Haswell Preview Benchmark

[Prime95]

Originally Posted by Prime95

Which do you think is better?

Surprisingly, in prime95's case it is the two instruction version with latency 8!!

Prime95 executes this code in a loop:

loop:
mul by DWT weight       Latency 5
add carry                     Latency 3
round value, calc next carry, apply inverse weight
goto loop

In the latency 8 case the next mul-by-DWT-weight can execute in parallel with the previous round-value,etc. code (essentially making the mul free). In the fused-mul-add case, the next FMA cannot start until the next carry has been computed. Thus, what I thought would be one of my most promising remaining Haswell optimizations (using FMA in the carry propagation code) turns out to be a dud.

[Prime95]

I just got the replacement CPU for the one I RMA'ed. I fired it up at 4.3GHz, 1.2V, DDR3-2400 and it is running a prime95 torture test (OK after 10 minutes). Temps are in the mid-80s doing large FFTs, mid-90s doing small FFTs.

Tomorrow, I'll play more with different voltage/speed combinations.

[kracker]

Originally Posted by Prime95

I just got the replacement CPU for the one I RMA'ed. I fired it up at 4.3GHz, 1.2V, DDR3-2400 and it is running a prime95 torture test (OK after 10 minutes). Temps are in the mid-80s doing large FFTs, mid-90s doing small FFTs.

Tomorrow, I'll play more with different voltage/speed combinations.

Good!

[sdbardwick]

That reminded me: Anandtech did some testing at various memory speeds and timings (but didn't include Prime95, unfortunately).

My TL;DR take is that 2400 at C9 is a good performance/time-spent-tweaking compromise.

[TheMawn]

Loving the memory overclock stats!

I might try looking to bring my 10-12-12-31 @ 2400 MHz down to 9-11-11-30 or something...

[kracker]

Originally Posted by TheMawn

Loving the memory overclock stats!

I might try looking to bring my 10-12-12-31 @ 2400 MHz down to 9-11-11-30 or something...

I can never run LL's overclocked, I always do DC. I guess paranoia of errors.

I mean, look at it this way: Even if you had 24 or 48 hours of stressing you're going to run more with constant LL's for how long. Probably rare, but still...

[db597]

Originally Posted by Prime95

I just got the replacement CPU for the one I RMA'ed. I fired it up at 4.3GHz, 1.2V, DDR3-2400 and it is running a prime95 torture test (OK after 10 minutes). Temps are in the mid-80s doing large FFTs, mid-90s doing small FFTs.

Does your system increase the voltage by +0.1V when AVX is in use? In which case your 1.2V could actually be 1.3V instead?

On my 4770K, after the Haswell optimisations, I can't get 4.3GHz stable at 1.2V (total after the AVX voltage bump is added). 10mins is OK, but overnight it would crash.

[Prime95]

My original Haswell did not bump 0.1V -- it has a BIOS option to turn off adaptive voltage. My new Haswell, the RMA'ed one, does bump the voltage.

Thus, when I reported 4.3GHz at 1.2V -- it was really running at 1.3V. Heat was an issue -- mid 80s during LL, 90s during torture.

Both CPUs are now running at 4.0GHz, 1.22V and 1.2V respectively, and DDR3-2400. The original running at 1.22V does have occasional unexplained reboots.

BTW, memory timings are weird. CPU-Z's reading of the SPD table was different than what the BIOS was setting the timings to from the SPD table. I manually changed the timings to the slightly slower one CPU-Z thinks is correct and my rare ROUNDOFF errors have disappeared (for now).

[xtruder]

I've seen mentions of speed measuring and CPU utilization at certain memory speeds in this thread, but how do i check those values? Iam using version 2.81.

I've been getting rounding errors on certain blocks in blend and have no idea how to interpret them, just to name a few:
- Rounding was 0.5 (or 0.455078125), expected less than 0.4 (fft 3360)
Whenever i increase CPU input voltage, iam getting this error

- Final result was BC12AB39, expected: 874392AC (fft 3584)
Giving 100MHz more uncore fixes this error, or lowering ring bus voltage at previous uncore setting.

Is there any sense in running particular fft size for more than 15 minutes? In blend, the fft sizes in question getting executed for only for 3 minutes. If i run them manually i get rounding errors 10-30min in.

[Prime95]

Originally Posted by xtruder

I've been getting rounding errors on certain blocks in blend and have no idea how to interpret them

You are not alone. I too am at a loss as to how to best address various overclocking problems in Haswell. The options are numerous, reduce the CPU clock, increase CPU voltage, uncore voltage, mem voltage, tweak memory timings, etc.

Prime95 does nothing to diagnose what your problem is -- it only tells you there is some kind of problem.

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.

[firejuggler]

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