Haswell Preview Benchmark

[chalsall]

Originally Posted by chappy

plus, it's diamonds! which are a thermal's best friend.

Not always...

Simple thermal dynamics would suggest that many "bucky-tubes" (AKA: nano-tubes) might present more surface area for electron transfer.

There's a reason soap bubbles almost immediately become spheres.

[chappy]

Originally Posted by chalsall

Not always...

Simple thermal dynamics would suggest that many "bucky-tubes" (AKA: nano-tubes) might present more surface area for electron transfer.

This kind of literal thinking is exactly why this forum is a sausage fest. :)

[only_human]

Originally Posted by kladner

IC Diamond 7 Carat claims to be mostly powdered industrial diamond. I have it on my CPU heatsink right now. I can't say it's better than Arctic Silver 5, but the break in period is stated as 2 hours, instead of 200 with a number of heat-cool cycles for AS5.

Originally Posted by chappy

seems to imply that Batalov is correct Toothpaste is better than Chocolate though.

According to the attachment in chappy's post (attachment), mayonnaise is as good as IC Diamond Carat 7. I'm just sayin'.

[ewmayer]

Originally Posted by chalsall

Not always...

Simple thermal dynamics would suggest that many "bucky-tubes" (AKA: nano-tubes) might present more surface area for electron transfer.

There's a reason soap bubbles almost immediately become spheres.

I don't think surface area of the molecules is what matters; rather IIRC heat conduction and electron mobility are almost-invariably closely correlated (think copper and diamond, which are structurally very different); in other words thermal/electrical conduction/insulation go together. But yes, some types the nanotubes have exceedingly high thermoelectrical conductivity, albeit only in the direction of the tube axis (and perhaps in the azimuthal 'spiraling round the tube' direction, which is useless for bulk heat transfer):

Because of the symmetry and unique electronic structure of graphene, the structure of a nanotube strongly affects its electrical properties. For a given (n,m) nanotube, if n = m, the nanotube is metallic; if n − m is a multiple of 3, then the nanotube is semiconducting with a very small band gap, otherwise the nanotube is a moderate semiconductor. Thus all armchair (n = m) nanotubes are metallic, and nanotubes (6,4), (9,1), etc. are semiconducting.[56]

However, this rule has exceptions, because curvature effects in small diameter carbon nanotubes can strongly influence electrical properties. Thus, a (5,0) SWCNT that should be semiconducting in fact is metallic according to the calculations. Likewise, vice versa—zigzag and chiral SWCNTs with small diameters that should be metallic have finite gap (armchair nanotubes remain metallic).[56] In theory, metallic nanotubes can carry an electric current density of 4 × 109 A/cm², which is more than 1,000 times greater than those of metals such as copper,[57] where for copper interconnects current densities are limited by electromigration.

Because of their nanoscale cross-section, electrons propagate only along the tube's axis and electron transport involves quantum effects. As a result, carbon nanotubes are frequently referred to as one-dimensional conductors. The maximum electrical conductance of a single-walled carbon nanotube is 2G₀, where G₀ = 2e²/h is the conductance of a single ballistic quantum channel.[58]

LOL: And from the "you couldn't make stuff better than this up if you tried" department, we have this inadvertent gem of ribaldry in the Wikipage on Fullerenes:

According to astronomer Letizia Stanghellini, "It’s possible that buckyballs from outer space provided seeds for life on Earth."[5]

Ooh, Letizia, you naughty double-entendring little minx, you. :)

[TheMawn]

Lol you guys and your heatsink issues. One small deviation from the book and things go wrong. I, on the other hand, once sawed off one of the four arms where the screws connect to the backplate because of a very inconveniently placed capacitor (just another one of my less than agreeable incidents dealing with Dell-tard tech) on the motherboard. I instead used a twist tie to tighten down that part of the heatsink. No issues that I know of.

I'm amazed any of you are talking about stock coolers with Ivy Bridge (or even Sandy Bridge and Haswell) in the first place as even a Hyper 212 which is one of the best and cheapest heatsinks out there makes such a huge difference.


Regarding thermal paste: I will say that stock is a noticeable step below aftermarket paste, usually. Particularly the crap the actual processor manufacturers send out with their equally crap heat sinks.

I replaced the roughly 6-month old paste on my old dell box's CPU heat sink and took off about 5C. Changing the heatsink took off another 10C.

I replaced the 18-month old paste on my GPU's heatsink and took probably 5C off the temperatures. It's hard to say because the actual change was 40C, but that's because the combination of cat hairs forming a mesh over the heatsink fins and dust filling in the holes in the mesh allowed for virtually 0 air flow until I (literally) peeled the dust off.

I replaced the 2 year old paste on my laptop's heatsink and took off 10C. The old stuff looked a bit dry.

Take this all with a grain of salt. My feeling is that a decent paste like AS5 makes enough of a difference, but it is just as likely that the stuff I replaced was just old and needed to be replaced anyway. $10 for a tube is probably all you want to spend, and only if you plan on applying paste enough times. I heard this new GELID stuff was a godsend so I picked up a tube of that just to see. I now have a 3/4 tube of AS5 and a 3/4 tube of the GELID stuff which I will probably never be able to finish in its lifetime.


P.S. Regarding RTFM: http://xkcd.com/293/

[kladner]

Originally Posted by only_human

According to the attachment in chappy's post (attachment), mayonnaise is as good as IC Diamond Carat 7. I'm just sayin'.

Yeah. I noticed that. It was a cheap experiment in any case. I'm going to have to have the board out in the fairly near term. I'll go back to AS5 then. I have both lying around. I would bet more on the durability of the diamond goop over mayo, though. Actually, mayo is as good as a whole swath of purpose-brewed products.

[kladner]

http://archive.benchmarkreviews.com/...&limitstart=12

[only_human]

Originally Posted by kladner

I would bet more on the durability of the diamond goop over mayo, though. Actually, mayo is as good as a whole swath of purpose-brewed products.

220% agree. Also mayo is putting acid and water right where you don't want corrosion. To consider the heat stability of the mayo would already be taking it too seriously. "My thermal paste spoiled." "Did you leave it out in the sun, hon?"

[ewmayer]

Got the Haswell fired up last night - last blocking piece of buffoonery was a not-completely-plugged-in SATA cable to the SDD. Been running my ongoing test of F28 @FFT length 15360K in 4-core mode for over 12 hours now, no signs of throttling. On the Sandy Bridge this was running at 115 ms/iter, now down to 84 ms. (This is using the same DDR3 memory sticks that were on the SB mobo - faster memory gets installed later).

More complete timing table in a few hours.

Thing is amazingly quiet with the stock cooler - running under full load, both side panels off the case and 4 case fans running at low speed it's no louder than my macbook, which is running similarly all-cores-loaded and sitting on the same desk.

I'll remove the heatsink (and fan from that) from the SB later and give the heatsink a good long soaking in soap and hot water - that'll also give me a chance to inspect the 18-month-old OEM thermal paste before I box it back up for storage-until-new-home-found.

[ewmayer]

Timings for Mlucas on Haswell.

The percentage-change datum to the right of each tabulated timing datum reflects per-cycle (i.e. scaled for the differing frequencies of the 2 chips I am using) throughput change for Haswell versus Sandy Bridge. The -3% data here in the 4-thread column are artifacts of the too-granular timing reporting for the self-tests, which causes "no apparent per-iteration timing change when rounded to nearest millisecond" to get processed as 100% * (3.3 GHz)/(3.4 GHz) * (T_sb/T_has - 1), which gives -3% when the reported tiings are the same-to-the-millisecond, i.e. when T_sb = T_has.

Test #3: SSE2 mode on 3.4 GHz Haswell quad, DDR3 SDRAM 1333 (PC3 10600):

                  Mersenne-mod:                             Fermat-mod:
FFT len    sec/iter [% throughput change]            sec/iter [% throughput change]
(Kdbl)    1-thread     2-thread     4-thread      1-thread      2-thread      4-thread
1024	.017 [+ 8%]  .008 [+21%]  .005 [- 3%]   .0158 [+10%]  .0082 [+10%]  .0044 [+10%]
1152	.020 [+26%]  .011 [+15%]  .006 [+13%]
1280	.023 [+ 5%]  .012 [+ 5%]  .007 [+11%]
1408	.027 [+37%]  .014 [+32%]  .008 [+21%]
1536	.027 [+ 8%]  .014 [+ 4%]  .008 [- 3%]
1664	.031 [+41%]  .015 [+49%]  .009 [+40%]
1792	.033 [+ 9%]  .017 [+ 8%]  .010 [- 3%]   .0313 [+ 6%]  .0161 [+ 5%]  .0088 [+ 6%]
1920	.036 [+40%]  .018 [+46%]  .010 [+36%]   .0338 [+43%]  .0177 [+37%]  .0095 [+38%]
2048	.037 [+ 8%]  .019 [+ 7%]  .010 [+16%]   .0353 [+ 6%]  .0182 [+ 5%]  .0098 [+ 6%]
2304	.042 [+20%]  .022 [+19%]  .013 [+12%]
2560	.048 [+ 9%]  .025 [+ 9%]  .014 [+25%]
2816	.058 [+31%]  .030 [+29%]  .017 [+26%]
3072	.056 [+11%]  .030 [+10%]  .018 [+ 2%]
3328	.066 [+38%]  .034 [+37%]  .019 [+38%]
3584	.069 [+10%]  .035 [+11%]  .022 [+ 6%]   .0656 [+ 8%]  .0342 [+ 5%]  .0191 [+ 8%]
3840	.077 [+37%]  .039 [+36%]  .022 [+32%]   .0741 [+35%]  .0377 [+34%]  .0201 [+36%]
4096	.074 [+ 9%]  .040 [+ 7%]  .024 [+ 9%]   .0733 [+ 8%]  .0382 [+ 6%]  .0216 [+ 6%]
4608	.088 [+22%]  .046 [+20%]  .028 [+14%]
5120	.096 [+ 9%]  .051 [+10%]  .032 [+15%]
5632	.121 [+32%]  .062 [+31%]  .038 [+20%]
6144	.114 [+ 9%]  .061 [+10%]  .043 [+13%]
6656	.138 [+34%]  .072 [+33%]  .043 [+29%]
7168	.139 [+10%]  .075 [+ 7%]  .052 [+10%]   .1349 [+ 6%]  .0710 [+ 5%]  .0497 [+12%]
7680	.164 [+35%]  .084 [+34%]  .049 [+29%]   .1580 [+32%]  .0822 [+29%]  .0445 [+32%]
8192	.157 [+ 9%]  .085 [+ 7%]  .059 [+12%]   .1503 [+ 9%]  .0790 [+ 7%]  .0540 [+15%]
----------------------------------------------------------------------------------------
Avg || Scaling:          1.929x       3.242x        ----           1.925x        3.387x

Avg per-cycle throughput gain, Haswell vs Sandy Bridge:
            1.203x       1.200x       1.169x         1.163x        1.144x        1.169x

Notice how the per-cycle throughput improvements occur disproportionately for the "awkward" FFT lengths involving odd components 9,11,13,15. The result is a much more monotonic (with increasing FFT length) timing profile on Haswell than on Sandy Bridge.

The reason for this rather striking FFT-length-smoothness-based timing disparity is unclear to me.

---------------------------

Next are timings on Haswell for the AVX-based Mlucas code. Here we see some really significant per-cycle throughput gains, and not just for large-odd-component FFT lengths, though those still exhibit larger relative speedups than their smoother brethren.

On Haswell, the throughput boost for AVX-vs-SSE2 is appreciably better than on SB for both Mersenne-mod (1.2-1.3x) and Fermat-mod (1.4-1.6x) convolution.

Test #4 AVX mode on 3.4 GHz Haswell quad, DDR3 SDRAM 1333 (PC3 10600):

                  Mersenne-mod:                             Fermat-mod:
FFT len    sec/iter [% throughput change]            sec/iter [% throughput change]
(Kdbl)    1-thread     2-thread     4-thread      1-thread      2-thread      4-thread
1024	.013 [+34%]  .007 [+25%]  .004 [+21%]   .0099 [+30%]  .0053 [+28%]  .0030 [+28%]
1152	.015 [+62%]  .007 [+66%]  .005 [+36%]   
1280	.017 [+37%]  .009 [+29%]  .005 [+36%]   
1408	.019 [+69%]  .010 [+65%]  .006 [+46%]   
1536	.021 [+34%]  .011 [+23%]  .006 [+29%]   
1664	.023 [+69%]  .012 [+62%]  .007 [+53%]   
1792	.026 [+34%]  .013 [+34%]  .008 [+21%]   .0200 [+20%]  .0104 [+19%]  .0061 [+16%]
1920	.027 [+65%]  .014 [+59%]  .008 [+58%]   .0212 [+47%]  .0111 [+45%]  .0064 [+38%]
2048	.029 [+24%]  .015 [+29%]  .009 [+29%]   .0231 [+22%]  .0121 [+21%]  .0070 [+19%]
2304	.033 [+53%]  .018 [+46%]  .010 [+46%]   
2560	.038 [+23%]  .020 [+21%]  .012 [+29%]   
2816	.044 [+52%]  .023 [+52%]  .013 [+49%]   
3072	.044 [+28%]  .024 [+21%]  .017 [+ 8%]   
3328	.051 [+58%]  .026 [+57%]  .015 [+49%]   
3584	.053 [+36%]  .029 [+31%]  .019 [+17%]   .0411 [+24%]  .0223 [+20%]  .0142 [+15%]
3840	.059 [+56%]  .031 [+50%]  .018 [+46%]   .0478 [+38%]  .0247 [+37%]  .0140 [+32%]
4096	.056 [+27%]  .032 [+21%]  .022 [+19%]   .0475 [+18%]  .0258 [+17%]  .0163 [+16%]
4608	.071 [+46%]  .038 [+43%]  .024 [+33%]   
5120	.070 [+29%]  .040 [+21%]  .028 [+28%]   
5632	.090 [+52%]  .048 [+50%]  .031 [+31%]   
6144	.084 [+33%]  .048 [+25%]  .042 [+11%]   
6656	.107 [+53%]  .056 [+49%]  .037 [+31%]   
7168	.102 [+39%]  .058 [+29%]  .048 [+19%]   .0797 [+26%]  .0451 [+20%]  .0400 [+18%]
7680	.125 [+54%]  .066 [+49%]  .042 [+36%]   .0996 [+39%]  .0525 [+35%]  .0331 [+25%]
8192	.121 [+25%]  .070 [+29%]  .057 [+19%]   .0922 [+28%]  .0522 [+22%]  .0443 [+25%]
----------------------------------------------------------------------------------------
Avg || Scaling:          1.877x       2.975x        ----           1.866x        2.938x

AvgGain, AVX
vs SSE2:    1.314x       1.280x       1.200x         1.588x        1.539x        1.371x

Avg per-cycle throughput gain, Haswell vs Sandy Bridge:
            1.437x       1.394x       1.321x         1.292x     1.264x     1.231x

Next up: Swap out the slower memory cannibalized from my SB quad for use in the above tests and replace it with the speedier RAM I bought from Newegg along with the rest of my new toys. I'll only be doing the AVX-build timing tests there.

[Prime95]

Originally Posted by ewmayer

Next up: Swap out the slower memory cannibalized from my SB quad for use in the above tests and replace it with the speedier RAM

Don't forget to set RAM to one of the XMP profiles in the BIOS. The BIOS boot screen will tell you the current memory speed.