Haswell Preview Benchmark

[xtruder]

Originally Posted by Prime95

In your particular case, the rounding errors are happening with small FFTs (assuming the FFT size was 3584 and not 3584K). FFTs 4K and below in size do not stress main memory much, meaning the problem is likely in the CPU or caches.

Sorry, forgot to put K behind it, those are large ffts iam having trouble with. Those particular ones come up after 5 hours blend (where it failed), so iam just running those to see what the deal is

Iam doing this on my CAD machine, stability goes above all. I already dropped from 4.6 tot 4.4 and may need bring it down more.

[db597]

Originally Posted by firejuggler

So, the easiest way for now is to not overclock Haswell?

It's still possible to overclock Haswell for long term endurance P95 testing. What I experienced was:

Using v28, a slight overclock around 4.0-4.1GHz is sustainable at sane voltages (<1.25V) and temperatures.

Using v27, an overclock of around 4.3-4.4Ghz is sustainable at sane voltages (<1.25V) and temperatures.

** assuming a decent heatsink. The above using DDR2133. If you have more memory bandwidth, I suppose the stress on the CPU will increase further.

[ewmayer]

Originally Posted by db597

It's still possible to overclock Haswell for long term endurance P95 testing. What I experienced was:

Using v28, a slight overclock around 4.0-4.1GHz is sustainable at sane voltages (<1.25V) and temperatures.

Using v27, an overclock of around 4.3-4.4Ghz is sustainable at sane voltages (<1.25V) and temperatures.

Precisely the point I made a few weeks ago in this thread - v27 [sans FMA] yields 5-10C lower temps, so should be more overclockable, or more stable at stock if v28 is too much even non-OCed.

Users should pick whichever option yields best overall stable throughput.

[TheMawn]

Just a shot in the dark...

Could it be that with different versions or clock speeds, the memory bottleneck is more or less severe putting more or less stress on the CPU?

I'm imagining the CPU doing a lot of waiting and frantic work and then more waiting and then more frantic work as it waits for data from the RAM and then rushes to do it with higher clock speeds and coding optimizations. This, as opposed to a more continuous and slower pace.

A runner doing 6 meters per second at a continuous pace versus a running doing 8 meters per second but stopping for one of every four seconds. Could it be then that the CPU can't handle the bursts of speed but CAN handle the slower, continuous pace?

[henryzz]

Originally Posted by TheMawn

Could it be then that the CPU can't handle the bursts of speed but CAN handle the slower, continuous pace?

Or the opposite.
I suspect a stress test needs to handle both situations.

[ewmayer]

Intel Delays Next-Generation Broadwell Chips Until 2014 Due to Manufacturing Issue - Mac Rumors

Intel yesterday confirmed that it is delaying the production of its next-generation Broadwell processors on account of a manufacturing issue. Intel CEO Brian Krzanich announced the news during Intel's earnings call, stating that production on the chips will begin during the first quarter of 2014 (via PCWorld).

Intel's Broadwell chips are designed to be the successor to its existing Haswell chips, manufactured on a 14-nanometer process as opposed to Haswell's current 22-nanometer process.

[figure caption]
Intel ran into some problems with the 14-nanometer process used to manufacture the chips and will have to fix them before it can resume production, CEO Brian Krzanich said during Intel’s earnings call on Tuesday.

Intel normally releases new chips like clockwork on an annual basis, and the manufacturing problems are a rare misstep for the company. Krzanich said there were problems with the "yield"—or the number of good chips the company gets per silicon wafer.

The Broadwell chips, which will eventually find their way into Apple's line of MacBook Airs and MacBook Pros, are said to be 30 percent faster and more power-efficient than Haswell, leading to even greater increases in battery life and performance. The inclusion of Haswell chips in Apple's MacBook Air boosted battery life to 12 hours on the 13-inch version and nine hours on the 11-inch version.

Luckily I'm not one of the folks eagerly waiting for this mobile incarnation of Haswell, but I expect a lot of MacAir fans [as in prospective-owners-of, not CPU-cooling-airflow-drivers :] will be disappointed.

I notice the process-size ratio for this shrink is appreciably (~11%) greater than the sqrt(2) which I believe serves as the long-term average of these. I wonder if that had something to due with the mfg problems.

[TheMawn]

14 nanometers. Holy crap. I find this is quite a leap from 22 also. At this rate, AMD is looking to be a giant in the wrong way.

[Mark Rose]

Originally Posted by TheMawn

14 nanometers. Holy crap. I find this is quite a leap from 22 also. At this rate, AMD is looking to be a giant in the wrong way.

To put that in perspective, a copper atom is about 140 pm in radius. So at 22 nm, features are only ~150 atoms wide.

[TheMawn]

Oh I know full well what 14 nanometers is in comparison to the atoms themselves. The "holy crap" was meant as a comment on the massive die shrink. For some reason I had thought 19 nm was the next one.

[only_human]

http://en.wikipedia.org/wiki/14_nanometer

Semiconductor
manufacturing
processes
10 µm — 1971
3 µm — 1975
1.5 µm — 1982
1 µm — 1985
800 nm — 1989
600 nm — 1994
350 nm — 1995
250 nm — 1997
180 nm — 1999
130 nm — 2002
90 nm — 2004
65 nm — 2006
45 nm — 2008
32 nm — 2010
22 nm — 2012
14 nm — 2013
10 nm — est. 2015
7 nm — est. 2017
5 nm — est. 2019
Half-nodes

It looks like Intel intends do try for 10 nm without Extreme Ultraviolet Lithography
http://en.wikipedia.org/wiki/Extreme...et_lithography

At the 2013 EUVL Workshop, Intel announced that EUV would still be under development in 2015, and hence would be targeted for 2017 7 nm HVM. Consequently, 10 nm would be carried out with ArF immersion multiple patterning.[178] TSMC[179] and GlobalFoundries[180] have made similar statements.

[nucleon]

After much stuffing around. I have my haswell up and running. (i5 4670k)

I left everything on auto. Just changed ratio for turbo modes.

4.2GHz - the CPU tries 1.4V-ish, hits 100degC on occasion, cpu clocks down. Eventually dies with display corruption.

4.0GHz - the CPU voltage selects around 1.231-1.236V, hits 87degC under intel burn in test, 90degC under prime95 default torture test (so far only about 10mins testing) Display corruption and OS hang on stopping torture test :)

I'll leave it at 4.0GHz burning for a while.

Benchmark (4GHz/2400 ram):

Compare your results to other computers at http://www.mersenne.org/report_benchmarks
Intel(R) Core(TM) i5-4670K CPU @ 3.40GHz
CPU speed: 3723.34 MHz, 4 cores
CPU features: Prefetch, SSE, SSE2, SSE4, AVX, AVX2, FMA
L1 cache size: 32 KB
L2 cache size: 256 KB, L3 cache size: 6 MB
L1 cache line size: 64 bytes
L2 cache line size: 64 bytes
TLBS: 64
Prime95 64-bit version 28.1, RdtscTiming=1
Best time for 768K FFT length: 2.702 ms., avg: 2.860 ms.
Best time for 896K FFT length: 3.332 ms., avg: 3.713 ms.
Best time for 1024K FFT length: 3.766 ms., avg: 3.994 ms.
Best time for 1280K FFT length: 4.785 ms., avg: 5.467 ms.
Best time for 1536K FFT length: 5.903 ms., avg: 6.373 ms.
Best time for 1792K FFT length: 7.027 ms., avg: 7.486 ms.
Best time for 2048K FFT length: 7.972 ms., avg: 8.684 ms.
Best time for 2560K FFT length: 10.133 ms., avg: 10.189 ms.
Best time for 3072K FFT length: 12.230 ms., avg: 12.534 ms.
Best time for 3584K FFT length: 14.712 ms., avg: 14.975 ms.
Best time for 4096K FFT length: 16.788 ms., avg: 17.002 ms.
Best time for 5120K FFT length: 20.999 ms., avg: 21.211 ms.
Best time for 6144K FFT length: 26.264 ms., avg: 27.713 ms.
Best time for 7168K FFT length: 30.541 ms., avg: 31.295 ms.
Best time for 8192K FFT length: 35.081 ms., avg: 36.452 ms.
Timing FFTs using 2 threads.
Best time for 768K FFT length: 1.419 ms., avg: 1.558 ms.
Best time for 896K FFT length: 1.788 ms., avg: 2.138 ms.
Best time for 1024K FFT length: 1.999 ms., avg: 2.170 ms.
Best time for 1280K FFT length: 2.560 ms., avg: 2.610 ms.
Best time for 1536K FFT length: 3.120 ms., avg: 3.193 ms.
Best time for 1792K FFT length: 3.744 ms., avg: 4.274 ms.
Best time for 2048K FFT length: 4.276 ms., avg: 4.878 ms.
Best time for 2560K FFT length: 5.388 ms., avg: 6.052 ms.
Best time for 3072K FFT length: 6.511 ms., avg: 6.785 ms.
Best time for 3584K FFT length: 7.742 ms., avg: 7.928 ms.
Best time for 4096K FFT length: 8.839 ms., avg: 9.518 ms.
Best time for 5120K FFT length: 11.165 ms., avg: 11.218 ms.
Best time for 6144K FFT length: 13.849 ms., avg: 14.098 ms.
Best time for 7168K FFT length: 16.145 ms., avg: 17.232 ms.
Best time for 8192K FFT length: 18.595 ms., avg: 18.748 ms.
Timing FFTs using 3 threads.
Best time for 768K FFT length: 0.991 ms., avg: 1.286 ms.
Best time for 896K FFT length: 1.205 ms., avg: 1.235 ms.
Best time for 1024K FFT length: 1.422 ms., avg: 1.450 ms.
Best time for 1280K FFT length: 1.833 ms., avg: 2.298 ms.
Best time for 1536K FFT length: 2.298 ms., avg: 2.759 ms.
Best time for 1792K FFT length: 2.686 ms., avg: 3.262 ms.
Best time for 2048K FFT length: 3.190 ms., avg: 3.953 ms.
Best time for 2560K FFT length: 3.946 ms., avg: 4.755 ms.
Best time for 3072K FFT length: 4.766 ms., avg: 4.824 ms.
Best time for 3584K FFT length: 5.716 ms., avg: 5.758 ms.
Best time for 4096K FFT length: 6.437 ms., avg: 6.525 ms.
Best time for 5120K FFT length: 8.272 ms., avg: 8.729 ms.
Best time for 6144K FFT length: 10.080 ms., avg: 10.150 ms.
Best time for 7168K FFT length: 11.684 ms., avg: 12.614 ms.
Best time for 8192K FFT length: 13.558 ms., avg: 14.366 ms.
Timing FFTs using 4 threads.
Best time for 768K FFT length: 0.827 ms., avg: 0.876 ms.
Best time for 896K FFT length: 1.025 ms., avg: 1.402 ms.
Best time for 1024K FFT length: 1.183 ms., avg: 1.563 ms.
Best time for 1280K FFT length: 1.894 ms., avg: 2.513 ms.
Best time for 1536K FFT length: 2.051 ms., avg: 2.149 ms.
Best time for 1792K FFT length: 2.334 ms., avg: 3.158 ms.
Best time for 2048K FFT length: 2.864 ms., avg: 3.515 ms.
Best time for 2560K FFT length: 3.506 ms., avg: 3.920 ms.
Best time for 3072K FFT length: 4.303 ms., avg: 4.501 ms.
Best time for 3584K FFT length: 5.057 ms., avg: 5.441 ms.
Best time for 4096K FFT length: 5.914 ms., avg: 5.992 ms.
Best time for 5120K FFT length: 7.454 ms., avg: 7.539 ms.
Best time for 6144K FFT length: 9.185 ms., avg: 9.314 ms.
Best time for 7168K FFT length: 10.625 ms., avg: 31.948 ms.
Best time for 8192K FFT length: 12.193 ms., avg: 12.712 ms.
Best time for 61 bit trial factors: 1.714 ms.
Best time for 62 bit trial factors: 1.772 ms.
Best time for 63 bit trial factors: 1.948 ms.
Best time for 64 bit trial factors: 1.866 ms.
Best time for 65 bit trial factors: 2.294 ms.
Best time for 66 bit trial factors: 2.667 ms.
Best time for 67 bit trial factors: 2.637 ms.
Best time for 75 bit trial factors: 2.569 ms.
Best time for 76 bit trial factors: 2.566 ms.
Best time for 77 bit trial factors: 2.565 ms.

Also I'll try discrete vid card.

-- Craig