i was just wondering the difference in the number crunching device that was in the cpu. I mean if its not gonna make a difference or is it. with the price taken into consideration. quad core should be better i think. and i plan on getting a couple 9950s but eh i could go cheap. intel chips are to exspensive.

i was just wondering the difference in the number crunching device that was in the cpu. I mean if its not gonna make a difference or is it. with the price taken into consideration. quad core should be better i think. and i plan on getting a couple 9950s but eh i could go cheap. intel chips are to exspensive.

Be careful when choosing your chip. AMD chips have had a few flaws, so first off make sure you are getting a newer chip, not one that's been sitting on the shelf for a couple months. There have been revisions to fix math errors and memory cache errors. AMD has also increased the power requirements on some of it chips (not sure which off hand). The AMD's also perform lower than the Intel chips. For the few dollars difference, it's better to go Intel. Price it up both ways, while you can find the AMD chip cheaper, the board usually costs more than the Intel board, and it ends up being a wash. And avoid the tri-core chips, they don't do much (if any) better than a dual core Intel of the same speed.

A quad usually does 2 - 2½ times the work of a dual core of the same speed.
Just make sure you don't skimp on memory. If you are building a quad, you should go with a minimum of 3GB memory.

---

"Every passing hour brings the Solar System forty-three thousand miles closer to Globular Cluster M13 in Hercules -- and still there are some misfits who insist that there is no such thing as progress." - Kurt Vonnegut

9950 Black edition has had errors in it? i feel unsafe buying from ebay now as i cant tell a difference.

9950 Black edition has had errors in it? i feel unsafe buying from ebay now as i cant tell a difference.

The bug was fixed long ago. Model numbers ending in "00" had it, such as the 9500. Model numbers ending in "50", such as the 9550 are fixed.

The bug rarely shows itself in normal workloads. In fact, I don't know of anyone who has definitly seen it outside of AMD's internal testing. I have had a 9500 with the bug, and with no fix, running four Rosetta tasks 24/7 since Jan 19, and the bug has never shown itself.

Well, he said he may go cheap, so i brought up the issues, though the 9950 does not have the errors of previous AMD quads, there are a lot of people out there selling "cheap" quad chips that have these L3/floating point errors.

The big issue in my opinion, with the 9950 is it's 140W power consumption. Many "cheap" boards are only designed to handle the previous 95W and 125W chips. So heat, under-voltage, and incompatibility issues pop up. Even some low end board that claim compatibility with the 140W chip have issues from what I have seen.

As far as I know the 9950 black edition is free of the errors of previous chips, it should be pointed out that as top of the line AMD chip it only does marginally better than the cheaper entry-level Intel Q6600, which can be built into a low priced board easily, whereas with the AMD, and it's higher power consumption, low priced boards can be a crap shoot. In the end, to get the same performance it may be cheaper to go Intel depending on where you shop.

I'm not saying it's a bad choice to go with a 9950, just that if you look into it, you will probably find the Intel's are a better value. I would use the 9950 for a midgrade system, but anything lower end, Intel seems to be a better choice, and anything higher end Intel is the only choice.

---

"Every passing hour brings the Solar System forty-three thousand miles closer to Globular Cluster M13 in Hercules -- and still there are some misfits who insist that there is no such thing as progress." - Kurt Vonnegut

A quad usually does 2 - 2½ times the work of a dual core of the same speed.
Just make sure you don't skimp on memory. If you are building a quad, you should go with a minimum of 3GB memory.

A quad of the same spec as a dual (i.e. same architecture, cache and speed) will do nearly 2x the work, assuming enough RAM.

---

OK, so im doing my first build. is wattage and socket all that you have to worry about when adding a 9950 cpu? cause they recently lowered the wattage on 9950s to 125, yet i still see boards that only support say 9700 series say here for cpu http://cgi.ebay.com/AMD-Phenom-X4-9950-BLACK-EDITION-4mb-2-6GHz-AM2-125W_W0QQitemZ250317767608QQcmdZViewItem?hash=item250317767608&_trkparms=72%3A1424%7C39%3A1%7C66%3A3%7C65%3A12%7C240%3A1318&_trksid=p3286.c0.m14

^^9950 Phenom black edition

here for motherboard
http://cgi.ebay.com/MSI-K9A2-PLATINUM-AMD-790FX-ATX-Socket-AM2-Phenom-QC_W0QQitemZ380075303083QQcmdZViewItem?hash=item380075303083&_trkparms=72%3A1424%7C39%3A1%7C66%3A3%7C65%3A12%7C240%3A1318&_trksid=p3286.c0.m14

^^MSI K9A2 Plantinum AMD 790FX

Does wattage and socket make them compatible or are there other factors that im not seeing.

Does wattage and socket make them compatible or are there other factors that im not seeing.

you need bios support too - check here:
http://global.msi.com.tw/index.php?func=prodcpu2&prod_no=1332&maincat_no=1
or if it's version 2, then here:
http://global.msi.com.tw/index.php?func=prodcpu2&prod_no=1395&maincat_no=1

i'd have thought you'd be better off buying it from newegg? unless it comes in much cheaper than the buy it now price on ebay...

---

Tom's Hardware Guide did a review of system boards that work with the new AMD quads and the news was not good. This may be old news as the BIOS is likely one of the issues.

I would check Tom's Hardware Guide. They have system board info and they test lots of configurations.

I would only buy a 9950 if you are OK with the purchase of a new mother board as well if it does not work. You can also check with your system board vendor and see if they have a BIOS update for the 9950.

Best of Luck and keep crunching....

Paul

BTW - Tiger Direct has a Q6600 complete build with OS and 2GB of RAM for $369. I don't know what kind of money you are looking to spend but I am really thinking of buying one of these systems: http://www.tigerdirect.com/applications/SearchTools/item-details.asp?EdpNo=4199883

---

Thx!

Paul

Good information in this thread; thanks. I am interested in a new machine dedicated to BOINC, powered by a quad core and I have definitely noticed a direct correlation between processing speed/time to complete a task and "floating point" and "integer speeds. And yet I have seen no reference to floating point or integer in advertising for any CPU. Am I correct that I should definitely acquire a machine with the highest floating point/integer speeds in order to sharply reduce the time required to complete a task?

EXAMPLE: Intel 2.66ghz machines; one with say 1300 floating point and 2900 integer speeds will require perhaps 40% or more time to complete identical tasks than the same 2.66ghz with say 2300 floating point and 5000 integer speeds. Any ideas? Bill

Good information in this thread; thanks. I am interested in a new machine dedicated to BOINC, powered by a quad core and I have definitely noticed a direct correlation between processing speed/time to complete a task and "floating point" and "integer speeds. And yet I have seen no reference to floating point or integer in advertising for any CPU. Am I correct that I should definitely acquire a machine with the highest floating point/integer speeds in order to sharply reduce the time required to complete a task?

EXAMPLE: Intel 2.66ghz machines; one with say 1300 floating point and 2900 integer speeds will require perhaps 40% or more time to complete identical tasks than the same 2.66ghz with say 2300 floating point and 5000 integer speeds. Any ideas? Bill

floating point performance does impact (integer less so i think), but the cache size is also v important. Of course, the crunching ability increases almost directly with the number of cores and the CPU speed. A Pentium 4 with the same float/int benchmark scores will be outperformed by a Core2 machine with the same scores (although the core2 machine will have to be at a lot lower frequency to get the same benchmark scores!)

I think the order of performance for a given clock speed will be something like:

Core2
Phenom/Opteron
Core2-Pentium (i.e. 2180 etc)
Athlon64/Pentium M
AthlonXP
Pentium 3
Pentium 4

On the up-side for the P4, the clock speeds do go quite high. I'm not sure where a Pentium D would fit in there.

HTH
Danny

---

RE: Benchmarks (floating point/integer speeds)

Toms benchmarks can be found here

I would say this:

1) The Boinc benchmarks are more of an estimate than a hard set number. Some of the numbers are just down right wacky, Example: This is one of my machines where the integer speed is probably a bit high:
Host 925439 XD

2) Make sure to look at performance in many different categories, just because a chip might perform better in 1 or 2 of them, does not mean that it is better at everything. Some benchmarks, (Whetstone for example) only load the process into the cache of the CPU and execute it from there, never factoring in RAM access time, or speed of getting information to the CPU. Though a chip might be faster when reading from it's own cache, this is likely to be a rare occurrence in real world usage.

---

"Every passing hour brings the Solar System forty-three thousand miles closer to Globular Cluster M13 in Hercules -- and still there are some misfits who insist that there is no such thing as progress." - Kurt Vonnegut

Danny&Steven; Thanks much for the information; most helpful.
Bill

Hi All,

I've built 2 Systems over the past while and want to optimize them.

First one is a Tyan Thunder K8WE w/ 2 X Opteron 252 @ 2.6 GHz and 3 GB RAM.

Second one is Asus M3A78-T with Phenom Quad @ 3.0 with 8GB RAM. (Bios 8.02)

The Phenom machine does about 200 Work Units/day, and I will restart the Tyan
one tomorrow...

Both are 100% stock settings.

Both Run Fedora 10 @ 64 Bit.

Will I be best off running the Boinc Client that comes with Fedora, or should I get or make one that runs with SSE3 or other optimization at compile time?

[root@phenom ~]# rpm -qa | grep boinc
boinc-client-6.4.5-2.20081217svn.fc10.x86_64
boinc-manager-6.4.5-2.20081217svn.fc10.x86_64


Thanks!

John W.

Hi All,

I've built 2 Systems over the past while and want to optimize them.

First one is a Tyan Thunder K8WE w/ 2 X Opteron 252 @ 2.6 GHz and 3 GB RAM.

Second one is Asus M3A78-T with Phenom Quad @ 3.0 with 8GB RAM. (Bios 8.02)

The Phenom machine does about 200 Work Units/day, and I will restart the Tyan
one tomorrow...

Both are 100% stock settings.

Both Run Fedora 10 @ 64 Bit.

Will I be best off running the Boinc Client that comes with Fedora, or should I get or make one that runs with SSE3 or other optimization at compile time?

[root@phenom ~]# rpm -qa | grep boinc
boinc-client-6.4.5-2.20081217svn.fc10.x86_64
boinc-manager-6.4.5-2.20081217svn.fc10.x86_64


Thanks!

John W.

....oh.. and there are 3 big fat PCI-Express slots on the M3A78-T. Is Cuda the hot ticket? Is there a "Best" Cuda card for buck?

# Make me a sandwich. no make it yourself. Sudo make me a sandwich. OK #

....oh.. and there are 3 big fat PCI-Express slots on the M3A78-T. Is Cuda the hot ticket? Is there a "Best" Cuda card for buck?

It depends on the project ...

For GPU Grid the consensus seemed to be the GTX 260 as dollar per CS ... I am not sure that is entirely accurate in that a dual GTX 295 gives you twice the tasks, the same speed on each core meaning twice the throughput for the same watts (roughly).

For SaH, INDICATIONS are that the faster cards don't give an advantage... that the 9800GT is as fast as the GTX 280 ... I am not sure anyone has conclusively proved this one way or the other ... my own limited tests showed this anomaly ...

My suggestion, if you can afford it is to go the route of GTX 295 cards, adding more as you can afford it ... but you will need a PSU that can handle three cards ...

I'm no expert on this (novice, more like) and my AMD 9850BE was the best I could afford 7 or 8 months ago. The comparison in processing power is heavily in favour of the Intel and when you bring hyperthreading and DDR3 RAM into the equation I'd guess it's a bit of a non-question as this review details.

For overclockers this article explains how an extra 50% and more can be gained on an i920.

But cost is always an issue, so AMD will continue to have its fans. There's a Phenom II now which is a halfway-house both in processing and cost terms. It's a trade off, which only you'll know whether you can afford. See here.

As ever, every extra month you can hold off making a decision will see the prices become more affordable.

---

....oh.. and there are 3 big fat PCI-Express slots on the M3A78-T. Is Cuda the hot ticket? Is there a "Best" Cuda card for buck?

My suggestion, if you can afford it is to go the route of GTX 295 cards, adding more as you can afford it ... but you will need a PSU that can handle three cards ...

Even the 9600GSO requires a 400 watt power supply, beyond most store bought mahcines of just a couple of years ago. That measn that you not only need to buy a card you need to buy a power suplly too! That could also mean you will be getting into the field of a do-it-yourselfer! Replacing power a supply is NOT hard, they are sorta like chicken mcnuggets, parts is parts! Not totally, but sorta.

---

westerj The actual computing work is not done in the BOINC Manager, so whether it is optimized or maximized or midasized doesn't much matter. It runs less then 1% of the time.

As for Rosetta, you might consider increasing your runtimes gradually over time. This setting is in your Rosetta preferences here on the website. This will save the overhead of scheduler hits and file downloads, and just crunch a few more models out of each task. So instead of managing 200 tasks per day to keep the machine busy, you could drop that to 1 task per core. Again, the overhead to manage the tasks and make more scheduler requests is negligable, but it helps the overhead on the project servers too.

...my standard disclaimer, do not change the preferred runtime all at once. The entire existing cache of unfinished work will take on the new runtime, but the BOINC manager will be requesting work based on the old runtime. So, typically you want to work down the size of the work cache, then increase the runtime a couple of notches each day, then bump the cache size back up.

Just watch the estimated runtimes BOINC shows, and compare to your current preference. Once they are roughly equal, then you know BOINC is properly understanding how much work to be requesting.

---

Rosetta Moderator: Mod.Sense