Haswell Preview Benchmark - Page 16

Prime95 · 2013-07-05, 21:30

Quote:

Originally Posted by ewmayer

George, have you noticed any temperature impact from using FMA in your development code?

I haven't checked temperatures yet using FMA FFTs. I do expect them to be higher.

The other thing to note is the the small in-place FFTs should have higher temps than the blend test. The blend test usually uses lots of RAM and bigger FFTs. The RAM access introduces stalls where the FPU has a chance to idle for a few clocks.

Oliver has given me a nice goal -- make prime95 as brutal as HPL :)

ewmayer · 2013-07-05, 21:59

@George: "If you don't see smoke, you're not there yet, dude."

Oliver may simply need to run a SuperCooler even without overclocking the CPU.

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

General thoughts on the commodity-CPU FMA offerings:

Apparently the reason Intel was willing to throw that much FMA hardware into their newest chips is that arithmetic-performing hardware is actually a very small % of die area of modern general-purpose processors. We are long past the days when multipliers represented huge chunks of silicon and "multiply is expensive" was the standard thought process, both in terms of die area and latency.

I do wish Intel or AMD had done one obvious enhancement which can be done with a miniscule cost in Silicon: fused floating-point add/sub, i.e. inputs x,y, return x+y and x-y together. (There is something that sounds like that in the SSE/AVX spec, but I mean doubling the add/sub throughput thereby, not merely returning e.g. (+,-+,-) in an output register in place of pure-vector-sum or pure-difference. Fused +- can be done relatively cheaply since the operand unpacking and mantissa-shift only need be done once for each input pair, then feed the normalized operands to fast integer-style add/sub hardware. Repacking the results must of course be done separately, but all this takes miniscule hardware compared to adding a big honking multiplier. As it is, we're doing many more MULs than we need to simply because that's the easiest way to get at the doubled add/sub throughput afforded by 2-per-clock-pipelined SIMD FMA.

Example: A with-twiddle radix-16 complex DFT needs 174 ADD and 84 MUL. (4*15 = 60 of those MULs are in the 15 complex twiddle-MULs ... the 16th, or better, 0th twiddle is unity so is a no-op). Targeting Intel AVX2 FMA it makes sense to convert all those ADDs to FMAs, leading to 174 FMAs, after suitable restructuring of the algorithm. Those 174 FMAs can be broken down as follows:

- 8 have a unity multiplicand, i.e. the only reason we use FMA for the them is because the CPU can issue 2 SIMD FMA per cycle, versus just one SIMD ADD per cycle;

- 166 come in add/sub pairs of form a*b +- c, i.e. we compute the same product twice because there is no option to compute it just once, followed (or accompanied) by a paired add/sub.

So basically what this means is that FMA causes global warming. :)

TheJudger · 2013-07-05, 22:43

Quote:

Originally Posted by Prime95

Oliver has given me a nice goal -- make prime95 as brutal as HPL :)

*hehe* Overclockers will either hate you or try to tell you that the new prime95 version is buggy: "Hey, my system has no problems to run prime95 v26.6 @ 4.5GHz but it crashes with v28.something. This must be a software bug! v28.something can't put more load on the CPU than v26.6 because v26.6 shows already 100% CPU usage in taskmanager!"

Oliver

ewmayer · 2013-07-06, 23:06

[Note all key elements/compounds below have their thermal conductivity appended in SI standard W/K/m, which in words expresses "(heat transfer rate in watts) per (temperature gradient in degree Kelvin per meter) per (unit area in meter squared)" form].

Attention all "hot chips fans" - it occurred to me last month that nearly all thermal pastes on the market perform dismally compared to what is desired. The ideal thing would be a seamless high-conducting metallic interface between CPU and heatsink/fan, but instead we are offered a rogue's gallery of "thermal pastes" or "greases", based on goop like silicone [0.5-10 W/K/m, depending on make and user gullibility :], which is also used in things like window caulk, i.e. as a thermal *insulator*. These goops start out with a heat conductivity typically two order of magnitude less than copper [which conducts at around 400 W/K/m], and Sekrit Magickal Additives like "microscopically fine industrial diamonds" (diamond of course is record setter here at ~1000-2000 W/K/m) allay this negligibly, because they swim in the thermally insulating paste. In other words, "it's the substrate, stupid".

That got me to thinking along the lines of low-T-melting metals. Hg [surprisingly low thermal conductivity for a metal of 8 W/K/m] is out due to toxicity and nonwetting behavior ... next up was Ga [40 W/K/m], which melts at around 36C, i.e. is "close to liquid" at room T. Ga is also nontoxic, but has other issues including rapid oxidation and its propensity to "poison" metals it comes into contact with. There is a commercial alloy called Galinstan (named based on its component metals Ga, In, Sn) which melts in the -10-0C range, and presumably addresses the oxidization issue via the addition of the tin [67 W/K/m], but that melting point is too low for my book. You want something that just starts to melt around 70-150C, i.e. around the T range of a hot chip. Indium melts ~130C, has a decently high thermal conductivity of 82 W/K/m) is close, and its high deformability at room T and slightly above makes it a favorite of electronics hobbyists who have home-brewed solutions to the above problem for years. In is also relatively nontoxic and has the advantage that it will deform at hot-chip T to allow it to gap-fill even to the microscopic level.

At this point you would be right in wondering "why the hell are people still bothering with shitty nonmetallic thermal pastes?" Well, this outfit apparently wondered the same thing.

Anyone who reads this and takes the plunge, please do report your results here. (My Haswell currently not running hot enough to justify further outlay ... FMA-enhanced code will hopefully change that).

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

Aside: Interesting factoid: the thermal conductivity of water @0C is actually 4x higher in the solid (ice, ~2 W/K/m) phase than the liquid. But note this ignores the enhanced thermal conductivity due to convection (when geometry admits it) in water. The reason I bring up the latter point is that in our present application we desire the filled gap to be as thin as possibility, which rules out convection-enhanced heat transfer in the liquid phase.

TheMawn · 2013-07-06, 23:31

The thing is, we're not using very much paste. At the end, you should have a layer of paste fractions of a millimeter thick. Yes, it would be nice to have a single seamless diamond crystal filling in the gaps, but it wouldn't help very much.

You have a linear network of thermal resistances which includes the resistance between the die and the heat spreader, the resistance between the heat spreader and the heatsink, and the resistance between the base of the heatsink and the fins.

No matter what kind of alien thermal tech you use as paste, you can't improve the conductivity of the spreader and of the heatsink design, which is why your "insulator" pastes actually perform as well as 99% silver pastes.

ewmayer · 2013-07-06, 23:59

Quote:

Originally Posted by TheMawn

You have a linear network of thermal resistances which includes the resistance between the die and the heat spreader, the resistance between the heat spreader and the heatsink, and the resistance between the base of the heatsink and the fins.

A linear network with one big fat obvious (albeit "waffer theen") resistor in there.

Quote:

No matter what kind of alien thermal tech you use as paste, you can't improve the conductivity of the spreader and of the heatsink design, which is why your "insulator" pastes actually perform as well as 99% silver pastes.

The conductivity of heatsink/spreader, typically Al/Cu, respectively) are not the limiting factors here. It's the stuff making it hard for the heat to *get* to all that wonderfully-conducting stuff that's the problem.

Think about your "silver paste" comment - what is the conductivity of the paste in question? Now compare that to the conductivity of pure silver. See the difference? That "tiny 1%" non-silver stuff has a disproportionate impact.

Here is a simple kitchen experiment one can easily make to literally "get a feel" for how much impact even a thin layer of low-conducting material can make: the next time you happen to have a pot of nearly-boiling water on the stove, dust your fingertips with a thin layer of talcum powder or corn starch and touch them to the exterior of the pot., and see how long it takes until things begin to feel painfully hot. Now rinse the powder off and with wet fingers touch the pot. Of course thermal grease will conduct much better than talcum powder - might be interesting to compare with water, while you're at the stove - but the conductivity *ratio* between talc/water is not dissimilar to that of thermal-grease vs copper.

Anyway, no one is forcing you to buy anything ... but for someone like Oliver who is unable to run his CPU even at stock speed, that 10x improvement in conductivity of the interface might make all the difference.

kladner · 2013-07-07, 02:03

This is a step beyond-
http://www.frozencpu.com/products/54..._x3.html?tl=g8

http://www.xbitlabs.com/articles/coo...roundup_7.html

However, I don't think Xbitlabs got a good application of the metal pad critters. There's too much difference between cores, and the metal does not look like the examples of demonstrably good applications I saw. The description of the burn-in period I read on another site, was positively scary, especially for non-Intel CPUs. That piece said you need to get the material up near 100 C. It gave specific instructions for stopping your HS fan or your water pump while monitoring temps with Speed Fan. The CPU surface needs to be perfectly horizontal or the metal will flow unevenly. If it does not flow correctly, it forms the wrong kind of metallic structure. I think the term used was "bi-phase". It made me think of crystallized solder joints. He was also specific that chickening out and not getting it hot enough would result in bad metallic structure.

I'll have to look further to see if I can find that report again. The author was an absolute fanatic, doing multiple applications of each substance and evaluating the coverage. He took multiple samples, discarded the high outlier, I think, and noted absolute best application results by temperatures achieved.

Xyzzy · 2013-07-07, 09:55

The metal cap (heat spreader) covering the core(s) uses a thermal compound.

http://www.tomshardware.com/forum/27...emoval-results

Click on "See full content" in the first post for pictures.

chip · 2013-07-07, 15:26

RightMark Multi-Threaded Memory Test v1.1
i7-2600K 4500, 4x4 DDR3-1866 9-10-10-27-1T, thread, [all]

16384KB, Read w/PF, 128-bit SSE2
1-core: 21000MB/s, [21000MB/s]
2-core: 13000MB/s, [26000MB/s]
3-core: 8800MB/s, [26400MB/s]
4-core: 6600MB/s, [26400MB/s]

16384KB, Write NT, 128-bit SSE2
1-core: 22700MB/s, [22700MB/s]
2-core: 14400MB/s, [28800MB/s]
3-core: 9400MB/s, [28200MB/s]
4-core: 7000MB/s, [28000MB/s]

kladner · 2013-07-07, 17:24

Quote:

Originally Posted by kladner

I'll have to look further to see if I can find that report again. The author was an absolute fanatic, doing multiple applications of each substance and evaluating the coverage. He took multiple samples, discarded the high outlier, I think, and noted absolute best application results (and appearance) by temperatures achieved.

More digging shows that there are two different products. The one I saw the detailed review of seems to be Indigo Xtreme. I am fairly certain that the review I saw was on a site called skinneelabs.com.
http://skinneelabs.com/thermal-paste/

Unfortunately, this site and its owner seem to have disappeared from the web. There are laments on OC sites about his absence. There are still some demo videos on YouTube, but not for the phase change metal thermal pads. http://www.youtube.com/user/SkinneeLabs

However, the Indigo site has the sort of instructions I was remembering.
http://indigo-xtreme.com/documentation.html

This performance chart has a skinneelabs watermark-
http://www.frozencpu.com/images/awar...erallTempa.jpg

It's too bad that the Coollaboratory Liquid MetalPad Thermal Interface does not appear in the chart. From other sources it seems that their Liquid PRO Thermal Interface Material has decent performance. As always, it is hard to get clear comparisons because application methods and skill vary widely among reviewers.

Both the Indigo and Coollaboratory products are at FrozenCPU, but the Liquid Pro is out of stock.
http://www.frozencpu.com/cat/l1/g8/T...Interface.html

ewmayer · 2013-07-07, 19:27

Quote:

Originally Posted by kladner

This is a step beyond-
http://www.frozencpu.com/products/54..._x3.html?tl=g8

http://www.xbitlabs.com/articles/coo...roundup_7.html

However, I don't think Xbitlabs got a good application of the metal pad critters.

Thanks...the 82 W/K/m they mention for the higher-T stuff exactly matches that for pure Indium metal, and the "preheat with hairdryer to apply to chip cap" also points to a melting point matching that of In. So there the key is to get a high-quality "soldered joint" with the Indium foil.

Quote:

Originally Posted by Xyzzy

The metal cap (heat spreader) covering the core(s) uses a thermal compound.

http://www.tomshardware.com/forum/27...emoval-results

Click on "See full content" in the first post for pictures.

*That* would be an ideal application for liquid-at-room-T metal alloy, because it's factory-sealed.

2013-07-05, 21:59	#167
ewmayer ∂²ω=0 Sep 2002 República de California 19·613 Posts	@George: "If you don't see smoke, you're not there yet, dude." Oliver may simply need to run a SuperCooler even without overclocking the CPU. ------------------------ General thoughts on the commodity-CPU FMA offerings: Apparently the reason Intel was willing to throw that much FMA hardware into their newest chips is that arithmetic-performing hardware is actually a very small % of die area of modern general-purpose processors. We are long past the days when multipliers represented huge chunks of silicon and "multiply is expensive" was the standard thought process, both in terms of die area and latency. I do wish Intel or AMD had done one obvious enhancement which can be done with a miniscule cost in Silicon: fused floating-point add/sub, i.e. inputs x,y, return x+y and x-y together. (There is something that sounds like that in the SSE/AVX spec, but I mean doubling the add/sub throughput thereby, not merely returning e.g. (+,-+,-) in an output register in place of pure-vector-sum or pure-difference. Fused +- can be done relatively cheaply since the operand unpacking and mantissa-shift only need be done once for each input pair, then feed the normalized operands to fast integer-style add/sub hardware. Repacking the results must of course be done separately, but all this takes miniscule hardware compared to adding a big honking multiplier. As it is, we're doing many more MULs than we need to simply because that's the easiest way to get at the doubled add/sub throughput afforded by 2-per-clock-pipelined SIMD FMA. Example: A with-twiddle radix-16 complex DFT needs 174 ADD and 84 MUL. (415 = 60 of those MULs are in the 15 complex twiddle-MULs ... the 16th, or better, 0th twiddle is unity so is a no-op). Targeting Intel AVX2 FMA it makes sense to convert all those ADDs to FMAs, leading to 174 FMAs, after suitable restructuring of the algorithm. Those 174 FMAs can be broken down as follows: - 8 have a unity multiplicand, i.e. the only reason we use FMA for the them is because the CPU can issue 2 SIMD FMA per cycle, versus just one SIMD ADD per cycle; - 166 come in add/sub pairs of form ab +- c, i.e. we compute the same product twice because there is no option to compute it just once, followed (or accompanied) by a paired add/sub. So basically what this means is that FMA causes global warming. :) Last fiddled with by ewmayer on 2013-07-05 at 22:54

2013-07-07, 02:03	#172
kladner "Kieren" Jul 2011 In My Own Galaxy! 2·3·1,693 Posts	This is a step beyond- http://www.frozencpu.com/products/54..._x3.html?tl=g8 http://www.xbitlabs.com/articles/coo...roundup_7.html However, I don't think Xbitlabs got a good application of the metal pad critters. There's too much difference between cores, and the metal does not look like the examples of demonstrably good applications I saw. The description of the burn-in period I read on another site, was positively scary, especially for non-Intel CPUs. That piece said you need to get the material up near 100 C. It gave specific instructions for stopping your HS fan or your water pump while monitoring temps with Speed Fan. The CPU surface needs to be perfectly horizontal or the metal will flow unevenly. If it does not flow correctly, it forms the wrong kind of metallic structure. I think the term used was "bi-phase". It made me think of crystallized solder joints. He was also specific that chickening out and not getting it hot enough would result in bad metallic structure. I'll have to look further to see if I can find that report again. The author was an absolute fanatic, doing multiple applications of each substance and evaluating the coverage. He took multiple samples, discarded the high outlier, I think, and noted absolute best application results by temperatures achieved. Last fiddled with by kladner on 2013-07-07 at 02:08

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Haswell-E Prelim. Benchmark	sdbardwick	Hardware	37	2015-02-10 18:49
Prime95 and Haswell	Pleco	Information & Answers	22	2014-07-13 16:03
Haswell Rig	Mini-Geek	Hardware	64	2014-05-27 13:22
Prime95 version 27.1 early preview, not-even-close-to-beta release	Prime95	Software	126	2012-02-09 16:17
Missing mouse-over preview text	retina	Forum Feedback	1	2011-09-12 15:32

2013-07-06, 23:06	#169
ewmayer ∂²ω=0 Sep 2002 República de California 11647₁₀ Posts	[Note all key elements/compounds below have their thermal conductivity appended in SI standard W/K/m, which in words expresses "(heat transfer rate in watts) per (temperature gradient in degree Kelvin per meter) per (unit area in meter squared)" form]. Attention all "hot chips fans" - it occurred to me last month that nearly all thermal pastes on the market perform dismally compared to what is desired. The ideal thing would be a seamless high-conducting metallic interface between CPU and heatsink/fan, but instead we are offered a rogue's gallery of "thermal pastes" or "greases", based on goop like silicone [0.5-10 W/K/m, depending on make and user gullibility :], which is also used in things like window caulk, i.e. as a thermal insulator. These goops start out with a heat conductivity typically two order of magnitude less than copper [which conducts at around 400 W/K/m], and Sekrit Magickal Additives like "microscopically fine industrial diamonds" (diamond of course is record setter here at ~1000-2000 W/K/m) allay this negligibly, because they swim in the thermally insulating paste. In other words, "it's the substrate, stupid". That got me to thinking along the lines of low-T-melting metals. Hg [surprisingly low thermal conductivity for a metal of 8 W/K/m] is out due to toxicity and nonwetting behavior ... next up was Ga [40 W/K/m], which melts at around 36C, i.e. is "close to liquid" at room T. Ga is also nontoxic, but has other issues including rapid oxidation and its propensity to "poison" metals it comes into contact with. There is a commercial alloy called Galinstan (named based on its component metals Ga, In, Sn) which melts in the -10-0C range, and presumably addresses the oxidization issue via the addition of the tin [67 W/K/m], but that melting point is too low for my book. You want something that just starts to melt around 70-150C, i.e. around the T range of a hot chip. Indium melts ~130C, has a decently high thermal conductivity of 82 W/K/m) is close, and its high deformability at room T and slightly above makes it a favorite of electronics hobbyists who have home-brewed solutions to the above problem for years. In is also relatively nontoxic and has the advantage that it will deform at hot-chip T to allow it to gap-fill even to the microscopic level. At this point you would be right in wondering "why the hell are people still bothering with shitty nonmetallic thermal pastes?" Well, this outfit apparently wondered the same thing. Anyone who reads this and takes the plunge, please do report your results here. (My Haswell currently not running hot enough to justify further outlay ... FMA-enhanced code will hopefully change that). ---------------------- Aside: Interesting factoid: the thermal conductivity of water @0C is actually 4x higher in the solid (ice, ~2 W/K/m) phase than the liquid. But note this ignores the enhanced thermal conductivity due to convection (when geometry admits it) in water. The reason I bring up the latter point is that in our present application we desire the filled gap to be as thin as possibility, which rules out convection-enhanced heat transfer in the liquid phase.

2013-07-06, 23:31	#170
TheMawn May 2013 East. Always East. 3277₈ Posts	The thing is, we're not using very much paste. At the end, you should have a layer of paste fractions of a millimeter thick. Yes, it would be nice to have a single seamless diamond crystal filling in the gaps, but it wouldn't help very much. You have a linear network of thermal resistances which includes the resistance between the die and the heat spreader, the resistance between the heat spreader and the heatsink, and the resistance between the base of the heatsink and the fins. No matter what kind of alien thermal tech you use as paste, you can't improve the conductivity of the spreader and of the heatsink design, which is why your "insulator" pastes actually perform as well as 99% silver pastes.

2013-07-07, 09:55	#173
Xyzzy "Mike" Aug 2002 5×17×97 Posts	The metal cap (heat spreader) covering the core(s) uses a thermal compound. http://www.tomshardware.com/forum/27...emoval-results Click on "See full content" in the first post for pictures.

2013-07-07, 15:26	#174
chip "Eugene" Mar 2013 Ukraine 3 Posts	RightMark Multi-Threaded Memory Test v1.1 i7-2600K 4500, 4x4 DDR3-1866 9-10-10-27-1T, thread, [all] 16384KB, Read w/PF, 128-bit SSE2 1-core: 21000MB/s, [21000MB/s] 2-core: 13000MB/s, [26000MB/s] 3-core: 8800MB/s, [26400MB/s] 4-core: 6600MB/s, [26400MB/s] 16384KB, Write NT, 128-bit SSE2 1-core: 22700MB/s, [22700MB/s] 2-core: 14400MB/s, [28800MB/s] 3-core: 9400MB/s, [28200MB/s] 4-core: 7000MB/s, [28000MB/s]