Haswell Preview Benchmark

[Batalov]

I see that George is on the new wave (a water wave that is), but I wasn't yet ready for this so I wanted to stick with the air coolers for now. The CM TPC 612 (which some people here on the forum already mentioned) turned out to have a very sane mounting system. Seems to do the job decently.

In comparison, I still remember the Tower kindly (with its spring loaded brass screws and all); kept it on the shelf. Some other coolers that I'd had have those "pull with all your might and fit that square opening on that rectangular peg... which peg? that peg... ah wait you cannot see under the cooler" ;-) type of brackets where while you are mounting you can never be sure if that horrible cracking sound in the middle of the procedure was the death for your mobo or not. And you never will know until you will power up again with baited breath.

[Prime95]

Originally Posted by Prime95

Some serious memory bandwidth problems at stock settings.

Some more back-of-the-envelope calculations. Scott was doing 320 iter/s, 1600K FFT size, 8 bytes per FFT double, an FFT iter does 2 passes over the FFT data, times two since the FFT data must be both read and then written. Putting it altogether:

320 * 1600K * 8 * 2 * 2 = 16.77GB/sec

Max theoretical bandwidth of dual channel DDR3-1600 is 25.6GB/sec (the test I saw at Anand or Tom's Hardware showed 21.33GB/sec in practice).

The FFT must also read code, some sin/cos data, and FFT weights which I have not quantified.

[Xyzzy]

When we do calculations like this how do we account for Intel's "turbo" function?

http://www.intel.com/content/www/us/...echnology.html

Our i7 runs at 3.9GHz with one core loaded but drops to 3.6GHz when all four cores are used.

Wouldn't the frequency difference make more cores being used look artificially worse?

[Batalov]

Originally Posted by Prime95

Mobo ($110): http://www.newegg.com/Product/Produc...82E16813130695
CPU ($250): http://www.newegg.com/Product/Produc...82E16819116899
Ram ($82): http://www.newegg.com/Product/Produc...82E16820231587
Heatsink ($118): http://www.newegg.com/Product/Produc...82E16835146028
OS ($100): http://www.newegg.com/Product/Produc...82E16832416550
SSD ($150): http://www.amazon.com/dp/B006EKJ8UI/..._26725410_item
Case ($60): http://www.newegg.com/Product/Produc...82E16811139018
Power supply ($100): http://www.newegg.com/Product/Produc...82E16817182083

$10 rebates on the case and power supply. Save $20 buying the CPU and mobo as a combo. Save $5 purchasing the RAM and OS as a combo. I'll try to save another 1% using fatwallet.com.

If you are buying the Rosewill Fortress PSU, you could grab the case combo. http://www.newegg.com/Product/ComboD...=Combo.1326789
I have the hawk, and it is decent (the place for the water radiator is compatible with your choice) and has a convenient HD docking slot... Just a thought.

[ewmayer]

I had been buggin George about his Haswell-related shopping list for roughly the past week - I was unable to find better deals for CPU/mobo combos at Frys and TigerE, erm I mean, -Direct, so yesterday shamelessly plagiarized his selections of CPU/mobo/RAM in my Newegg order. George, thanks for the shopping list. This has the further advantage that we can use a team approach to any technical issue related to this particular hardware marriage.

[George noted that Microcenter had even better deals, but on an in-store-only basis, and the MC website indicated the closest store location to me (NoCal) is Orange county (SoCal), which I found surprising, as I would've expected an outfit like that to have at least one Silicon Valley location. But any prospective buyers should check their site for store locations near you.]

Total for CPU+MoBO+RAM+[CA-sales-tax]+shipping comes in below my (somewhat arbitrary, but still useful) personal price limit for this upgrade of $500. If I want to overclock the CPU I may have to shell out extra for a better cooling solution, but my price target was based on non-OCed CPU. Will first see how hot that runs in my ATX case, then take it (or not) from there. I hope to get some good cooling mileage via clever use of aerodynamics (and perhaps also of mirrors, as the magicians like to say).

I intend to first plug the RAM from my current Sandy Bridge quad (purchased certified-pre-owned from Mike last December) into the Haswell mobo, in order to get a set of comparative timings of my Mlucas code which takes memory speed out of the equation. After that I'll switch to the speedy newer memory.

After I do my first round of AVX->AVX2 code porting and FMA-related optimizations, I will contemplate attempting an ATX case surgery to mount both the SB and Haswell Mobos in the same case. If that proves too involved, I'll entertain offers for the SB gear. (George already expressed interest, by way of upgrading one piece of his home CPU farm).

[kladner]

Originally Posted by Batalov

I see that George is on the new wave (a water wave that is), but I wasn't yet ready for this so I wanted to stick with the air coolers for now. The CM TPC 612 (which some people here on the forum already mentioned) turned out to have a very sane mounting system. Seems to do the job decently.

Perhaps the TPC 812? The one with vertical vapor chambers? I paid about the same for the Hyper 612 as the 812 goes for. Wish I'd seen the 812 first. Mine is a nice enough cooler, but it makes connecting to the fan headers around the upper edges (in a mid-tower) challenging. Low profile RAM is preferable, too.

[Prime95]

Originally Posted by Batalov

If you are buying the Rosewill Fortress PSU, you could grab the case combo. http://www.newegg.com/Product/ComboD...=Combo.1326789
I have the hawk, and it is decent

I'm too cheap :)

BTW, you rightly point out that the radiator for my CPU cooler does NOT fit in very many cases. I've never used water cooling and wouldn't have this time except that all the reviews say Haswell runs HOT. I'm looking forward to trying something new.

BTW2, I calculated that the Platinum certified power supply is likely to save $20/year over a Bronze certified power supply -- more if I move my power-hungry GTX 570 to this box.

[pepi37]

If I may ask:will enabling AVX2 give performance boost, or that is all we can expect from latest Intel chip?

[ewmayer]

Originally Posted by pepi37

If I may ask:will enabling AVX2 give performance boost, or that is all we can expect from latest Intel chip?

Simply having an AVX2-capable chip means little, except insofar as Haswell improves overall performance via other means. The main potential benefit for Prime95-like code is the significantly better floating-point crunching capability conveyed by the addition of FMA (fused multiply-add) to the FPU instructions. Theoretically this could yield as much as roughly-doubled throughput, but only if the memory can keep the FPU pipes full, which is almost certain to not be the case. The question thus becomes how much of the theoretical improvement is realizable in practice.

My gut tells me this will be akin to what we saw for the SSE2->AVX upgrade: theoretically 2x more throughput per cycle, in practice 1.3-1.4x, and obviously only for code updated to use the new instructions.

The major reason I consider shelling out $400-500 for Haswell upgrade to be a no-brainer (again, this is speaking from the perspective of a float-heavy code developer) is that the AVX->AVX2 code-upgrade effort will be much less: I expect 1-2 months' work for the "big picture' optimizations, compared to the 5-6 months I've spent refactoring my SSE2 inline assembler to use AVX.

[pepi37]

Originally Posted by ewmayer

Simply having an AVX2-capable chip means little, except insofar as Haswell improves overall performance via other means......

My gut tells me this will be akin to what we saw for the SSE2->AVX upgrade: theoretically 2x more throughput per cycle, in practice 1.3-1.4x, and obviously only for code updated to use AVX.

So first we need to wait until code for AVX2 is written and implemented in Prime95, then we can do final benchmarking and say final words ( and go or not go to shop to buy it) :)

[ewmayer]

Originally Posted by pepi37

So first we need to wait until code for AVX2 is written and implemented in Prime95, then we can do final benchmarking and say final words ( and go or not go to shop to buy it) :)

Well, Scott "Speedy Gonzalez" Bardwick's Haswell timings above show a nice gain already for the unmodified code: I'm speaking of (hopefully) 1.3-1.4.x additional improvement versus that, i.e. for AVX-on-Haswell upgraded to AVX2/FMA-on-Haswell.

It's quite possible that the overall speedup due to the combination of [faster-overall chip, typically faster RAM in the newer systems, code using AVX2/FMA] on a cycle-per-cycle basis vs Sandy/Ivy Bridge will approach 2x. We shall see.