Low latency, HIGH HTT, who wins??

[charlie]

I see all the threads, all the pushin' and shovin'...

Noone can agree.

What's better?

Most will agree that low latency kills high latency (duh!) but where's the line?

It looks like the much heralded OCZ VX4000 will hit 270 for most peeps.... that's nice
Where does that put VX compared to TCCD? We need someone to do a REAL test review, not a phoney baloney Ram#1 with VIA333 chipset and ram#2 with NF4, lol....
I'm confident there's a point where TCCD (2.5-3-3) will pass the performance of VX (2-2-2)...
But where's that point? If that magic number is 335, well then we'd better stick to VX and tight timings. If the number is 310, many of us are hitting that rarified air.
Now there's the issue of high HTT for a 24/7 system... components may not "like" high HTT, although my 24/7 rig runs 9 x 302, 2.5-3-3 with SATA drive....
In fact I think I'll do the little mini review myself... just need to head down to Fry's and grab some OCZ VX sticks (since I don't HAVE any, never had an opportunity to GET any) and I'll use my G.Skill "LA" 4400 sticks (since I HAVE THEM)

I will test A-B:

Sandra b/w buffered/unbuffered

3DMark01/03/05

Real life Doom3 GameDemo Bench

superPi

THEN, we will see where the chips fall... and we can all put this debate to bed.

C

[Rauf]

yes, a test between 10x270 VX/BH-5 VS 9x300 TCCD would be nice.

[Ashe]

Good idea
Wanted to ask that question to bu never had the heart to do so.

Get some vx and show us how the burn down TCCD

[perkam]

making weird sounds while passing your window at night...

Too much information:

Perkam

[bachus_anonym]

@Charlie,

The thing is that with 1GB kits of TCCD, it's very hard to get them rock-stable @ DDR600 7-3-3-2.5-1T. I would be very happy if someone proved me wrong and posted some honest screenshots of Prime95 (LargeFFT and Blend) running for at least few hours.
In some of the everyday tasks, TCCD having 30Mhz advantage over BH-5 is not enough to make up for the loss causes by relaxed timmings and higher latency.
I've got 2x512MB and 2x256MB TCCD kits and I can tell you that none of them hit DDR600 7-3-3-2.5-1T stable enough to pass Prime95. I will continue trying to push them hard...

One of the tests I performed quickly is WinRAR 2.42 Hardware benchmark. 2x256MB TCCD vs 2x2565MB BH-5. I run this benchmark for about 90 seconds and it processed 60MB of data... Compression speed is the result:

335x8 = 2682MHz 7-3-4-2.5-1T ---> 739 KB/s

298x9 = 2683MHz 6-3-3-2.5-1T ---> 727 KB/s

268x10 = 2684MHz 6-2-2-2.0-1T ---> 741 KB/s

Results are clearly in favor of low latency mems (BH-5 in this case...). In that benchmark TCCD would need a lot more speed @ 7-3-3-2.5-1T to beat BH-5.
Even relaxing timmings further, to 7-3-4-2.5-1T and 67MH difference in pure speed doesn't quite cut it... More over, DDR670 7-3-4-2.5-1T is absolutely not doable as far as 100% Prime95 stability goes.

Unfortunately, I sold my OCZ VX PC3200 some time ago so can't compare it to my 2x512MB Gskill... But I might pick some of OCZ's new VX as soon as prices go down...

[charlie]

@Charlie,

The thing is that with 1GB kits of TCCD, it's very hard to get them rock-stable @ DDR600 7-3-3-2.5-1T. I would be very happy if someone proved me wrong and posted some honest screenshots of Prime95 (LargeFFT and Blend) running for at least few hours.
In some of the everyday tasks, TCCD having 30Mhz advantage over BH-5 is not enough to make up for the loss causes by relaxed timmings and higher latency.
I've got 2x512MB and 2x256MB TCCD kits and I can tell you that none of them hit DDR600 7-3-3-2.5-1T stable enough to pass Prime95. I will continue trying to push them hard...

One of the tests I performed quickly is WinRAR 2.42 Hardware benchmark. 2x256MB TCCD vs 2x2565MB BH-5. I run this benchmark for about 90 seconds and it processed 60MB of data... Compression speed is the result:

335x8 = 2682MHz 7-3-4-2.5-1T ---> 739 KB/s

298x9 = 2683MHz 6-3-3-2.5-1T ---> 727 KB/s

268x10 = 2684MHz 6-2-2-2.0-1T ---> 741 KB/s

Results are clearly in favor of low latency mems (BH-5 in this case...). In that benchmark TCCD would need a lot more speed @ 7-3-3-2.5-1T to beat BH-5.
Even relaxing timmings further, to 7-3-4-2.5-1T and 67MH difference in pure speed doesn't quite cut it... More over, DDR670 7-3-4-2.5-1T is absolutely not doable as far as 100% Prime95 stability goes.

Unfortunately, I sold my OCZ VX PC3200 some time ago so can't compare it to my 2x512MB Gskill... But I might pick some of OCZ's new VX as soon as prices go down...

interesting....
but I think P95 is useless. If a rig can do everything that it's required to do, it's STABLE. If you run a network of college 'puters researching the DNA sequence, that's ONE definition of stability, if yuo play doom3 that's another definition of stability.
From dictionary el-Charlie:
Stability= The ability to run reliably during a required task.
In fact, if all you do is stare at the desktop, a suicide screenshot IS 100% stable. If you're benching sPi, a completed run IS 100% stable.

And as far as 512mB vs. 1gB... for benching (racing) you don't need 1gB, except for Bank interleave. So it's a valid comparo to pit 512mB tccd against 1gB VX. And my G.Skill go 310ish at 2.5-3-3

[CrimeDog]

when my dfi nf3 and mobile 3400+ arrive next week i'm going to pit my 1gb mushkin lvl2 vs 512mb 0431 tccd on bp (thanks charlie ). been looking forward to this test for a while.

[nebuchanezzar]

I'm just glad someone is gonna finally do this and test them both in the same system. I do believe in p95 but thats me. I run it with winamp or mp going cuz the 2 nFIIs I still have lose sound usually. I did try some KHX TCCD once and gave up on it cuz my 2*256BH5 beat the 512 stick but I just didn't have the headroom to give the TCCD a fighting chance. Thanks charlie..maybe this will help a lot of people on some buying decisions and kill a few arguements

[bachus_anonym]

Charlie, If Prime95 is useless then I can tell you that TCCD's chances are now higher
I will have some interesting information for you by the end of the day

[jikdoc]

from an everyday usage standpoint, i find 2225 @ 250 faster than 2.5437 @ 310 using utt and tccd. while i don't go overboard with p95, 3 hours is good enough for me, i do think that having some confidence in an overclocked system's stability is important in trouble shooting errors when running programs.

[uwackme]

What matters is MB/s transfered into the CPU chip. Whether that happens with 300Mhz 7,3,3,2.5 or 250Mhz 5,2,2,2.0 is irrelevent. HOW MANY MB per SECOND is delivered is all that matters.

A further issue of random access LATENCY comes into play depending on the nature of the accesses. If all the CPU does is read in cache-blocks 4KB at a time into the internal cache, then random access latency will be a minor factor as 99% of the time the cpu is just bursting in data in a straight line with very little address changes requiring re-issuing RAS.

Further, HTT doesnt have diddly to do with any of this directly. The Hypertransport bus is the connection between the CPU and the IO controller known as NForce4 (for example). How fast DMA access to/from ram from peripherals, and how fast PCI-E access to/from ram are governed by this HTT speed, but in the grand scheme of things, unless you have 10 Raptors in raid0 all being accessed simultaneously, you wont be taxing the 800-1000Mhz HTT bus's bandwidth. CPU to memory access has nothing at all to do with the HTT, HTT can be 800 or 900 or 1000 or 1200 but this doesnt impact the MB/s between the CPU core and the DDR. It can impact NUMBERS you see in bench programs but this will only be due to slight IO throughput and/or responsiveness changes from the higher/lower HTT on a system-wide basis...not because the actual MB/s transfer rate from CPU to memory was changed.

Remember, the HTT is a "DDRlike" bus in that data is clocked on the rising AND falling edges, and it is a 16bit differential bus like PCI-express(meaning two wires PER bit) allowing much higher speeds of operation than the DDR memory bus. So 1000Mhz HTT is like DDR @ 2000Mhz..... a long way from our 500-600Mhz (250-300Mhz FSB) numbers on the DDR memory bus. PCI-express is only a 100Mhz HTT bus.

HTT should be set to the max the CPU<->NForce4 chips are 100% stable at, any higher is not improving usefulness, simply taunting HappyInstabilityBall.

DONT TAUNT HAPPYINSTBILITYBALL!

So if 250Mhz 5,2,2,2.0 gives the same MB/s as 300Mhz 7,3,3,2.5 there should be very little difference in realworld use between the two settings, assuming the same internal NET clock ie: 250Mhz x12 vs 300Mhz x10. The only difference you'd measure from here would be the impact of random-access latency on realworld use.

The real impact for users is what CPU you have, since all but the FX's are LOCKED multiplier's on the upside, to get 300x10 results from 250Mhz ram you need to be able to set a x12 multiplier. If this is not posible, then your decisions is clear

[HARDCORECLOCKER]

I see all the threads, all the pushin' and shovin'...

Noone can agree.

What's better?

Most will agree that low latency kills high latency (duh!) but where's the line?

It looks like the much heralded OCZ VX4000 will hit 270 for most peeps.... that's nice
Where does that put VX compared to TCCD? We need someone to do a REAL test review, not a phoney baloney Ram#1 with VIA333 chipset and ram#2 with NF4, lol....
I'm confident there's a point where TCCD (2.5-3-3) will pass the performance of VX (2-2-2)...
But where's that point? If that magic number is 335, well then we'd better stick to VX and tight timings. If the number is 310, many of us are hitting that rarified air.
Now there's the issue of high HTT for a 24/7 system... components may not "like" high HTT, although my 24/7 rig runs 9 x 302, 2.5-3-3 with SATA drive....
In fact I think I'll do the little mini review myself... just need to head down to Fry's and grab some OCZ VX sticks (since I don't HAVE any, never had an opportunity to GET any) and I'll use my G.Skill "LA" 4400 sticks (since I HAVE THEM)

I will test A-B:

Sandra b/w buffered/unbuffered

3DMark01/03/05

Real life Doom3 GameDemo Bench

superPi

THEN, we will see where the chips fall... and we can all put this debate to bed.

C

It depends on the definition of "better" - got my fastest SuperPi at 12 x 275 with BH-5 at 2-2-2-5, it also was the best RAM delay in EVEREST (35,5 I think) but my best SANDRA MEMBENCH is done with the TCCD at 11 x 305 with 2,5-4-3-7.
So if You want to bench the highest possible MEMBENCH in SANDRA drop the multi, go for TCCD and raise the HTT up to the limit, timings are not so important then.
If You want to get the fastest SuperPI it's another thing, there is to do a lot more of work - if You try with BH-5 at 2-2-2-5 it depends on multi AND HTT, but BH-5 wont do much above 280 with a multi of 12 - if You use TCCD it depends on the board and the BIOS - with a FX-55 on phase You'll get maybe near 10 x 340 and that will be faster even at CL 2,5 I think.

I got almost exactly the same SuperPI with BH-5 at 12 x 275 2-2-2-5 as with the TCCD 60 MHz higher at 12 x 280 2,5-3-3-5.

But 2,5-3-3-5 killed my TCCD that time at 3.360MHz, too much voltage was required. So with 2,5-4-3-7 I think You have to get much more speed and the difference to BH-5 will be more than 100MHz.

Yet I'm still waitin' for my RMA'd DFI NF4 - on NEO2 I can't get 10 x 336 with the TCCD 1:1 (my FX-55 maxes out stable at 3.360MHz) and I also tried 3 kits of BH-5, all maxed out at 12 x 275 with 3,8/3,9 Vdimm, even to push the Vdimm further to the limit of 4,1 didn't help so I think the NEO2 also maxes out at that point - let's see what the DFI NF4 will do.......

[SynGroW]

I did some tests a while back

BH-5 vs TCCD


Pifast with BH-5.
CPU runs 4MHz faster, but I can't help that.


Pifast with TCCD.


Sisoft with BH-5.
Bandwith is lower as with Gskill, but the efficiency is higher because of the MP 11


Sisoft with TCCD.

I didn't use A64 Tweaker in any of the tests, so the results are a bit low for the speeds

Pifast @ higher speeds:


BH-5


TCCD

BH-5 is still the king

[mcnbns]

What kind of voltage are you throwing at that BH-5 compared to the TCCD?

TCCD is the mainstream choice nowadays. If you get some good speed out of it you can nearly match BH-5 performance with less voltage.

I still like the looks of 270MHz X-2-2-X, though...

[Dissolved]

i didnt read every post here..

But IMHO, Low Latency and high HTT can be both good in some ways..

In terms of benchmarking, Low Latency will favor some and higher HTT will favor other benchmarks.

In any case, id take 2225 250+ fsb(100%stable) then 300fsb 2.5-3-3-7 if the total cpu mhz are equal.

[macci]

SynGroW, Bh-5 seems to win there indeed and I have a feeling that the difference would have been slightly bigger if you had used A64Tweaker and ran both at the very optimal settings.

charlie, add memtest to your benchmark list?
also running just GT1 of Mark03 and skipping Mark05 completely might not be a bad idea (will save you some time). Same goes to Sandra Buffered.
Also please remember to bench both useing the optimal A64Tweaker settings (w/ VX you can pretty much max out everything - not sure if that can be done w/ TCCD?).

But at the end of the day 3D2001 ORB #1 pretty much tells you what is the fastest stuff out there..

[Revv23]

subscribed

this makes me feel so good to hanging on to my 2gb of BH-5/6

[skate2snow]

Macci, you cannot put everything to the lowest with TCCD.... Read-Write, Write-Write cannot be put to the lowest.. And there might be more though...

[macci]

The real impact for users is what CPU you have, since all but the FX's are LOCKED multiplier's on the upside, to get 300x10 results from 250Mhz ram you need to be able to set a x12 multiplier. If this is not posible, then your decisions is clear

or then you just use a mem divider of 12 at 300x10 and run the RAM at 250 2-2-2-5..

skate2snow, pretty much what I thought. Another reason why TCCD needs quite a bit more MHz in order to match VX/BH5.

[bachus_anonym]

Here you go folks...

This is what I spend last couple hours on

Couple notes:

1. Memory used: 2x256MB Mushkin LvL II PC3200 BH-5 vs 2x256 Gskill PC4800LA TCCD
2. Mobo: DFI NF4 Ultra-D
3. All DRAM settings are EXACTLY SAME for all three comparisons.

BH-5]


TCCD


4. For 3DMark2001 I used P.O.S. card with no 3D acceleration You can laugh hard at the scores... I excluded Car Low as very unpredictible Test due to physics involved.

Once I get better video card I will show some more extensive testing



EDIT: WinRAR benefits HUUUGELY from tweaked BH-5

[charlie]

SynGroW, Bh-5 seems to win there indeed and I have a feeling that the difference would have been slightly bigger if you had used A64Tweaker and ran both at the very optimal settings.

charlie, add memtest to your benchmark list?
also running just GT1 of Mark03 and skipping Mark05 completely might not be a bad idea (will save you some time). Same goes to Sandra Buffered.
Also please remember to bench both useing the optimal A64Tweaker settings (w/ VX you can pretty much max out everything - not sure if that can be done w/ TCCD?).

But at the end of the day 3D2001 ORB #1 pretty much tells you what is the fastest stuff out there..

macci,
yeah, that's the conventional wisdom... and my leaning also.
HOWEVER!!!
There MUST be a point that TCCD at 2.5-3-3 PASSES BH5/VX at 2-2-2...

Just to make an illustration, would oh let's say 400HTT, 1:1 2.5-3-3 beat the BH5/VX?? I think YES
How about 350? Or maybe 340? Where will high HTT memory beat BH5/VX??
That's the goal here

P.S. Oh and I always look for an excuse to bench '03 and '05.... something about those 2kg solid blocks of unobtainium called ASUS N6800U's staring at me

P.S.S. Oh and I hope you enjoyed your first visit to the US.... I know everyone who saw you was really happy to have met you.
C

[charlie]

Here you go folks...

This is what I spend last couple hours on

Couple notes:

1. Memory used: 2x256MB Mushkin LvL II PC3200 BH-5 vs 2x256 Gskill PC4800LA TCCD
2. Mobo: DFI NF4 Ultra-D
3. All DRAM settings are EXACTLY SAME for all three comparisons.
4. For 3DMark2001 I used P.O.S. card with no 3D acceleration You can laugh hard at the scores... I excluded Car Low as very unpredictible Test due to physics involved.

Once I get better video card I will show some more extensive testing



EDIT: WinRAR benefits HUUUGELY from tweaked BH-5

Bachus,
Oh come on....
293 @ 2.5-3-3??
My 24/7 'puter with full eyecandy/auto-hide toolbar/w'papers/ss's/24/7 realtime HW monitoring and everything will run sPi 8M at 311....2.5-3-3!!
Crank up the TCCD!

[bachus_anonym]

@Charlie...

Who said this is MAX... ??? I can bench way higher but that requires dropping multi as Winnie survives only about 2.72GHz...

The point of this comparison is to show what numbers I get at same CPU speed... Not getting as high clocks as possible...

[charlie]

@Charlie...

Who said this is MAX... ??? I can bench way higher but that requires dropping multi as Winnie survives only about 2.72GHz...

The point of this comparison is to show what numbers I get at same CPU speed... Not getting as high clocks as possible...

whew! good.... what is your max benchable HTT @ 2.5-3-3 sPi1M stable? Because usually sPi 1M OC's are also 3D stable or CLOSE to it (1 or 2mHz)
And trcd seems to give quite a performance hit at 4....
Maybe try 8 x 315 vs. 10 x 252

or 8 x 330 vs 10 x 260??

[bachus_anonym]

or 8 x 330 vs 10 x 260??

No offense but I think you need to stay on the ground little bit harder
You mean, DDR660 7-3-3-2.5-1T ??? SuperPi 8M ??? I'll let you try that first With reasonable timmings this is gonna be VERY HARD...