Thanks :D can't wait your test
question: how can I edit my system information?
-> Oh I just realized how to edit signature. :up:

very nice test...very good for many beginners in benchmarking:)

Some good info here, thanks for posting this up! :up:

Nice testing, but, bare in mind, THIS varies ALOT on platforms, its not that simple, unfortunately.

Quote:

---

Originally Posted by M.Beier

Nice testing, but, bare in mind, THIS varies ALOT on platforms, its not that simple, unfortunately.

---

yup ;) maybe other platforms (ex: INTEL 1156) have different results. I wish I could do & compare results of further tests among several different platforms.
(and... though... still I think this benchmark is valid, at least for AM3 platform!)

Blue Storm, zalbard, M.Beier - thanks a lot for reading my post :)

very good result, good info about tREF a tREC for me!

Quote:

---

Originally Posted by FlanK3r

very good result, good info about tREF a tREC for me!

---

Glad to hear that :D

DGLee -

Welcome to XtremeSystems :welcome: Thank you for the great review. :D

very nice breakdown. well done excel sheet!

Quote:

---

Originally Posted by runmc

DGLee -

Welcome to XtremeSystems :welcome: Thank you for the great review. :D

---

Hello admin. especially glad to receive a reply from you :D

Quote:

---

Originally Posted by ZenEffect

very nice breakdown. well done excel sheet!

---

thanks! another benchmark is reserved for this topic ;)

:ROTF::ROTF:Hey, DgLee;)
Many people refer to memory timing will be good.
Very good.:up:
Welcome to XS.:welcome:

곱등이 대근이 환영!:rofl::rofl:

Quote:

---

Originally Posted by namegt

:ROTF::ROTF:Hey, DgLee;)
Many people refer to memory timing will be good.
Very good.:up:
Welcome to XS.:welcome:

곱등이 대근이 환영!:rofl::rofl:

---

Hi bro lol :D
I'm not a 곱등이 :(

ôw, thanks for this nice job, all done on your own ? you get :up::up::up::up:

Quote:

---

Originally Posted by aslanwas

ôw, thanks for this nice job, all done on your own ? you get :up::up::up::up:

---

Yup :D Thanks for appreciation ;)

Nice, and very thorough! Someday you'll have to go more indepth and see what different combinations provide for performance.

One thing I've noticed was that, at least with my Micron D9 (JNL I think) OCZ Reaper HPC 1600 7-7-7-30 1.9v sticks is the tWRWR (Write to Write Delay) and tRDRD (Read to Read Delay) needs be the same for best memory performance. If they are even one apart it will drop my Copy and Write bandwidth down about 200-300MB/s.

Just did a quick run of a few different sub-timings, from within Windows using CPU-Tweaker 1.5 and MaxxMEM² Preview (Note: I'm running 1333 because my BIOS is weird and some mystery setting/s bump it down from 1600 to 1333, even though 1600 is selected. Also the MaxxMEM² sys info does not update unless you restart the program. Run-to-run variance is about ±20MB/s Copy, ±1MB/s Read, ±8MB/s Write)
Specs:
ASRock 890FX Deluxe3
PhII X2 555BE @ 3200MHz, 2400MHz CPU-NB (CPU-NB brought the score up from 7.7GBs)
OCZ Reaper HPC 2x2GB 1600 7-7-7-30 1T. tRFC@90ns (any others just ask)

Row 1: (assuming 1680x1050 heh)
Image 1 - Base run (Maximum Read Latency@60)
Image 2 - Maximum Read Latency dropped to 55
Image 3 - CPU speed bumped up 100MHz via multiplier (0.5x increase), Max Read Lat@60



Row 2:
Image 1 - CPU +100MHz, Max Read Lat@55
Image 2 - CPU +200MHz, Max Read Lat@60
Image 3 - CPU +200MHz, Max Read Lat@55

Note: No clue why 3.4GHz dropped performance so much compared to 3.3Ghz

Thanks a lot for your info, Formula350 :)
Frankly I'd never thought of them - tWRWR & tRDRD. Really impressive that they affect performance that much.
Maybe I'll retry on them sooner or later ;)

Welcome to XS :)

Very nice thread, good info, effort well done.

If you would be kind enough to post some info on other subtimings, Im having trouble getting these sticks to run at 2000mhz, maybe with some input from your subtimings i can get them to work, Thanks.

E.g here is a cut and paste from aida/everest under motherboard/chipset/Northbridge AMD K10 IMC

Memory Controller
Type Dual Channel (128-bit)
Active Mode Dual Channel (128-bit) - Unganged

Memory Timings
CAS Latency (CL) 8T
RAS To CAS Delay (tRCD) 8T
RAS Precharge (tRP) 8T
RAS Active Time (tRAS) 26T
Row Cycle Time (tRC) 27T
Command Rate (CR) 1T
RAS To RAS Delay (tRRD) 4T
Write Recovery Time (tWR) 8T
Write To Read Delay (tWTR) 4T
Read To Precharge Delay (tRTP) 4T
Four Activate Window Delay (tFAW) 18T
Write CAS Latency (tWCL) 6T
Refresh Period (tREF) 7.8 us
DRAM Drive Strength 1.0x
DRAM Data Drive Strength 1.0x
Clock Drive Strength 1.5x
CKE Drive Strength 1.5x
Idle Cycle Limit 16

Error Correction
ECC Supported, Disabled
ChipKill ECC Supported, Disabled
RAID Not Supported
DRAM Scrub Rate Disabled
L1 Data Cache Scrub Rate Disabled
L2 Cache Scrub Rate Disabled
L3 Cache Scrub Rate Disabled

Quote:

---

Originally Posted by DGLee

Thanks a lot for your info, Formula350 :)
Frankly I'd never thought of them - tWRWR & tRDRD. Really impressive that they affect performance that much.
Maybe I'll retry on them sooner or later ;)

---

I had a friend change them on his DDR2 and didn't show any benefit in Sandra for RDRD and WRWR(For some reason it won't install on my computer so I can't compare that way, but I'm going to reinstall 7 soon anyways), but dropping down just the Max Read Latency by 2ns resulted in a 250-300mb/s improvement!

Quote:

---

Originally Posted by Sanousie

Welcome to XS :)

Very nice thread, good info, effort well done.

If you would be kind enough to post some info on other subtimings, Im having trouble getting these sticks to run at 2000mhz, maybe with some input from your subtimings i can get them to work, Thanks.

E.g here is a cut and paste from aida/everest under motherboard/chipset/Northbridge AMD K10 IMC

Memory Controller
Type Dual Channel (128-bit)
Active Mode Dual Channel (128-bit) - Unganged

Memory Timings
CAS Latency (CL) 8T
RAS To CAS Delay (tRCD) 8T
RAS Precharge (tRP) 8T
RAS Active Time (tRAS) 26T
Row Cycle Time (tRC) 27T
Command Rate (CR) 1T
RAS To RAS Delay (tRRD) 4T
Write Recovery Time (tWR) 8T
Write To Read Delay (tWTR) 4T
Read To Precharge Delay (tRTP) 4T
Four Activate Window Delay (tFAW) 18T
Write CAS Latency (tWCL) 6T
Refresh Period (tREF) 7.8 us
DRAM Drive Strength 1.0x
DRAM Data Drive Strength 1.0x
Clock Drive Strength 1.5x
CKE Drive Strength 1.5x
Idle Cycle Limit 16

---

What brand and model of DDR3 are you running? With out knowing the factory specs it's hard to say what to really adjust.

I won't pretend to know what I'm talking about with the DDR2/3 setups since I upgraded from S939 with DDR1, BUT...

General rule of thumb is: tRAS + tRP = tRC, so for you 26+8=34. You have it set at 27. I don't think that it's preventing you from hitting 2000MHz, but it's a start.

A few other things I'd be try changing to:
CL-tRCD-tRP - 9-9-9
Command Rate - 2T
Write CAS Latency (tWCL) - 8T (If you can change this)

Increase the voltage of the memory a little
Increase CPU-NB voltage (that is the memory controller)
Increase the actual NB voltage (have heard it helps a little)

That's all I can think of w/o knowing you memory model. GL! :up:

Quote:

---

Originally Posted by Formula350

I had a friend change them on his DDR2 and didn't show any benefit in Sandra for RDRD and WRWR(For some reason it won't install on my computer so I can't compare that way, but I'm going to reinstall 7 soon anyways), but dropping down just the Max Read Latency by 2ns resulted in a 250-300mb/s improvement!




What brand and model of DDR3 are you running? With out knowing the factory specs it's hard to say what to really adjust.

I won't pretend to know what I'm talking about with the DDR2/3 setups since I upgraded from S939 with DDR1, BUT...

General rule of thumb is: tRAS + tRP = tRC, so for you 26+8=34. You have it set at 27. I don't think that it's preventing you from hitting 2000MHz, but it's a start.

A few other things I'd be try changing to:
CL-tRCD-tRP - 9-9-9
Command Rate - 2T
Write CAS Latency (tWCL) - 8T (If you can change this)

Increase the voltage of the memory a little
Increase CPU-NB voltage (that is the memory controller)
Increase the actual NB voltage (have heard it helps a little)

That's all I can think of w/o knowing you memory model. GL! :up:

---

Patriot 2000mhz LL kit, rated at 8 8 8 26 , but thats on a p55 platform, with a Trc of 90 in the XMP profile, I found out the ram does not like low TRC, its been a night mare tr ying to get the sy stem stable with this ram, I cant find nothing on google lol.

I can get 1800mhz 7 7 7 20 stable, but anything past that, nothing seems to work, it cant be the cpu-NB its running at 1.475 at 3.4ghz, Im on SS phase.

All i would like is some sub timings info, for some reason i can never find anything on the net about sub timings, So im led to believe that everyone uses the Auto settings?

Nice info.

You could add a memory benchmark specific graphics test like Dragothic Low detail (game 2 test 1) from 3dmark 2001 to see FPS changes. I wouldnt use all the benchmarks of 3d2001, thats a lot of running time lol. Although if you do, be careful with the low detail car test (game 1 test 1) because it has "good runs" and "bad runs" due to the way it was made.

For anyone interested... Using MaxxMEM (I really need to figure out why Sandra errors when loading...) and all the same timings and CPU speed, here is something to consider:
CPU 3.2GHz, CPU-NB 2400MHz, DDR3 @ 1333 6-6-6
Mem Copy: 11405MB/s
Mem Read: 9889MB/s
Mem Write: 7971MB/s
"Reached Memory Score: 8.93GB/s"
Latency: 57.0ns

CPU 3.2GHz, CPU-NB 2000MHz, DDR3 @ 1600MHz 6-6-6
Mem Copy: 12020MB/s
Mem Read: 9548MB/s
Mem Write: 6712MB/s
"Reached Memory Score: 8.13GB/s"
Latency: Forgot to check, but it would've only been around 55ns


So, with thus far being able to run at a lower memory voltage, I've decided to ditch 1600 for now till I work the quirks out with my system :\

Nice review. I'm going to do a little testing soon to see how I compare. I will try and give results if I remember.

Good bump Guru. This is very helpful when trying to figure out how to cut down Pi times!

Nice to see someone has the time to do the benches we all want to see.
This certainly flies in the face of what I previously thought to be true - tRCD being a (relatively) major factor in performance on AM3 platforms. Interesting.

Maybe I won't have to go and drop some money getting a SuperTalent Hyper kit after all.

nice review

on a Sandy bridge system, will the results be the same or nearly the same , or not ?

nice info!