My old '07 vintage Core 2 Quad is probably dead. Somehow I got it running stably with 2 GB of RAM, but that's far from ideal. Assuming I can't get it fully operational again, I'm looking at a new machine for daily use, development, and crunching (in that order of priority.)

Normally what I do when I build a machine is to find a good, brand name, pre-built machine on sale that has good upgrade opportunities. For many years that's been cost effective (I'm rarely interested in having the fastest, bleeding edge technology.) However, I'm finding now that it's difficult to find a desktop that's made with standard parts, and I need to be able to install a full size GPU and replace the power supply.

Accordingly, for the first time in a number of years I'm looking at building the system rather than buying.

I'm not looking to build "The Mother of All Crunchers". Crunching is only part of what it will be used for, and I only crunch during the cooler part of the year. Cost definitely IS a concern.

The old computer is donating the following parts:

600W power supply (brand new).
GTX 460 GPU
2x 1TB SATA 7200 RPM drives.
DVD burner.

I'll be installing Windows 7 Professional 64 bit on this system.

New parts:

Case: Cooler Master HAF 912
http://www.amazon.com/Cooler-Master-HAF-912-Computer/dp/B00BCXF6O4/
http://www.newegg.com/Product/Product.aspx?Item=N82E16811119233

I like CM's HAF series cases because their focus is on cooling. This isn't a huge case, but it's large enough. It comes with two 120 mm fans, but has room for up to 6 120 mm fans or 2 120 mm fans and 2 200 mm fans. I envision having up to four fans in the case (200 mm in front and on top, 120 mm in back, and possibly 120 mm on the side blowing on the GPU), and the motherboard coincidentally has headers for 4 case fans, so that works well.

One minor drawback is that the front panel ports are USB 2.0 rather than 3.0, so if I want USB 3.0 on the front I'd need to add them later.

I'm guessing this is one of their newer cases, since despite its low price it has a nice selection of features such as 2.5" SSD drive bays, top mounts for either 2x120 mm fans (or a 240 mm radiator) or a 200 mm fan, and a rotatable/removable 3.25" drive cage, and removable intake filters for the front and bottom fan openings.

The price is certainly right: $59.99.


Motherboard: ASUS Z87-Plus
http://www.amazon.com/Asus-Z87-PLUS-DDR3-1600-Motherboard/dp/B00CRJSX5Q/
http://www.newegg.com/Product/Product.aspx?Item=N82E16813131980

I've always had good luck with Asus, and this seems like a decent middle of the road motherboard. Lots of SATA ports if I need it, tons of USB ports, etc. It seems to have decent overclocking abilities, although that's not a priority right now.

$159.

CPU: Core i5 4670K
http://www.amazon.com/Intel-i5-4670K-Quad-Core-Desktop-Processor/dp/B00CO8TBOW/
http://www.newegg.com/Product/Product.aspx?Item=N82E16819116899

This was a tough decision, as I was considering everything from a Haswell Core i3 through an Ivy Bridge-E Core i7 4820K.

I wanted four real cores, so the Core i3 was out. The choice between two Core i7s -- the Ivy Bridge-E 4820K and the Haswell 4770K -- was a tough one. They're about the same price, but the 4820K uses quad channel memory while the 4770K is dual channel. LLR is usually bottlenecked by memory bandwidth so the 4820 is very tempting. However, motherboards for the 4820 are more expensive, so even though the CPUs are about the same price, the entire system will be more expensive. In the end, some benchmarks posted here with an Ivy Bridge-E running in both dual and quad channel mode indicate that while the quad channel memory is important when running six cores, it's not that big of a factor when running 4 cores.

Once the decision was made to go with Haswell and dual channel memory, I decided a Core i5 was a better choice. It's about $100 cheaper than the Core i7, and I won't be using hyperthreading for crunching anyway. The only real disadvantage is that the L3 cache is 6 MB vs 8 MB.

I was going to get the Core i5 4670, but the unlocked 4670K is only a few dollars more and it doesn't pay not to get the K version. The only disadvantage is the "K" chips lack one of the VM features (VT-X, I think), which I believe means that you can't run a 64 bit guest OS inside of a 32 bit host OS. Since I'll be running 64 bit Windows, that shouldn't be a problem.

$225

(Note: At least at the beginning, I'll be using the Intel stock cooler. Depending on temps, I can add a better air or water cooler later on.)

Memory: Corsair Vengeance 8 GB ( 2 x 4 GB ) DDR3 1600 MHz (PC3 12800) 240-Pin DDR3 Memory Kit
http://www.amazon.com/Corsair-Vengeance-240-Pin-Platforms-CMZ8GX3M2A1600C9/dp/B004CRSM4I/
http://www.newegg.com/Product/Product.aspx?Item=N82E16820233144

This ram is on the motherboard's QVL (qualified vendor list). 8 GB is enough for now, and I can always fill the other pair of slots with another 8GB if I need. I like having finned heatsinks on the ram for better cooling.

$87

Comments?
____________
My lucky number is 75898⁵²⁴²⁸⁸+1

Looks for a nice machine ;) I've got the Cooler Master HAF X 922, and it's really nice. It's huge, and currently has 3x 200mm fans, and can fit a dual 120mm fan radiator. It's got plenty of room for a cpu heatsink, i've got a cooler master hyper 212+. It is definitely upgrade friendly, with a bottom mounted psu and plenty of room for a gfz card. I like the look, too.

I can't comment much about the i5-4670(k), but I have an i5-2500k and it's nice and surprisingly cool. I've heard the newer chips (haswell, ivy, etc) don't o/c as well as sandy.

As for the RAM, it seems they are most stable at speeds one step under what they are rated for. I'd probably get 1866 RAM, some 1866 corsair RAM is actually less expensive that the 1600 you listed. http://www.amazon.com/dp/B008HK4ZAG

Looks really nice, good luck.
____________

275*2^3585539+1 is prime!!! (1079358 digits)
Proud member of Aggie the Pew

As for the RAM, it seems they are most stable at speeds one step under what they are rated for. I'd probably get 1866 RAM, some 1866 corsair RAM is actually less expensive that the 1600 you listed. http://www.amazon.com/dp/B008HK4ZAG

Nice setup. Just repare for an aftermarket cooler if you want to use the k on the cpu and/or crunch avx WU's on the four cores.

Point taken.

After reading reviews (I assumed ram was ram; apparently Corsair isn't as reliable as it used to be) Corsair is out and G.skill is in. New ram:

http://www.amazon.com/G-skill-RipjawsX-DDR3-1866-Memory-F3-14900CL9D-8GBXL/dp/B004JM1ZG8/

$88

Mike, since you are getting an ASUS board, their AI software has a great Auto Overclocking set up nowadays. It will automatically set the highest STABLE clocks and voltages. Depending on the quality of the CPU that would range from 3.9 Ghz to 4.2 for core i5 or i7 SB, IB, or Haswell. Temps will of course be higher but not anywhere close to trying to get the best manual overclock settings. I'd use it before deciding on what aftermarket cooler you eventually decide on. I personally would go with a water cooling setup because they have less CPU temperature variance vs. ambient temps.

Edit: The Auto Overclocking setup does take temps into account so the result will be higher with better cooling.
____________
Largest Primes to Date:
As Double Checker: SR5 109208*5^1816285+1 Dgts-1,269,534
As Initial Finder: SR5 243944*5^1258576-1 Dgts-879,713

I have a HAF-XM case for my 3770K. I like it a lot and the airflow is good... 2x200mm fans (one front, one on top), 2x120mm on the side blowing on the GPUs, and 2x120 on the rear (set up push-pull) for the CPU (liquid) cooler radiator. One minor gripe is that sometimes the side panel doesn't want to latch tightly... it tends to "pop out" 1/4 inch or so on one lip and I have to fiddle with it to get it to stay put. No biggee. It's marketed as a "mid-tower", but it is at the large end of that range.

+1 on the faster memory.

I seem to remember that back when we were shaking out the first versions of LLRavx, temps were running 3-5C warmer than with the prior LLR, all other things being constant (not to say that's surprising). That was with SB chips only, at that time.

Good luck and let us know how it goes. Post photos! (a.k.a. geek porn)

--Gary

If 1883 ram is better for stability at 1600, then why not buy 2400 speed memory and have LOTS of overhead? The difference in price isn't very large.

With that in mind, I found this at Newegg:

http://www.newegg.com/Product/Product.aspx?Item=N82E16820231574

What's interesting is that it comes with some sort of dual fan gizmo for cooling the ram.

Never saw that before.

(Upon searching further, there's a handful of memory fans out there.)
____________
My lucky number is 75898⁵²⁴²⁸⁸+1

If 1883 ram is better for stability at 1600, then why not buy 2400 speed memory and have LOTS of overhead? The difference in price isn't very large.

With that in mind, I found this at Newegg:

http://www.newegg.com/Product/Product.aspx?Item=N82E16820231574

What's interesting is that it comes with some sort of dual fan gizmo for cooling the ram.

Never saw that before.

(Upon searching further, there's a handful of memory fans out there.)

If 1883 ram is better for stability at 1600, then why not buy 2400 speed memory and have LOTS of overhead? The difference in price isn't very large.

With that in mind, I found this at Newegg:

http://www.newegg.com/Product/Product.aspx?Item=N82E16820231574

I'm using this:

http://www.newegg.com/Product/Product.aspx?Item=N82E16820231585

Working great so far!

Buy a Xeon 1230 V3. 4 cores / 8 threads, 8 MB cache ( this model is without iGPU ). In boinc, especially in PG ( where cpu temp are much bigger than in other project ) there is small margin for overclocking. So, "K" version isn't as good as somewhere else.

You can easily @ it to 3.7 GHz ( all 8t running ) on Z87 based motherboard. In addition, fast memory and you're good to go. :)

I was in the same place you are, like 3/4 months ago.
My config:

Xeon 1230 v3 @ 3.7
Msi Z87M-G43
2x4 G Skill 2400 MHz 10/12/12/31

Some PSP ( HT off ):
http://scr.hu/0rlv/vmry9

"*Intel Xeon Processor Family is designed for servers. Some features may not support when installed on 8 series chipsets. For more details, refer to ASUS support site at http://support.asus.com."

If 1883 ram is better for stability at 1600, then why not buy 2400 speed memory and have LOTS of overhead? The difference in price isn't very large.

With that in mind, I found this at Newegg:

http://www.newegg.com/Product/Product.aspx?Item=N82E16820231574

Pulled the trigger and ordered some parts:

PCPartPicker part list / Price breakdown by merchant / Benchmarks

CPU: Intel Core i5-4670K 3.4GHz Quad-Core Processor ($219.99 @ Amazon)
Motherboard: Asus Z87-PLUS ATX LGA1150 Motherboard ($159.99 @ Newegg)
Memory: G.Skill Ares Series 8GB (2 x 4GB) DDR3-2400 Memory ($76.99 @ Newegg)
Case: Cooler Master HAF 912 ATX Mid Tower Case ($59.99 @ TigerDirect)
Case Fan: Cooler Master Megaflow 110.0 CFM 200mm Fan ($8.99 @ Newegg)
Other: SilverStone 140mm Fan Filter with Magnet for Case Fan/Power Supply Fan and Panel Air Vent FF141B ($8.99)
Total: $568.93
(Prices include shipping, taxes, and discounts when available.)
(Generated by PCPartPicker 2014-02-14 23:26 EST-0500)

From my point of view. After changing from triple to quad channel you should see a 25% improvement in peak bandwidth. But remember this is only the peak performance.

I had decided to stick with 1866 memory so the memory could run at 1.5 volts. That's all the CPU officially supports. I ended up buying 2400 speed memory. Why? I have the option of running faster if I want to, and if not I can always run the memory at 1866 and 1.5 volts rather than the 1.65 volts required for 2400. Finally, 2400 speed memory was $1 less expensive than the otherwise identical 1866 ram. Go figure.

The new computer is up and running, at least the hardware is. I still have a laundry list of software that needs to be installed before things are back to normal.

Some observations:

Cooling:

Those HAF (high airflow) cases really make a difference. I'm using the same GPU as before, and it's running a good 20 degrees C cooler than in the old computer. I didn't blow the dust out of it while the computer was apart, either. Wow. Just wow. The case is a Cooler Master HAS 912 with the front 120 mm fan moved to the side and a CM 200 mm fan in the front. The rear 120 mm exhaust fan is unchanged and the top vents are left open for passive convective cooling.

Overclocking:

The Asus motherboard comes with some interesting utilities. You can do a lot of the BIOS configuration stuff from within Windows, and you can restart directly into the BOIS without having to sit there pounding on the Del key like a bored chimpanzee in a lab experiment. That's a nice touch.

It has an "auto overclock tuning" button in Windows software. It works, sort of.

It happily raised the memory from 1600 MHZ to 2400 MHZ, and the CPU from 3.4 GHz to 4.6 GHz, while testing it "under load" as it gradually increased the frequencies and voltages.

Even with the stock cooler (something better is on order), the cpu temp remained ok.

Not surprisingly, their definition of "under load" does not mean "continuously running AVX instructions on all cores". Starting up LLR rapidly brought the CPU temps into the 80s and it hit 90 within a few seconds.

Back to stock speeds for now. :)

As expected, LLR runtimes (stock clocks) are about half of what they were with the Core2Quad Q6600.

Memory:

We all know that LLR and GFN are usually constrained by memory bandwidth. The Asus utilities show the CPU power consumed. Running PPSE (which uses cache more and memory less) consumes 5 to 10 watts more power than SR5 does, and the CPU runs a few degrees hotter as well.

Once I get the new cpu cooler, I'll try overclocking the memory first to see what effect that has.

Gripes:

Although I'm very happy with the purchases, there's a couple of things I don't like:

HAF 912: You seem to have to remove BOTH the left and right side panels in order to release the plastic clips holding the front bezel. Therefore, to clean the front air filter, you need to remove both sides as well as the front.

Asus Z87-Plus: Not sure if this is specific to this motherboard or if it affects all new "UEFI" BIOSs. You can't get into Widows Safe mode by hitting F8 anymore. F8 just brings you to a menu that lets you select which disk to boot from. To get to safe mode, you have to already be in Windows and run the msconfig program to configure Windows to boot into safe mode on the next boot.

I think the BIOS itself is putting the computer to sleep, above and beyond Windows' own power settings. I need to dig deeper into this. There are all sorts of power saving options, including powering down unused ports, but I think it's overdoing it.


Nice touches:

The motherboad has a power switch on the motherboard itself. It's a good idea to hook everything up on the bench before mounting it all in the case, so to actually turn it on without the case's power switch you usually need to short two pins on the motherboard to simulate the switch. Having an actual button makes it simpler.

The CPU wattage display is nice.

The system tells you how effective your cooling system is by displaying the degree/watt ratio, i.e., how much hotter your cpu gets for each extra watt used. The lower the number, the better the cooling. It's 0.59 degrees C per watt with the stock cooler.

Lots of control of the fans, including turning them off under low loads.

Acknowledgements:

First a very special thank you to a few people who, unasked, donated hardware to help me rebuild. That was incredibly generous. You know who you are!

I'd also like to thank everyone here for their help and comments. I spent a lot of time reading website reviews, but I also listened to the advice here and the advice here was very, very good.

____________
My lucky number is 75898⁵²⁴²⁸⁸+1

I just installed the software for the Cyberpower UPS this computer is plugged into. It shows how much power the UPS is supplying on the battery backed up plugs.

This computer can crunch LLR over twice as fast as the Core 2 Quad it replaced, yet the max power consumption dropped from about 390 watts to 265 watts. That's with the same GPU, and includes two monitors and a network switch.

That's a big reduction in power!
____________
My lucky number is 75898⁵²⁴²⁸⁸+1

After some timing experiments with my i5-3330, I have come to the conclusion that dual channel (DDR3-1333 CL9) is not sufficient for a quad core :-(

These are the rough timings from running SR5 workload:

1 instance: 2.5 ms/iter
2 instance: 2.6 ms/iter
3 instance: 3.1 ms/iter
4 instance: 3.7 ms/iter

As a result of this timing exercise, I have gone from running BOINC on all four cores to just 3 cores with about < 10% dop in aggregate throughput.

The result of "no difference" from the other thread must've been due to the humongous L3 cache|(15MB) which would be enough to hold two FFTs in their enitrety plus change.

Mike, do you have scaling timings similar to mine? Especially against different memory speeds?

PS:- I've ordered two sticks of DDR3-1600. Hopefully, that'll bring some improvement in performance.

Mike, do you have scaling timings similar to mine? Especially against different memory speeds?

Just allocate a mineral oil bath for the system, cooling problems solved!

:)

Looks like a nice box!
Can you post a summary of the final parts list, or is the original post up to date?
____________
My lucky #: 60133106^131072+1 (GFN 17-mega)

Just allocate a mineral oil bath for the system, cooling problems solved!

:)

Looks like a nice box!
Can you post a summary of the final parts list, or is the original post up to date?

This afternoon I replaced the Intel stock cooler with the Noctua NH-U14S (with a second fan in push-pull configuration). It's, um, slightly larger.

To say there's a big difference in temperatures is a significant understatement. With all 4 cores running LLR, it's about 20 degrees cooler AND the fans are barely spinning. Forget the temperatures; the computer is MUCH quieter than before. In fact, I had a scare when I powered it up because the BIOS screamed at me that there was a CPU fan failure. The fans were spinning way below the default warning threshold.

I decided to mount it vertically, with the exhaust pointing straight up through the vents in the top of the case. When I looked at the actual sizes, it just made no sense to have those 150 mm fans pointing at the rear 120 mm exhaust fan. Trying to connect a fire hose to a drinking straw is the analogy that comes to mind.

Did I mention it's quiet? I f I didn't hear the hard drives moving every now and then, the only way I'd know it's turned on is that I can see one of the CPU fans turning through the top vent. And it's crunching full bore at the moment.

I can thoroughly recommend this CPU cooler. Attaching it was easy (and the instructions are well written), it come with a 6 year warranty, and the fans have a 150,000 hour MTBF. The company has a history of providing free adapter kits to existing owners when new CPU sockets come out.

In other news, I swapped the GTX 460 with a superclocked GTX 580. Time to give that a test drive and see if it can run Genefer reliably with the factory overclocking.

Real numbers:

4 cores running PSP-LLR, GPU idle:

Asus AI Suite 3 is reporting a CPU temp of 52 degrees (core temp reports 64 degrees).
CPU fan (150 mm 1500 RPM Noctua) is spinning at 646 RPM
2nd CPU fan (150 mm 1200 RPM Noctua) is spinning at 531 RPM
Rear fan (120 mm Cooler Master) is spinning at 885 RPM
Side fan (120 mm Cooler Master) is spinning at 831 RPM
Front fan (200 mm Cooler Master) is spinning at 464 RPM<
____________
My lucky number is 75898⁵²⁴²⁸⁸+1

In other news, I swapped the GTX 460 with a superclocked GTX 580. Time to give that a test drive and see if it can run Genefer reliably with the factory overclocking.

GPU=GeForce GTX 580
Global memory=1610612736 Shared memory/block=49152 Registers/block=32768 Warp size=32
Max threads/block=1024
Max thread dim=1024 1024 64
Max grid=65535 65535 65535
CC=2.0
Clock=1594 MHz
# of MP=16
No project preference set; using AUTO-SHIFT=8

Starting initialization...
maxErr during b^N initialization = 0.0000 (14.286 seconds).
Testing 370268^1048576+1...
Estimated total run time for 370268^1048576+1 is 8:59:33

maxErr exceeded for 370268^1048576+1, 0.5000 > 0.4500
MaxErr exceeded may be caused by overclocking, overheated GPUs and other transient errors.
Waiting 10 minutes before attempting to continue from last checkpoint...

Try the memory first. I bet you can run it with the shader overclock with a lower memory setting.

Try the memory first. I bet you can run it with the shader overclock with a lower memory setting.

Friend of mine just found his pair of 580s in his closet. I need to see if I can talk him out of one...
____________
My lucky #: 60133106^131072+1 (GFN 17-mega)

Stock Intel cooler:



... next to the Noctua NH-U14S:



____________
My lucky number is 75898⁵²⁴²⁸⁸+1

Try the memory first. I bet you can run it with the shader overclock with a lower memory setting.

First thing I did was put everything back to stock -- or at least I tried to do that. I installed EVGA precision from the disk that came with the GTX 580, and it immediately said "There's a newer version!". So I installed the latest. Then I uninstalled the latest and went back to the old one because the new Precision won't let you lower the clocks below factory settings.

Even at stock, however, it still errored out, so the next thing I did was lower the memory clock about 10 % below stock. Now it seems to be running stable. I'll let it complete this GFN before playing with the clocks again.

Actually, I lied. I left out a step. Somewhere in there I set all the fans to full speed to see if it made a difference. Noise went up, temps went down. Error occurred at the exact same place.

During the GFN challenge I was able to OC ALL my Fermis by lowering the memory clocks 10-15%. With my GTX 460 (at 1884 Mhz) & 570 (at 1844 Mhz) I was able to OC 25-30% even without my A/C cooling setup I had at my old house, although my 580 OC'ed the least (at 1644 Mhz).

So memory clock IS the key.

After some timing experiments with my i5-3330, I have come to the conclusion that dual channel (DDR3-1333 CL9) is not sufficient for a quad core :-(

These are the rough timings from running SR5 workload:

1 instance: 2.5 ms/iter
2 instance: 2.6 ms/iter
3 instance: 3.1 ms/iter
4 instance: 3.7 ms/iter

As a result of this timing exercise, I have gone from running BOINC on all four cores to just 3 cores with about < 10% dop in aggregate throughput.

The result of "no difference" from the other thread must've been due to the humongous L3 cache|(15MB) which would be enough to hold two FFTs in their enitrety plus change.

Mike, do you have scaling timings similar to mine? Especially against different memory speeds?

PS:- I've ordered two sticks of DDR3-1600. Hopefully, that'll bring some improvement in performance.

Cores core 1 core 2 core 3 core 4 Iter/sec I/S/core Incremental Loss 1 8.515 117.4398121 117.4398121 n/a 0.00% 2 8.761 8.763 228.2583915 114.1291957 110.8185794 2.82% 3 9.495 9.488 9.509 315.8784081 105.2928027 87.6200166 10.34% 4 11.073 11.896 12.001 11.056 348.1466467 87.03666167 32.2682386 25.89%

Cores core 1 core 2 core 3 core 4 Iter/sec I/S/core Incremental Loss 1 7.893 126.6945395 126.6945395 n/a 0.00% 2 8.065 8.066 247.9697486 123.9848743 121.2752092 2.14% 3 8.336 8.350 8.334 359.7124921 119.904164 111.7427435 5.36% 4 9.248 8.882 9.244 9.048 439.4186831 109.8546708 79.70619097 13.29%

During the GFN challenge I was able to OC ALL my Fermis by lowering the memory clocks 10-15%. With my GTX 460 (at 1884 Mhz) & 570 (at 1844 Mhz) I was able to OC 25-30% even without my A/C cooling setup I had at my old house, although my 580 OC'ed the least (at 1644 Mhz).

So memory clock IS the key.

That certainly fits with everything we know (or think we know). When I get a chance I'll put the 580 back at factory clocks except for the memory.

Here's some real life results showing the difference the 2400 Mhz memory makes. This is the task page for the Core i5's PSP tasks. The first 4 at the bottom are with the memory set to 1600 MHz and the newer ones are with the memory running at 2400 MHz.

http://www.primegrid.com/results.php?hostid=424445&offset=0&show_names=0&state=0&appid=8

(If you're reading this sometime in the future, that link might not show anything. Sorry.)

The CPU time dropped from about 160K seconds to 136K seconds AND these newer tasks are also about 5% larger than the earlier tasks that took 160K seconds. (Ignore the elapsed time measurements; I have way too much junk running in the background.)


____________
My lucky number is 75898⁵²⁴²⁸⁸+1

Cores core 1 core 2 core 3 core 4 Iter/sec I/S/core Incremental Loss 1 7.893 126.6945395 126.6945395 n/a 0.00% 2 8.065 8.066 247.9697486 123.9848743 121.2752092 2.14% 3 8.336 8.350 8.334 359.7124921 119.904164 111.7427435 5.36% 4 9.248 8.882 9.244 9.048 439.4186831 109.8546708 79.70619097 13.29%

That looks like very interesting data.. how are you getting it?

I'm on my phone, so this will be a bit brief.

Your first command line looks correct. Is the llr64.exe executable in the directory where you're running the command? It may also be called cllr64.exe.

When it runs, you'll see the iteration timing numbers.

-b is for genefer, not llr.
____________
My lucky number is 75898⁵²⁴²⁸⁸+1

1, command prompt does not accept commands while running LLR Currently unable to run additional instances of either LLR or cmd.exe simultaneously for testing

2, the data I'm getting (time per bit) is that averaged over the total run time? If it is, that presents a potential for misleading data, if the run time is not great enough to average out the startup time you cited.

Except for the clock speed thing you seem to have figured it out. The clock speed problem is a bit bizarre.
____________
My lucky number is 75898⁵²⁴²⁸⁸+1

That makes sense. & have run some tests. But the results are puzzling. It was puzzling enough having the single core test not fully clock the CPU, but its actually going faster with 5 workunits than it was with just 1. (5.85 ms now)

if I run the test more than once does the system cache any part of the benchmark?

That running 5 units simultaneously yielded more daily units complete (by about 3/4ths a task) than running just 4 (1 per core)