Haswell Preview Benchmark

[Prime95]

Originally Posted by sdbardwick

1 Worker      8.2ms
2 Workers     8.7ms
3 Workers    10.32ms
4 Workers    12.50ms

Ouch. To restate your findings:

1 Worker     122 iter/s
2 Workers    230 iter/s
3 Workers    291 iter/s
4 Workers    320 iter/s

Some serious memory bandwidth problems at stock settings which will only get worse with FMA3 optimizations.

I'm purchasing DDR3-2400 for my Haswell build. That should help somewhat. In an ideal world I should get 50% more memory bandwidth (right?) which means up to 320*1.5 = 480 iter/s in the 4 worker case. This is close to 4 times the single worker case, 4*122 = 488 iter/s. Even in this idealistic analysis, overclocking the CPU and FMA3 improvements will mean improvements in the one worker case which means I'm again bandwidth limited in the 4 worker case.

[chalsall]

Originally Posted by Prime95

Even in this idealistic analysis, overclocking the CPU and FMA3 improvements will mean improvements in the one worker case which means I'm again bandwidth limited in the 4 worker case.

This is why we all enjoy working with you George.

You push the envelope.

[sdbardwick]

Originally Posted by Prime95

[...]
I'm purchasing DDR3-2400 for my Haswell build. That should help somewhat. In an ideal world I should get 50% more memory bandwidth (right?) which means up to 320*1.5 = 480 iter/s in the 4 worker case. This is close to 4 times the single worker case, 4*122 = 488 iter/s. Even in this idealistic analysis, overclocking the CPU and FMA3 improvements will mean improvements in the one worker case which means I'm again bandwidth limited in the 4 worker case.

The delta in bandwidth from the higher frequency will be less than expected - in some cases it can be a lot less. As the frequency increases, so do the memory sub-timings (CAS, etc.) which can cause latency issues; my 1600MHz RAM will provide lower iteration times than some cheap 1866MHz modules.

EDIT:
It's late, and I think I am playing a little too lose with the technical details. The maximum bandwidth will increase by 50%, but the effective bandwidth will be less. The pipe can pour out (64 x 2400 x 100000) bits per second, but you might need to wait a bit (latency) for the flow to start.

[Chuck]

Originally Posted by Prime95

I'm purchasing DDR3-2400 for my Haswell build. That should help somewhat.

Could you please tell us the main components (MOBO, CPU, RAM) you plan to use?

[Prime95]

Originally Posted by Chuck

Could you please tell us the main components (MOBO, CPU, RAM) you plan to use?

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.

[fivemack]

The heatsink that comes in the box with the 4770 is not capable of keeping the machine cold enough to run at above 2.7GHz when running eight threads of gnfs-lasieve4I14e.

[sdbardwick]

Originally Posted by fivemack

The heatsink that comes in the box with the 4770 is not capable of keeping the machine cold enough to run at above 2.7GHz when running eight threads of gnfs-lasieve4I14e.

The standard cooler is shockingly weak but my boxed 4770K HSF allows 4x LL jobs @3.6GHz without thermal throttling. Did one of the <expletive deleted> push-pins come loose?

[ewmayer]

Originally Posted by fivemack

The heatsink that comes in the box with the 4770 is not capable of keeping the machine cold enough to run at above 2.7GHz when running eight threads of gnfs-lasieve4I14e.

Would you expect a 4670 to run any cooler?

[sdbardwick]

Originally Posted by ewmayer

Would you expect a 4670 to run any cooler?

At a given clock speed, sure. With 2MB less cache and no hyperthreading it should run cooler (despite having the same 84W TDP). HT (at least on SB/IB) is an effective heat generator.

[Batalov]

One thing that got me (I've used the stock cooler just now on the 3570K for a week while the CM612 cooler came in the mail*; and then before that, last time I used the similar stock cooler was 5 years ago ... so I forgot the trick) was that you are only supposed to push the pins and not do anything else, when you mount.

Instead I also turned the thingies 90 degrees according to their arrows. (And apparently I am not the only idiot who doesn't RTFM - see some "manuals" on the web: they say that these are the arrows that will help you make it "tight".) That's not what you want to do - this is for dismounting. Luckily it worked fine for the week and the cooler didn't just pop off in the middle of the night.

I could have waited for the Haswell, but then I didn't. Wasn't sure about the price and/or the newest mobo childhood diseases like that previous generation nightmare... But of course, it doesn't happen all the time. And when George will write some magic that will make Haswells fly around the 3rd gen in circles, we will all be sitting here, biting our elbows. ;-)

_________
* I had the Tuniq Tower (the "old faithful"), but the mounting holes have now changed compared to the Q6600, so I had to burn in the with the stock cooler

[Xyzzy]

It costs way too much but the backplate and screws on our Noctua cooler makes life bearable.

http://www.newegg.com/Product/Produc...82E16835608019

The OEM pushpin thingy is almost worse than the Pentium FDIV bug. Seriously, what were the engineers thinking? Oh right, cost.

Even if the Noctua was not more efficient than the stock cooler we would buy it!

They are pretty good about having adapter plates for various socket configurations, too.