Sandy Bridge vs Haswell AVX2 L-L Completion Times
 

Hi,

With better performance achieved with Haswell over Ivy Bridge and Sandy Bridge in Prime95 especially because of AVX2, is there really a considerable increase in completion times of L-L tests?

For example my X5470 3.33GHz quad core (2 models above Q9650 eqv) was showing an ETA of ~27 days (24/7) to complete a 77M number. Assuming an i5-2500 would take ~21 days, would an i5-4430 (i5-2500 eqv in terms of performance) finish many days faster because of AVX2 or would it only make a 2 to 3 day difference?

Thank you,

On my Haswell-E 5960X with 8 cores on a 77M exponent (4M FMA FFT) it is showing ETA ~59h after the first 100K iterations (~2.76 ms/iter)

If I disable AVX2 (FMA3) with CpuSupportsFMA3=0 in local.txt and start over, it is using 4M AVX FFT. Now it is showing ETA ~ 67hours after the first 100K iterations (~3.13 ms/iter).

[QUOTE=ATH;449071]On my Haswell-E 5960X with 8 cores on a 77M exponent (4M FMA FFT) it is showing ETA ~59h after the first 100K iterations (~2.76 ms/iter)

If I disable AVX2 (FMA3) with CpuSupportsFMA3=0 in local.txt and start over, it is using 4M AVX FFT. Now it is showing ETA ~ 67hours after the first 100K iterations (~3.13 ms/iter).[/QUOTE]

I really appreciate this comparison, even though it makes me itch to switch.....to Intel. Maybe for Xmas. :unsure:

[QUOTE=ATH;449071]On my Haswell-E 5960X with 8 cores on a 77M exponent (4M FMA FFT) it is showing ETA ~59h after the first 100K iterations (~2.76 ms/iter)

If I disable AVX2 (FMA3) with CpuSupportsFMA3=0 in local.txt and start over, it is using 4M AVX FFT. Now it is showing ETA ~ 67hours after the first 100K iterations (~3.13 ms/iter).[/QUOTE]

Thank you!

This type of information is truly a gem which seems to be very hard to find around here in terms of ms/iter benchmarks!

59h is truly an amazing ETA and something I would love to have but I'm only going to be getting an i5 at most I think.

From your comparison, your Haswell-E 5960X with 8 cores and AVX2 is getting ~59h ETA and with AVX2 disabled is getting ~67h ETA. If we compare a similar performing Sandy Bridge CPU (E5-2687W is kind of close but 5960X is much better but if we assume a fictional Sandy Bridge CPU equally matching the 5960X in performance for the sake of numbers) it should also be completing it in ~67h because of not having AVX2 but with AVX2 (Haswell and newer) the ETA is ~59h! That's a ~12% increase in performance which is quite substantial!

With AVX2 Assuming 24/7 then 720h in a month / 59h is ~= 12.2 tasks of ~77M exponent
With AVX2 Assuming 24/7 then 8760h in a year / 59h is ~= 148.5 tasks of ~77M exponent

Without AVX2 Assuming 24/7 then 720h in a month / 67h is ~= 10.75 tasks of ~77M exponent
Without AVX2 Assuming 24/7 then 8760h in a year / 67h is ~= 130.75 tasks of ~77M exponent

AVX2 does have an impact (positive) on the completion times of L-L testing but the increase although there, doesn't seem to justify the cost of a Haswell i5 over a Sandy Bridge i5 in my case!

An equally performing Sandy Bridge processor to a Haswell or newer (AVX2) processor would only be ~12% slower then because of not having AVX2? Unless my math is wrong I could really save with a Sandy Bridge setup!

[QUOTE=danmur;449110]
This type of information is truly a gem which seems to be very hard to find around here in terms of ms/iter benchmarks![/QUOTE]

[URL="http://www.mersenne.org/report_benchmarks/"]More benchmarks than you would ever want... [/URL]

I tried to remove AVX as well with CpuSupportsAVX=0 in local.txt and now the ETA is 119 hours. So:
77M AVX2: ~59h
77M AVX ~67h
77M no AVX ~ 119h

[QUOTE=sdbardwick;449111][URL="http://www.mersenne.org/report_benchmarks/"]More benchmarks than you would ever want... [/URL][/QUOTE]

Thanks! I've checked out those benchmarks but I'm not sure what the ms values are supposed to represent? For example under 68.13M to 77.91M (4096K) there are all of those ms values. Do they represent the time to complete 1 iteration?

[QUOTE]I tried to remove AVX as well with CpuSupportsAVX=0 in local.txt and now the ETA is 119 hours. So:
77M AVX2: ~59h
77M AVX ~67h
77M no AVX ~ 119h[/QUOTE]Thanks!

Even without AVX that's incredibly fast and something that I dream of having instead of an i5 which I don't even have yet! Maybe a Sandy Bridge 2011 8 core but the motherboards are too expensive! It's important though that you mention AVX as that was only introduced with Sandy Bridge (as I've just found out). My current X5470 doesn't even have AVX so going to an equally performing Sandy Bridge processor from what I have now will provide a big performance increase just there with AVX! AVX2 does provide a big increase over AVX but the cost of a Haswell CPU is much greater than a Sandy Bridge CPU and I would definitely get great results with an upgrade to Sandy Bridge from my current system.

Thanks again

[QUOTE=danmur;449115]Thanks! I've checked out those benchmarks but I'm not sure what the ms values are supposed to represent? For example under 68.13M to 77.91M (4096K) there are all of those ms values. Do they represent the time to complete 1 iteration?

Thanks!

Even without AVX that's incredibly fast and something that I dream of having instead of an i5 which I don't even have yet! Maybe a Sandy Bridge 2011 8 core but the motherboards are too expensive! [/QUOTE]

Bargainhardware.com has dual 8-core Sandy Bridge machines with 64G memory for £300 - they are old enough that data centres have just hired people to haul the racks away and sell the individual machines. I have four.

[QUOTE=fivemack;449116]Bargainhardware.com has dual 8-core Sandy Bridge machines[/QUOTE]

Do you mean [url]http://www.bargainhardware.co.uk/[/url] ?

It looks like Bargainhardware.com is a parked domain.

[QUOTE=fivemack;449116]Bargainhardware.com has dual 8-core Sandy Bridge machines with 64G memory for £300 - they are old enough that data centres have just hired people to haul the racks away and sell the individual machines. I have four.[/QUOTE]

Thanks that's a great site with really good deals on E5-2600 machines but sadly I'm forced to configure all of the options (can't select 0 or 1 CPUS but have to select 2 CPUS) and it comes out a little too expensive ~$285. There are Dell 2011 single socket boards for ~$100 which would be a better way to go for me as the total would be around $215 but I could buy it in individual parts and it would consume a little less power.

[QUOTE=danmur;449110]An equally performing Sandy Bridge processor to a Haswell or newer (AVX2) processor would only be ~12% slower then because of not having AVX2? Unless my math is wrong I could really save with a Sandy Bridge setup![/QUOTE]

No. There is a more pressing factor than AVX2, and that is the memory bandwidth. Sandybridge fastest supported memory is DDR3 1333, while Ivy Bridge / Haswells support DDR3 1600. This makes a straight up 20% increase in thruput for Haswells, even before you factor in AVX2 gains.

However Skylakes support DDR4 2133. If you can, go for the cheapest skylake.