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

[Assimilator1]

I've just been looking at the prices of used SB mbrds & it seems I can get a mbrd for far less than £100 , looking on ebay they range from £25(!)-£60+, although I'm sure some have unsuitable chipsets, which chipset/mbrd should I be looking for to o/c a SB?

Also a cheaper alternative to the 2500k is the 2400/2380P which go for £80-£100, of course these are o/c limited (max 3.6 GHz) but the SB option is back in contention .

Both Lynnfield & the above SB option would cost about £175 to £210 (CPU, mbrd & RAM, cheapest to average price).
With the i5 2400 having a max o/c of 3.6 GHz & the i5 860 (Lynn) approx. max of 3.8-4 GHz, the performance of those 2 would be close (looking in ATs bench, at the same clock the SB has a small overall advantage). So there I would be better off with the SB (primarily cos I could upgrade latter with a K & maybe an IB CPU) although that hinges on being able to get the right SB mbrd for £50-£60.

The Bloomfield setup would cost £215-£240 so that pretty much rules it out (unless I got lucky on getting a very cheap mbrd ...... that reminds me Mark made me an offer ).

RD
Got ya now

 

[biodoc]

Both machines are dedicated to DC,

2600K@3.6 GHz, 64-bit linux

Asteroid "long WUs", with avx app

HT on: 5037 seconds (average of 36 WUs)
HT off: 2667 seconds (average of 30 WUs)

Q9550@3.4 GHz, 64-bit win7

Asteroid "long WUs", with sse3 app

4078 seconds (average of 9 WUs)

Asteroid "short WUs", with sse3 app

2049 seconds (average of 18 WUs)

 

[GLeeM]

Thanks for the information, biodoc

For comparison, I averaged the last ten WUs sent in from my i7 920 at 3948Mhz.

5720 seconds or about 29,000 ppd

If the Q6600 averages 4660 seconds per WU that would be about 17,800 ppd

 

[Assimilator1]

Thx for the info guys

Ok so to convert to seconds, my Q6600 @3.24 GHz does "long" WUs in about 4666s (averaging by eye only ), short WUs in ~2400s.

Biodoc
So with HT on your 2600k rig's doing 8 WUs in 5037s (that's just cool anyway!).
With HT off it would take it 5334s to do 8 WUs, so HT cuts 8 WU output time by ~6%.

Nice to have but not worth me paying a huge amount for, I beginning to wonder about the i5 750 or maybe 760. 750s go for £40-£50, 760s £50-£60.

Your Q9550 is 14% faster than my Q6600, of which 5% is due to your CPUs higher clock speed. So it seems A@H benefits from the Penryn architecture right at the upper end of its speed up. What FSB is your Q9550 running btw?

Gleem
I assume that's with HT on?
Anyway your rig's ppd is 63% greater than mine! (of which 22% is due to it's higher clock speed).
Comparing your i7 920 @3.95 GHz to Biodocs i7 2600k @ 3.6 GHz, his is 12% faster inspite of a 9% lower clock.
So clock for clock the SB is 21% faster with A@H than Bloomfield, interesting.

And Biodocs 2600k (HT on) is 46% faster than my rig with just a 11% higher clock rate.
Hmm this mixing of ppd & WU times is confusing the picture! lol....... what's Biodocs ppd? (HT on).

 

[Assimilator1]

So anyone know what's a good cheap S1155 mbrd to have?
And S1156 for that matter.

 

[GLeeM]

what's Biodocs ppd? (HT on).

seconds x minutes x hours = number of seconds in a day 60 x 60 x 24 = 86,400 seconds

ppd = 86,400 / 5037 seconds per WU x 240 points per WU x 8 cores = 32,933 ppd

31,100 ppd with HT off

32,933 - 31,100 = 1,833 ppd more

1,833 x 365 = 669,045 more points per year with HT on than with HT off!!!

WOW! Sold me on it

 

[biodoc]

Your Q9550 is 14% faster than my Q6600, of which 5% is due to your CPUs higher clock speed. So it seems A@H benefits from the Penryn architecture right at the upper end of its speed up. What FSB is your Q9550 running btw?

400MHz with 8.5X multiplier so the FSB is running at 1600Hz.

When I first got the chip I was able to clock it to 3.8 GHz stable but for 24/7 DC, I reduced it to 3.4 GHz.

Update on the 2600K stats:

with avx app (from above post):
HT off: 2667 seconds (average of 30 WUs)
with sse3 app:
HT off: 2560 seconds (average of 13 WUs)

So not much difference between the 2 apps. Perhaps the SSE3 app is slightly faster.

I don't have enough data to compare the 2 apps with HT on.

 

[Assimilator1]

Gleem

Yea, but for example would you pay 25% more for a 6% gain?
You've got HT on anyway haven't you? (looking at an earlier post by you it seems you didn't realise you had it? )
Have you ever compared A@H ppd with HT on & off?

I wonder if Lynnfields lower bandwidth compared to Bloomfield affects HT performance?

Just looking on ebay again I see i7 2600k's going for £120-£160 vs £100-£130 i5 2500k
So taking the average of each £140 vs £115, HT (& an extra 100 MHz) costs you 22% more.
(btw I see that prices for both those CPUs has shot up this month! What's that about? Haswell's release?? But I would of thought the opposite would happen as people sell off their old gear )

Ah it's the ppWU that I was stuck on, I just realised I could of looked at tasks for my client, lol.

So Biodocs 2600k ppd is 85% more than my Q6600!

********************************************************************
Let's see if I can summarise the different upgrade options (using average prices, speed boost for A@H).....

Q9550 o/c @3.55 GHz £65 for a 20% speed boost = £3.25/%.

i7 920 o/c @4GHz £45, HSF ~£30, mbrd £90, 12GB DDR3 £105 = £270* for a 63% speed boost = £4.29/%

i7 860 o/c @4GHz £75, HSF ~£30, mbrd £65, 8GB DDR3 £70 = £240 for a 58% speed boost = £4.14/%.
(going by Anands bench, & assuming that all DC apps are bandwidth heavy, which might be wrong! in which case same performance or slightly faster than the Bloomfield's).

i5 750 o/c @4GHz £45, HSF ~£30, mbrd £65, 8GB DDR3 £70 = £210 for a 52% speed boost = £4.04/%.

i5 2400 @3.6 GHz (o/c limited ) £90, HSF ~£30, mbrd £45, 8GB DDR3 £70 = £235 for a 70% speed boost = £3.36/%
(compared to Biodocs 2600k, less 6% for no HT & ~9% for the smaller cache(!) using ATs bench again http://www.anandtech.com/bench/product/287?vs=288 )

i5 2500k o/c 4.5 GHz £115, HSF ~£30, mbrd £45(?), 8GB DDR3 £70 = £260 for a 95% speed boost = £2.74/%
Ok that one's by far the best bang for your buck (or pound ), but it's getting pricey, but it would be worth saving for if I don't get a bargain on 1 of the others.

i5 3350P @3.6 GHz (o/c limited & run hotter than SB!) £130 (new), HSF ~£30, mbrd £45(?), 8 GB DDR3 £70 = £275 for a 80% boost = £3.44/%
(assuming the upper end of its advantage over SB of 10%, see ATs bench).

i5 4440 fixed turboed @3.3 GHz (bar turbo, totally o/c locked! , & they run hotter than IB! What a load of c**p!), £141 (new), mbrd .... do you know what it's not even worth looking up, no way is its better IPC (best case 10% faster) going to make up for 300 MHz, & even if it did the newer platform is just going to be more expensive.

Maybe the more expensive but faster i5 4570S? Fixed turbo @3.6 GHz £150, HSF ~£30, mbrd (Asus Z87) £95, 8GB DDR3 £70 = £345 for a 90% boost = £3.83/%

i5 4670k o/c @4.5 GHz £180, HSF ~£30, mbrd £95, 8GB DDR3 £70 = £375 for a 115% boost = £3.26/%
That's near enough equal 2nd best for £/% gained but that's getting really expensive!

* special offer by Mark,

 

[Assimilator1]

Biodoc
How come you reduced your Q9550 to 3.4 GHz? What vcore does it need for 3.4 & 3.8 GHz btw?

Seems the SSE3 app is 4% faster than the AVX, worth having for free!
Would be interesting to see if HT still improves output by 6% with the SSE3 app

 

[biodoc]

Biodoc
How come you reduced your Q9550 to 3.4 GHz? What vcore does it need for 3.4 & 3.8 GHz btw?

Seems the SSE3 app is 4% faster than the AVX, worth having for free!
Would be interesting to see if HT still improves output by 6% with the SSE3 app

The HSF on the Q9550 is mid-range so it may have been because of heat. In the summer, I usually clock back my computers so my 1090T and 3930K are in the family room and are not overclocked due to heat and noise. It's getting cold so I may overclock the 3930K again. The Q9550 and 2600K are in the basement so they have modest overclocks in the summer.

My wife is always complaining about the electricity bill so clocking back saves some power too.

The Q9550 is at 1.25 Volts @3.4GHz. I don't remember the voltage at 3.8GHz.

The 2600K is a good chip. It'll run stable at 4.1 GHz. I'll ramp it up to 4.1 GHz for the FAH race. The 3930K is a SB-E chip so it consumes a lot of power and runs hot when I overclock it. I can go to 3.75 GHz with the fan on "low" (Corsair H-100 watercooled). Any thing higher and I have to crank up the fan and the noise is prohibitive.

 

[GLeeM]

Yea, but for example would you pay 25% more for a 6% gain? You've got HT on anyway haven't you? (looking at an earlier post by it seems you didn't realise you had it? ) Have you ever compared A@H ppd with HT on & off?

1. I keep forgetting that you don't pay monthly electric bill. For the rest of us, the original hardware cost is rather minor.

2. I only turn off HT for PrimeGrid. I think after the last PrimeGrid race I ran A@H for a short while with HT off ... my results were like biodoc results, a little better with HT on.

 

[Assimilator1]

Err, I don't follow your connection between elec bills & HT, having HT on will increase power consumption, yes you gain ~6% ppd for A@H, but it still draws a little more power http://www.anandtech.com/show/2658/8 (table at the bottom).
True the CPU's efficiency, watts/ppd, (probably) improves, is that what you meant?

And although I don't pay a monthly electricity bill I don't throw caution to the wind , hence my 2nd PC is not on 24/7 & is a dual core Pentium (as well as cost reasons), 99% of the time it's just on on sunday afternoons when my son visits, very occasionally I leave it on crunching in the cooler months. It was on yesterday following some maintenance, then I left it crunching DPAD , 1st time since the beginning of the year.
Also for power & heat reasons I down clocked my quad from ~3.34 GHz to 3.24 GHz a couple of years ago or so, IIRC it dropped power consumption by ~15w.
I'm debating whether I should upgrade the 2nd PC with the Q6600 from this machine when I upgrade it, obviously that would substantially increase it's power consumption!

Anyhow, my point wasn't about power usage but rather would you pay 25% more for a CPU that performed just 6% faster?, assuming HT was the only advantage of course. HT doesn't help your power bills .

HT doesn't work with Primegrid?

Biodoc
Got ya , what temps does the Q9550 run at?

 

[GLeeM]

Err, I don't follow your connection between elec bills & HT

I'll try to explain how I look at it.
Not sure what the 25% comes to but let's say it is $24. If I run that system for only 48 months then the cost is $0.50 per month. If it costs $20.00 per month electricity to run the system, then the increase in price is only 2.5%
Now the question is if I would pay 2.5% increase in cost for 6.5% increase in points.
I'm not sure I am making sense ... can someone else explain?

 

[Assimilator1]

Err you've lost me even more now lol.

The extra cost of owning an HT capable CPU isn't relevant to the electricity bill (bar the extra cost of running HT).

You compare the cost of a non HT CPU, then the HT CPU, work out the extra cost of having HT in terms of % of cost over non HT CPU & then work out performance gained & then decide if your happy with that. It's got nothing to do with the elec bill.

The electrical cost of running HT is an addition to the elec bill, you could also calculate ppd gain vs extra power used (for total ownership cost, bar depreciation costs which is as good as impossible to work out) & see if the extra cost is worth it to you.

The 25% figure I gave was just an example (as mentioned), as in if a HT CPU cost 25% more than a non HT CPU would you buy it, as I said .

As for lumping the cost of the CPU & electrical cost (over 4yrs) together & then working out what % cost HT is of that, I think that's missing the point. Which is HTs affect on CPU cost & performance & whether it's worth it.
For me I wouldn't want to go much beyond 1:1, so if HT made the CPU cost 24% more but the performance gain was only 12% I probably wouldn't bother, 12 & 12% I definitely would, unless I was already slightly over my budget

 

[biodoc]

I thinks GLeeM's point is for those of us participate in DC, the initial cost of the hardware is fairly low compared to the cost running the computer over time. My electricity cost me about $200 per month to run my computers. That's about $12,000 over 5 years. :|

My 2600K@3.6 w/GTX460 (idle) draws 155 watts full load
My Q9559@3.4 w/GTX570 (idle) draws 174 watts full load

Both have the same power supply and both are running Rosetta right now.

For us paying for our power it makes sense to spend extra to get a more powerful chip that's also power efficient. The 2600K gets a lot more work done than the Q9550 for about the same investment in electricity.

IMHO that's why buying a sandy bridge or Ivy bridge is a good investment and that initial higher cost of the chip is relatively insignificant over 5-7 years.

AS1, the Q9550 is running at 51C (rosetta), but the GPU is idle.

 

[GLeeM]

hopefully I'll get my own place next year so the elec bill will be more of an issue then!

For us paying for our power it makes sense to spend extra to get a more powerful chip that's also power efficient.

Everyone is advising you to save for the newest energy efficient technology!

I think we are worried about you quitting DC after you get your own place. We enjoy your posts - this place is slow enough already

 

[GLeeM]

Back on topic.

I was just thinking about underclocking to save cost and remembered something about HT from many years ago.
I had a dual Xeon 1.6 that I OCed to 2.4 that had HT.
At 1.6 the increase from HT was like 30%, but at 2.4 it was only about half that.

So the slower the speed the better HT does. (it has to do with that ~30 instruction long pipeline hits and misses, the virtual core can do work while the pipeline reloads)

Has anyone ever underclocked and how far could you get?

 

[Rattledagger]

Has anyone ever underclocked and how far could you get?

Well, the old 386sx-16 MHz was very easy to underclock to half the speed... just hit the speed-button on the chassis... :whiste:

If my recollection isn't too fuzzy, actually had a game where it was neccessary to downclock the 386, due to the computer being "too fast".


Being more on-topic, maybe my recollection is too Fuzzy here, but isn't power-usage following a square law, meaning double the speed you'll use 4 times as much Power. If this is true, overclocking example an old i7-920 from 2.67 GHz to 3.5 GHz to have the same speed as an i7-2700K would be 31% overclock but 72% higher power-usage. If also includes the i7-920 has a TDP of 130 W and i7-2700K is 95 W, an additional 72% means 223 W power-usage, more than 2x the i7-2700K.

While it won't be so much extra usage when includes the rest of the system and actual power-usage will be less than 223 W for the cpu, overclocking can still give a significant increase in power-usage.

 

[Assimilator1]

Biodoc
Yea understood now , although my point/question is still valid, as in the value of HT, especially as it increases power usage! Personally speaking if a particular CPU's only advantage was HT (giving say a 6% performance boost) I wouldn't pay more than ~10% extra for it, where possible I'd go for a higher clock speed part. But each to their own

My Q6600 draws 206-210w running F@H, so it's rather juicy! (and judging by the ~4C temp increase A@H gave over F@H I'm sure it's drawing more now!).

Gleem
Don't worry I'd never stop DC , worse case scenario if my finances were up against the wall I'd still run DC but not 24/7, maybe only as I had my PC on.
But I've recently just changed job & my wages have gone from ~£19k to ~£24k (& saving ~£800/yr fuel!) , so once I've gotten over a couple of recent bills I should be fine there, & will be better still when my son gets a job!
When I do look at getting my own place I just got to make sure I don't get a place that's cutting it too close for my finances .

Btw I forgot to mention that I do actually contribute towards electricity for running my PC 24/7, something I worked out & offered to my landlady/friend some years ago, so I pay £15/mth on top of my rent for that. Must admit it was a few years ago I worked that out & I haven't recalculated it to take into account price rises :$

Thought I'd already mentioned that I've decided not to buy a Nehalem system (but I'd forgotten to post it! lol), instead what I'll be doing is for a short term mild upgrade I'll be buying a Q9550 (my ebay bidding limit is £65). Then at a latter date, probably late in Q1 next year, I'll buy a Sandybridge system, probably with an i5 2500k.
When I upgrade my main rig I'll switch the mbrd, Q9550 & RAM to my 2nd rig.

I've also just bought a Thermalright U120 extreme cooler, it's a latter revision one that supports S1366, 1156 & 1155 as well as S775, & if I get the cooling improvement I'm expecting then I may be able to o/c my Q6600 to 3.35-3.4 GHz again .

Underclocking? Don't swear!
Only ever done that to find out what the highest FSB is on a mbrd.
Let us know what the outcome is.

RD
I can't remember the laws on heat increase, but what I do remember from overclocking various CPUs is that even a very large o/c gives a moderate increase in CPU temps, but what really rapidly increase temps is voltage increases, could that be what the square increases are from?

 

[biodoc]

My Q6600 draws 206-210w running F@H, so it's rather juicy! (and judging by the ~4C temp increase A@H gave over F@H I'm sure it's drawing more now!).

Rosetta seems to be running cool on all my computers so maybe it's a good summer time project. If I get some time this weekend, I'll switch the Q9550 back to asteroids and check heat and power consumption. Maybe project apps with SSEx/AVX optimizations run relatively hot and consume more power?

 

[Assimilator1]

Yea I'd imagine as it's using more of the cores it's producing more heat, I'm going to do a power test of A@H in a moment.

Well the figures came up much higher than I expected!
Wall power test:-
A@H - 250w!
F@H - 237w
Idle - 163w (speedstep etc. disabled, it never idles anyway )

I was puzzled particularly by the F@H figures, I knew I'd added 128Gb SSD & a BR burner end of last year, but I can't see them adding much (apparently my SSD uses ~4w!) then I remembered I'd upgraded my HD 4830 to a HD 4870 beginning of this year, that'll be why! I bet my rig draws well over 300w loading the graphics card up too! (if you want to know then LMK).

 

[Sunny129]

Maybe project apps with SSEx/AVX optimizations run relatively hot and consume more power?

while i'm no expert, chatter on the A@H forums and elsewhere seem to indicate that some instruction sets task the CPU more than others, resulting in better run times, but also more heat generation. supposedly the AVX2 instruction set on Haswell CPUs is even more brutal than AVX, and makes Haswell chips run very hot when being used.

 

[Assimilator1]

I wonder how hot?

Running A@H my Q6600s 2 hottest cores are at 65C right now, ambient's 23C.
Cooler's a Noctua NH-U12P.

 

[biodoc]

Q9550@3.4 is 174 watts full load with Rosetta (optimizations unknown after search at Rosetta site)
Q9550@3.4 is 176 watts full load with A@H (SSE3 app)

The temps went up 2-3 C with Asteroids.

CPU voltage is 1.225 for both tests.

Now I'm more confused than usual.

 

[Sunny129]

well the difference between 174W and 176W corresponds to a marginal difference in voltage...so my guess is that Rosetta is optimized using an instruction set other than SSE3, hence the difference in temps...but don't quote me on that.