How much does Hyper Threading help DC? (contemplating upgrade)

[Assimilator1]

Although I'm too scint atm, I'm thinking about upgrading in the nearish future.

How much does DC benefit from Hyper threading? Especially 4 vs 8 HTs
Here's the options I'm thinking about so far.

I'll probably be buying 2nd hand gear & probably not a current platform. It's regarding my main rig (the Q6600 in my sig).

At 1st I thought a simple CPU upgrade to my main rig with a Q9650 or Q9550, however the 9650's still go for silly money (£100-150 on eBay) to probably give me ~3.75 GHz (9 x 417), so not worth it for a ~25% boost (16% faster GHz & upto 10% from better architecture).

Whilst the 9550's go for a more reasonable £60-70 the lower multiplier would limit an o/c to ~3.55 GHz (8.5 x 417), just about worth it for a ~20% speed boost (10%+10%) I guess. Although I will sell my Q6600 for ~£25 so I suppose it would be a cheap though mild upgrade then.

************************
My 2nd thought after reading various reviews on AT & THG, recent & otherwise, was the i7 920.
Quad core, 8 threads, dirt cheap on ebay @ £40-60! & from a quick read probably good for 4 GHz (anyone confirm that?). Though as I looked into it their are some major downsides (I don't care it's a dead end platform btw, if the price is right), 1stly it's a triple channel memory platform, so I assume to get the most from the CPU I'd need to populate 3 channels? In which case more expense on memory.

2ndly there are no new mbrds & they are extremely rare on ebay (found 1 atm going for £100, which I would think would match the price of newer platforms new mbrds).
3rdly, am I right in saying that because these are o/ced by the BLCK it's like the bad old days of over clocking where the graphics & HDD buses are overclocked too? (same for S1156?).

***********************
My 3rd thought (& not necessarily my last ) is the S1156 platform with a i7 860 (quad core, 8 threads, 8 MB cache). The CPUs go for £60-90 on ebay & theirs loads of them . ATs review showed they hit 4 GHz with 1 http://www.anandtech.com/show/2832/18 , I've no idea how typical that is, any1 know?
Not sure which mbrd I'd need but Asus ones for example go for £50-£100 on ebay depending on model.
(I've got to re-visit the Nehalem vs westmere performance story, I can't remember offhand what the score is there).

Btw S1155 onwards seems to only allow overclocking for the k series CPUs , which for the 2500K new is £260! And even on ebay their £110-130, so too expensive. The i5 3330 go for £134 new, £90-£110 2nd hand, which is more than the i7 860 & slower factoring in overclocking on the i7.

Haswell S1150 quads start at £142 for the i5 4440 + mbrd, too much.

I can get 8GB (4x2) of Crucial DDR3 1600 for £70 (I know the mbrds are only rated @ 1333 MHz but as long as they run it that's fine).

Oh so many choices!

What do you folk reckon?

 

[petrusbroder]

I have a i7-920 running at 4.0 GHz, 24/7 without any problems at all ... closed-loop water cooling.
I have a i7-860 running at 3.7 GHz with some problems when running some of the more sensitive PrimGrid applications (I usually cut down below 3.4 GHz), but no problems in any other DC. However, I run it 3.2 GHz when folding@home because of temps ..

Here is a good guide to OC LGA1156 Core i7-CPUs.

 

[petrusbroder]

Oh, I forgot: I turn hyperthreading off in PrimeGrid only ... use it in all other projects. It is nice to get 8 tasks done ...
I am not sure if thet is the most efficient way of using the processsors ...

 

[Markfw]

Well, I think it does wonderful. and for F@H, here is my latest suggestion: This dual socket motherboard for $231
http://www.newegg.com/Product/Produc...2E16813182190R
with 2 if these @ 2.93 ghz for $150
http://www.ebay.com/itm/321220167713?ssPageName=STRK:MEWNX:IT&_trksid=p3984.m1439.l2649

So for less than $400 USD you get over 100k ppd (pretty sure) and 16 threads I have the same setup and 2.4 ghz, but they just miss the deadline for the bonus, so I am upgrading to 2.93 ghz

Now the heatsinks are not your standard socket 1366, requires a special HSF. The old style TRUE 120's will work without the backplate.

 

[GLeeM]

3rdly, am I right in saying that because these are o/ced by the BLCK it's like the bad old days of over clocking where the graphics & HDD buses are overclocked too? (same for S1156?).

No. Don't have to worry about much except maybe changing ratio to keep ram at spec.
You need high end CPU cooler to get 4.0. Also, if you go this route, make sure the i7 920 is Revision D0, not C0.

HT can help if your many tasks use different parts of the CPU. This is why I usually try to run at least two different projects at the same time.

Are you sure you can't afford newer technology? The savings in electricity with newer tech would probably pay for itself in short time

 

[Assimilator1]

Thanks for the input guys

Gleem
Re BCLK, so what's Anand on about here?

Unfortunately the PCIe controller on Lynnfield is tied to the BCLK. Increase the BCLK to overclock your CPU and you're also increasing the PCIe controller frequency. This doesn't play well with most PCIe cards, so the first rule of thumb is to try and stay at 133MHz multiples when increasing your BCLK.
From here http://www.anandtech.com/show/2832/19

Thx for the cpu rev tip , although I probably won't get a i7 920 due to difficulty of getting a mbrd + the extra RAM expense.
Btw I take it running a i7 920 on just 2 RAM sticks would cripple it's performance?

Atm I don't pay electricity directly (via rent), though I don't want to take liberties so probably another good reason to stay away from the 920 . Oh and hopefully I'll get my own place next year so the elec bill will be more of an issue then!

Petrus
Do you know what sort of performance boost you get with HT? I assume that each WU time takes longer with HT but that more WUs get done, curious to know what sort of boost it gives.

Oh & thanks for the guide ....... which you forgot to link

Btw I don't plan to water cool, is your 860 on air or water?

Mark
Hmm dual socket would be nice! however even at $400 (£248) that's getting steep, and anyway I can't get those bits at those prices, if at all.

I can't get those CPUs at $75ea, none are available in the UK & to ship them from the states (even if I was happy about that) would cost me $243 for the pair (from another guy).
Btw the link to the mbrd doesn't work, but I copied & pated the item no. Is this the right 1? no-redirect!-www.newegg.com/Product/Product.aspx?Item=13-182-190R&Tpk=N82E16813182190R (AT please don't redirect! - grr they have! - that fixed it!).
Anyway I can't get that mbrd here or any ATX dual mbrd it seems (at least under £250, didn't bother looking over that).

***********************************************************

So atm it looks like I could either go for a mild upgrade with a Q9550 @ £60-70 (-£25) giving me a ~20% boost.

Or an i7 860 £60-£90, 8GB RAM £70, an unknown mbrd £50-£100? oh & of course a new cooler ~£30 giving a 25-35% boost (comparing the Q9650 to the i7 870 in the AT review I linked, the Q9650 will be very close to what my o/ced Q6600 does & I knocked off 5% of the i7 870 CPU intensive scores to estimate i7 860 scores, (I'll look at the bench latter).
Plus of course even if I can only overclock to 3.5 GHz that'll be an additional 25% speed boost (err assuming it scaled linearly of course!).

So using averaged prices, ~£220 (less money I get for old bits) will give me a 50-60% speed boost.

I'll look at newer platforms again latter.....

 

[GLeeM]

Unfortunately the PCIe controller on Lynnfield is tied to the BCLK.

920 is Bloomfield. I can change BCLK by one to tweak the OC

 

[Assimilator1]

Tweak the OC?

Anyhow, like I said in my previous post I'm probably not getting a Bloomfield 920 (partially due to problems getting a mbrd but also the RAM issue , unless I can get a cheapish mbrd) [edit] on 2nd thoughts the mbrds aren't as rare or as expensive as I 1st thought, anyone know what the Gigabyte or MSI 1366 mbrds are like?.

Talking of RAM, that reminds me, you didn't answer that question , I assume Bloomfield will be quite badly crippled if it only had 2 RAM channels used? (especially seeing as it's only supports 1066 DDR3, how well can it o/c RAM though?)

So, until I evaluate new options (which offhand I think will be just too expensive to gain a major boost in speed, but I'll look into it) atm I'm looking at a i7 860, 8GB DDR3 1600 (for future proofing & o/cing purposes) & whatever 1156 mbrd is suitable.
So seeing as I may well go the Lynfield route, would I then have problems with the PCIe & SATA bus being o/ced??
Hmm, wondering if I should be looking at the i7 870 for it's higher multiplier........ they mostly go for £100-£120, maybe not then .

Btw any1 any suggestions for a good overclocking S1156 mbrd? (I don't care about SLI or XF)

 

[Sheep221]

for dirt price you can get regular non k i7 haswell and cheap mobo or used SB/IB unlocked and z77, p67 or z68 combo, both options will be nearly same price but much more faster

it is not reasonable to invest in old equipment when you want to increase your ppd significantly

 

[GLeeM]

Talking of RAM, that reminds me, you didn't answer that question

Yeah, I don't know for sure.
I think I remember someone saying if you had four sticks to use them all because having more memory was better than triple channel.

Tweak the OC?

I can tweak the OC by changing BCLK from 190 to 189.
I'm not sure how you could "stay at 133MHz multiples when increasing your BCLK"?

Make sure you carefully check out newer technology - even i3 or low end i5.

 

[Assimilator1]

I seem to recall looking at i3 SBs & they were slower or no faster than my Q6600 @3.24 GHz
I'll look into i5 though.

for dirt price you can get regular non k i7 haswell and cheap mobo or used SB/IB unlocked and z77, p67 or z68 combo, both options will be nearly same price but much more faster

it is not reasonable to invest in old equipment when you want to increase your ppd significantly

I guess you haven't read my earlier posts, but used unlocked SBs are anything but cheap , I've found the 2500k going for £110-£130, I can't imagine the latter IBs ks would be cheaper?
True it's going to o/c more, & per MHz it's going to be a little faster than i7 Lynnfield but it'll cost a lot more too, at least the CPU anyway, I'd imagine the mbrds too?

Re non k i7s, I don't know where you got the idea that non k i7s are dirt cheap , they start at ~£227! (i7 4770S or i7 3770), that's new.

I have wondered about the i5 3350P (3.1 GHz, £133), but with no HT it's going to be left trailing an o/ced i7 860 @3.5+ GHz surely?
Anyhow, even if not it'll be quite a bit more expensive.

Btw are non K IBs overclocked limited to +4 bins like the non K SBs? (seems they are http://www.anandtech.com/show/7003/the-haswell-review-intel-core-i74770k-i54560k-tested/3 )
(yet to look at non K SB/IB performance vs o/ced Lynnfield).

What a minefield!

 

[Bubbleawsome]

I fold with an i7 870 and while it is fast, it's not that fast. If that is the money you have, it's ok. But if you will be able to save up over time I would wait until you can afford haswell. I only manage 22k ppd with the 870 and a hd 7770 using beta units. Also, they are hot little chips, running 24/7 with no overclock (dell stock board, 1.2-1.3 vcore, 3.6 turbo.) I hover around 80-83c. Max fan. Loud.
4500rpm stock intel 80mm hsf (loud) and 2000 rpm 80cfm 92mm fan I found in an old Alienware. (Silent) stock dell case though, so take what you will.
The system was a studio xps 8100.

 

[Sheep221]

I seem to recall looking at i3 SBs & they were slower or no faster than my Q6600 @3.24 GHz
I'll look into i5 though.


I guess you haven't read my earlier posts but used unlocked SBs are anything but cheap , I've found the 2500k going for £110-£130, I can't imagine the latter IBs ks would be cheaper?
True it's going to o/c more, & per MHz it's going to be a little faster than i7 Lynnfield but it'll cost a lot more too, at least the CPU anyway, I'd imagine the mbrds too?

Re non k i7s, I don't know where you got the idea that non k i7s are dirt cheap , they start at ~£227! (i7 4770S or i7 3770), that's new.

I have wondered about the i5 3350P (3.1 GHz, £133), but with no HT it's going to be left trailing an o/ced i7 860 @3.5+ GHz surely?
Anyhow, even if not it'll be quite a bit more expensive.

Btw are non K IBs overclocked limited to +4 bins like the non K SBs? (seems they are http://www.anandtech.com/show/7003/the-haswell-review-intel-core-i74770k-i54560k-tested/3 )
(yet to look at non K SB/IB performance vs o/ced Lynnfield).

What a minefield!

you can get used 2600k+p67 combo for £250 or less, it would be huge mistake to buy nehalem for this price

 

[GLeeM]

but with no HT it's going to be left trailing an o/ced i7 860 @3.5+ GHz surely?

It would probably be close or better.

You have to remember - HT makes two tasks take almost twice as long. The improvement is in the single digit percent. (if the two tasks are using different parts of the CPU)
Years ago, on P4 Northfield, I could get pretty good improvement by running one integer (or single precision) task and one double precision floating point task from F @H. I would guess that today you could get good improvement running half AVX and half any other task. If you ran all AVX tasks you would probably see no improvement.

 

[Assimilator1]

you can get used 2600k+p67 combo for £250 or less, it would be huge mistake to buy nehalem for this price

Does that include RAM? Because the £220 I quoted for the i7 860 system includes 8GB of DDR3 1600.

Gleem
Re HT, ah thanks, that's the kind of info I was after .
Atm I'm running Asteroids@h & not F@H, I will probably be running A@H for sometime & AFAIK their is currently no choice of WU type. So in which case am I right in saying that HT would be worthless? That would change my choice of CPU a lot!
And even when I return to F@H & assuming I could chose to run 2 different types of WU then the gain is less than 10%?

Bubbleawesome
Thx for the info, worth knowing for sure , just FYI I won't be using a stock HSF regardless of which CPU I get.
What ppd do you get from the CPU alone?

 

[petrusbroder]

Petrus
Do you know what sort of performance boost you get with HT? I assume that each WU time takes longer with HT but that more WUs get done, curious to know what sort of boost it gives.
Oh & thanks for the guide ....... which you forgot to link

Well, the performance is slightly better with HT than without. The WUs take longer, but I mix projects and thus gain a little. Does not make much in the short run, but 3-5% shorter crunching for 2 WUs makes a lot of WUs looking over serveral months of 24/7 crunching.

The link? Let me check ... here it is!

 

[GLeeM]

So in which case am I right in saying that HT would be worthless? That would change my choice of CPU a lot!

No it would not be worthless - there would be some improvement.
Now you went from one extreme to the other! You were expecting more then you would get so I try to bring you to the middle and now you are at the other end

There is also most GPU WUs use some CPU. HT helps alot with this and other automatic tasks that a computer runs. Also sometimes there is a user browsing, emailing, typing, etc. that will use some CPU time. HT helps with all these.

When I update I will get HT.

 

[Assimilator1]

I think you mis-understood me some what , I was asking if it was worthless where running only 1 type of project with no choice in WU types.
And at my 1st post I had no idea how useful HT was for DC hence me asking .

Petrus
Thx

 

[biodoc]

I don't think it's as simple as HT vs non-HT. My 2600K (3.6 GHz) crunch times per thread on Asteroids with HT on were a bit better (less time) than my Q9550 @ 3.4 GHz. You can compare your Q6600 crunch times with other processors by going through the "top hosts' list at Asteroids. http://asteroidsathome.net/boinc/top_hosts.php

The problem with top hosts browsing is you have no idea what the clock speeds are but I find it's useful.

Roughly, my 2600K PPD on asteroids was about 2-fold better than the Q9550 and my 3930K was 3x better with HT on (12 threads).

I have no idea why but at least on Asteroids, CPU architecture must have a significant impact.

If you want, I can do some tests on the current WUs at asteroids (Q9550 vs 2600K (with and without HT on) at the same clock speed.

I think you'd find that your Q6600 vs a Q9550 would be nearly identical at the same clock speed. I know it's a lot of money, but it might be worth waiting and saving up for a sandy bridge or ivy bridge processor.

 

[Sheep221]

Does that include RAM? Because the £220 I quoted for the i7 860 system includes 8GB of DDR3 1600.

Gleem
Re HT, ah thanks, that's the kind of info I was after .
Atm I'm running Asteroids@h & not F@H, I will probably be running A@H for sometime & AFAIK their is currently no choice of WU type. So in which case am I right in saying that HT would be worthless? That would change my choice of CPU a lot!
And even when I return to F@H & assuming I could chose to run 2 different types of WU then the gain is less than 10%?

Bubbleawesome
Thx for the info, worth knowing for sure , just FYI I won't be using a stock HSF regardless of which CPU I get.
What ppd do you get from the CPU alone?

As I said, you can do that, but I highly recommend to save abit more and get IB/SB rig. And no I didn't count the RAM.

The HT is advantage only in some projects, for example Boinc DC project allows use of HT, so you can run up to 8 or 12 instances respectively, depending how many threads you operate.
So if your project doesn't benefit from HT much, I would just save the cash and get an i5 instead.

 

[Rattledagger]

No idea about overclocking, the i7-920 ran very hot with the Stock cooler (uhm, how was it to turn-off the auto-erroring what screws-up every second Word in IE, it seems upgrading to Windows 8.1 has turned it back on ) but as far as memory goes you'll definitely want to use either 3 or 6 sticks (no idea about how the 4-stick-mainboards really are for the old i7-920.)

As for HT, my only tests for this was with Climateprediction, where if ran 4 Hadam3p-models at once the i7-920 performs as a 3-core cpu while running 8 it performs like a 3.5-core cpu, meaning HT gives an additional 20%.

Still with CPDN, a quick look reveals the old i7-920 uses 2,02 s/TS running Hadcm3n, while the i7-2700K uses 1,45 s/TS. Appart for faster clock-speed the i7-2700K has roughly 5% extra advantage over the older model.

If also includes the 920 is a 130W cpu while the 2700K is a 95W, I'm not so sure buying the old 920 is a good choise even if you can get it somewhat cheaper. Even the 2700K is by now 2 generations old, while the i7-920 is it ... 4 generations old?

 

[Assimilator1]

Biodoc
If you don't mind doing those test on your Q9550 & your 2600k then yes thanks that would be very useful, especially HT vs non HT .
My Q6600 is doing the (now rare) short A@H WUs in ~39 mins, the long 1s in ~1hr 17mins.
Oh & as for big speed ups in A@H vs different architecture it's probably because the latter CPU supports a higher level of SSE which is being used by A@H.

Btw clock for clock the Q9x50s are typically 5-10% faster than the Q6x00 CPUs, not much but it's there .

Sheep221
I hear what your saying about getting a better bang for your buck on the SBs but at the end of the day it is another £100 on top of the Lynnfield system (going by your price for CPU & mbrd).
Though the SB k system would be quite a bit more powerful especially seeing as the k o/cs easily to ~4.5 GHz as I understand it?
The Lynnfield i7 860 to 4 GHz (& without further digging, that maybe the upper end?), & what about Bloomfield?

It seems an i7 Bloomfield at the same clock as a i5 SB is very similar performance (bar single thread apps which I'm not bothered about) the SB in the biggest gaps is 10-15% faster http://www.anandtech.com/bench/product/288?vs=99 , I had thought the 3 RAM sticks needed for the 920 would make it too expensive but the cheaper CPU counters that (relative to an i5 Lynnfield that is). Of course like you say I could wait, save more money & with SBs better o/cing & better core get a faster & more economical rig........
Assuming an SB i5 k @4.5 GHz & a i7 920 @4GHz, is the SB worth the extra 45% cost to gain 22-27% performance? Hmm.....

Theirs so many variables to this my head's hurting!

I'm running Asteroids@home atm which is a BOINC project, so HT would be useful, hopefully Biodoc can give us some numbers on that .

Hi RD
My Q6600 @ 3.24 GHz runs very hot, I'm not going to be using the stock cooler.

Re climatepred on the 920, I don't get ya, why are you saying it performs like a 3 core cpu when it's a quad core? Are you comparing it to another architecture?
Useful to know about the HT boost anyway :thumbsup:

What's s/TS?? (you've got a 2700k? nice rig, even if it is 2 gen's old ).

 

[biodoc]

Biodoc
If you don't mind doing those test on your Q9550 & your 2600k then yes thanks that would be very useful, especially HT vs non HT .
My Q6600 is doing the (now rare) short A@H WUs in ~39 mins, the long 1s in ~1hr 17mins.
Oh & as for big speed ups in A@H vs different architecture it's probably because the latter CPU supports a higher level of SSE which is being used by A@H.

Ok, I started collecting data with my 2600K@3.6 GHz with HT on. Then I'll turn HT off and collect more data. I think 30-40 WUs of each should be enough to see if there's a difference in overall ppd. We'll see.

 

[Sheep221]

if astroids are on boinc than you can get nehalem and run it 8x in a row then, where virtual threads will be slower due to nature of HT but you will end up fine in this case so you just can get it then

 

[Rattledagger]

Hi RD
My Q6600 @ 3.24 GHz runs very hot , I'm not going to be using the stock cooler.

Re climatepred on the 920, I don't get ya, why are you saying it performs like a 3 core cpu when it's a quad core? Are you comparing it to another architecture?
Useful to know about the HT boost anyway :thumbsup:

What's s/TS?? (you've got a 2700k? nice rig, even if it is 2 gen's old ).

My tests was all done on the i7-920, using the exact same wu as benchmark-wu. First ran only the benchmark-wu and nothing else, this gave a total production of N wu/day. Afterwards ran benchmark-wu + 3 other wu's alongside. If the Production was fully linear, this means total Production should now have increased to 4N wu/day, but in reality the total Production was only 3N wu/day. Lastly ran benchmark-wu + 7 other wu's alongside, this gave 3.5N wu/day as total production.

As for s/TS, this means Seconds per timestep and should be lowest possible. All CPDN-wu's of a given type has the same number of timesteps. For this particular wu-type where's roughly 1 million timesteps, meaning the 2700K uses 1.45 Seconds to finish 0.0001% of the wu, meaning roughly 17 days to finish a wu...