Slashdot Mirror

← Back to Stories (view on slashdot.org)

Building A High End Quadro FX Workstation

Posted by Hemos on from the making-it-faster-and-better dept.

An anonymous reader writes "FiringSquad has an article detailing some of the differences between building a high-end workstation and a high-end gaming system. They go into things like ECC memory, and the difference between professional and gaming 3D cards. The Quadro FX 2000 coverage is particularly interesting -- the system with the Quadro FX 2000 was never louder than 55 dB!"

29 of 89 comments (clear)

  1. interesting by Sh0t · · Score: 2, Interesting

    Compared to the 75 db GFFX, that's a whisper

  2. ECC Memory? by Proc6 · · Score: 3, Insightful
    Can someone tell me why ECC memory is a good idea? I don't think I can remember in all my years of computing a machine crashing due to a memory error, or a machine not crashing because ECC memory saved it. Maybe I wouldnt know if it did, but, I've always felt like ECC memory was slow, more expensive, and necessary about like UFO insurance. Personally Id rather have regular memory, that taco's the machine completely when there's a problem, so I know there's a problem, than I would ECC constantly correcting memory errors without my knowing, untill I go to leave on vacation, then the whole DIMM gives out.

    I Am Not A Memory Expert though.

    --

    I'm Rick James with mod points biatch!

    1. Re:ECC Memory? by evilviper · · Score: 2, Interesting
      I don't think I can remember in all my years of computing a machine crashing due to a memory error

      Either you just haven't recognized when it happened, you don't work with any significant number of computers, or you've been INCREDIBLY lucky.

      Memory isn't perfect. If your uptime is important, you need ECC.
      --
      Slashdot gets worse every day... Pipedot: News for nerds, without the corporate slant
    2. Re:ECC Memory? by e8johan · · Score: 5, Informative

      RTFA Read The F**king Article!

      "Two to twelve times each year, a bit in memory gets inappropriately flipped. This can be caused by cosmic rays flying through your RAM or a decay of the minute radioactive isotopes found in your RAM - the impurity need only be a single atom. Most of the time, this flipped bit is unimportant. Maybe it's a flipped bit in unallocated memory, or maybe it just altered the position of a pixel for a fraction of a second. If you're unlucky though, this flipped bit can alter critical data and cause your system to crash. In our situation, a flipped bit could potentially alter our results significantly."

      Quoted from the second paragraph of the fourth page.

    3. Re:ECC Memory? by larien · · Score: 4, Informative
      OK, two points:
      1. If you're aiming for stability, you try to remove all such possible causes; even if Windows will crash once a week, there's no point making it worse by risking memory failure.
      2. Even if your machine doesn't crash, a flipped memory bit could invalidate your data results by altering a crucial figure. In some cases, it's not important, but a flipped bit at the higher range could alter a conclusion significantly and you wouldn't notice.
      Depending on your target audience, the latter may be more important than the former.
    4. Re:ECC Memory? by rtaylor · · Score: 2, Insightful

      It's subtle corruptions most people worry about. If you're doing financial transactions, you do everything you can to ensure that 4 doesn't turn into an 8 accidentally.

      --
      Rod Taylor
    5. Re:ECC Memory? by e8johan · · Score: 4, Insightful

      Large simulations (such as this, or car crash simulations, etc.) take days, if not weeks to run. Since the ECC ram isn't 100% slower (i.e. time of fast memory times two is more than time of ECC memory) there is no need to run it twice.

      Anyhow, if the two simulations differ, you'll have to do it a third time to check if you get a match, and still you only know that you are *likely* to have gotten it right. With ECC the chance of getting it right increases.

    6. Re:ECC Memory? by kperrier · · Score: 2, Insightful

      Wouldn't running the tests twice be a better way to ensure this kind of thing doesn't happen?

      So, if it takes 4-6 DAYS for a test to run, you want to run it again to verify the results? They don't have the time to do it again. Take this from some one who manages a 190 node Linux cluster. We use this for seismic data processing. Our processing run times are 3 to 4 days, each and there are multiple runs for each job. We have project schedules that we need to meet, and running each step in the processing schedule twice is not an option.

      Depending on what you are doing, the money is better spent on the front end for quality hardware, than to double the time for a project to process the data. You could double the initial cost of the hardware, have two clusters and run the tests in parallel and compair the results. This may be the best thing to do, depending on what you are modeling/processing, but its much cheaper to invest in the quality hardware up front.

      Kent

    7. Re:ECC Memory? by still_nfi · · Score: 2, Interesting

      Um....since a large number of memory accesses come from cache, wouldn't it be more important to have an ECC cache than main memory? Certainly, that is where it is most likely that a flipped bit is going to cause a problem. I have doubts that any of the processors use ECC code in the L1 or L2 caches?

      Also, it's been a while but don't most non-ECC memory use parity bits? So that a single flipped bit will be noticed...hence the isolated blue screens of death/ kernel panic on very rare occasions. Or is a parity bit what passes for ECC these days?

      --
      "I have been around the world and found that only stupid people are breeding" -- Harvey Danger
  3. Not entirely surprising by kruetz · · Score: 5, Informative

    Let's face it - the main focus in a games PC is a blindingly fast GPU that can do umpteen hundred frames/sec at 1600x1200x32 or whatever, so you also need your system to be able to give the data to your video card as fast as possible. (Sound is another consideration, but not quite so major).

    But "honest-to-goodness computation" (numerical analysis, ...) doesn't use a GPU too intensively, except for displaying graphical data, for which the high-end OpenGL cards are ideal. The main focus here is CPU's performance in doing complex numerical tasks, not just passing data to the AGP slot. And let's face it, multiple-CPU PCs don't necessarily do anything for gaming, but they're great for this sort of stuff.

    However, most if not all of the points in this article are quite informative - did YOU know the difference between Athlon XP and MP. I thought I mostly did.

    And his choice of ECC RAM - Two to twelve times each year, a bit in memory gets inappropriately flipped ... If you're unlucky though, this flipped bit can alter critical data and cause your system to crash. In our situation, a flipped bit could potentially alter our results significantly. Geez.

    We come to the video card - a hacked GeForce isn't the same thing as a Quadro - bet some of the FPS freaks might be a little surprised, but the GeForces and Radeons aren't made for this sort of stuff. No real surprise, if you think about. But, as he says, why not a FireGL? Everything comes back to the lesson of the day: know your task. And boy, he certainly does.

    Anyway, enough of regurgitating some of the finer points of this great article. Read it for yourself. And don't post comments about how 1337 your Radeon 9700 Pro or Ti4800 is. Know your task.

    --

    This sig intentionally left bla... dammit!
    Who's got the whiteout?
  4. SGI should be put out of its misery by Proc6 · · Score: 4, Funny
    Not intending to start a Holy War, I realize the 64 CPU monsters have their place but their workstations are just ignorant (this is coming from a previous SGI only owner)...

    "These systems were around $40,000 when first released. Each R12000 400MHz has a SpecFP2000 of around 350-360 and so it's approximately equal to an Athlon 1.2GHz. The caveat is that the SpecFP2000 benchmark is actually made up of a bunch of other, smaller, tests. For computational fluid dynamics or neural network image recognition, the 400MHz SGI CPU is 2.5 to 5 times faster than the Athlon!"

    WOW! 2.5 times faster than a 1.2Ghz Athlon!? Man, you'd almost need a $168 2.4 Ghz Athlon to keep up! I wish they made them!

    P.S. The 3.06 Ghz P4 is just under 1000 on the SpecFP benchmark.

    --

    I'm Rick James with mod points biatch!

    1. Re:SGI should be put out of its misery by lweinmunson · · Score: 2, Interesting

      Not intending to start a Holy War, I realize the 64 CPU monsters have their place but their workstations are just ignorant (this is coming from a previous SGI only owner)...
      "These systems were around $40,000 when first released. Each R12000 400MHz has a SpecFP2000 of around 350-360 and so it's approximately equal to an Athlon 1.2GHz. The caveat is that the SpecFP2000 benchmark is actually made up of a bunch of other, smaller, tests. For computational fluid dynamics or neural network image recognition, the 400MHz SGI CPU is 2.5 to 5 times faster than the Athlon!"

      WOW! 2.5 times faster than a 1.2Ghz Athlon!? Man, you'd almost need a $168 2.4 Ghz Athlon [pricewatch.com] to keep up! I wish they made them!

      P.S. The 3.06 Ghz P4 is just under 1000 on the SpecFP benchmark [specbench.org].

      Lets see, the last generation that we have SPEC numbers on for SGI is the 600Mhz R14K. It clocks in at 529 PeakFP compared to 656 Peak FP for the 2.4 GHz MP that was used in the benchmark. That's about a 20% difference in Speed. The original CPU's that he was dealing with, the R12K 400 and the 1.2 K7 are 407 and 352 respectively. That actually gives the SGI a lead by about 15%. Now if the 2.5x increase in an application holds true, I'd say the SGI is still a good deal if you can afford it. Now granted I don't have $40,000 to spend on a workstation, but there are plenty of companies who are willing to spend the extra $30,000 once to get double the performance out of their $60,000 a year engineers for the next two or three years. Also, as is pointed out in the article, the P4 is insanely optimized for SPEC. It's numbers have no real meaning to most realworld applications. If you want to get right down to it, SGI can give you 512 CPU's run through a single Infinite Reality module. No one would actually do this, but it's nice to dream about it once in a while :)

  5. Biased? by Gheesh · · Score: 4, Interesting

    The article carefully explains the choices made. However, we find the following line at the end of it:

    Special thanks to AMD, NVIDIA, TYAN, and Ryan Ku at Rage3D.com for helping me with this project.

    Well, maybe they had no influence at all, but then how come that most of the chosen products match this 'special thanks' line?

    1. Re:Biased? by sweede · · Score: 3, Insightful

      perhaps the author of the article did research and picked out the componants of the system BEFORE contacting vendors and buying them.

      you dont order food or car parts without knowing what is there and what you want/need do you??

      Oh, and if you also notice that the rest of the site is based on new hardware reviews and performance, you'd think that they would have good experiances with what works and what doesn't.

      If you went out and researched companies or people for a project you where doing, would you not include them in a `special thanks to' section of the paper?

      --
      I follow the SDK and GDN principles.. Spelling Dont Kount, Grammer Dont Neither
    2. Re:Biased? by vivIsel · · Score: 3, Insightful

      Welcome to the world of "hardware review" sites. Bias is their collective middle name.

  6. Re:Easy by Molt · · Score: 2, Interesting

    You may like to read the article. This is a scientific visualization workstation being built with a seriously nice Quatro FX graphics card.


    The author even benchmarks UT2k3 on it, and the scores are.. umm.. impressive.

    --
    404 Not Found: No such file or resource as '.sig'
  7. ISV Certification by vasqzr · · Score: 5, Informative


    If it's not ISV certified it doesn't do you much good, as for as a workstation goes.

    From Ace's Hardware:

    When you look at the typical price ($4000-$6000) of a workstation built by one the big OEM's you might ask yourself why you or anyone would pay such a premium for a workstation.

    In fact if you take a sneak peek at the benchmarks further you will see that a high-end PC, based upon a 1400MHz Athlon, can beat these expensive beasts in several very popular workstation applications like AutoCAD (2D), Microstation.

    Yes, it is possible that you are better served by a high-end PC, assembled by a good local reseller. Still, there are good reasons to consider an OEM workstation.

    Most of the time, a workstation is purchased for one particular task, and sometimes to run one particular application. Compaq, Dell and Fujitsu Siemens have special partnerships with the ISV's (Independent Software Vendor) who develop the most important workstation applications. In close co-operation with these ISV's, they verify if the workstation is capable of running each application stablely and fast. In other words, you can ask the OEM whether or he and the ISV can guarantee that your favorite application runs perfectly on the OEM's workstation. ISV certification is indeed one of the most critical factors that distinguishes a workstation from a high-end desktop.

    Secondly, it is harder to assemble a good workstation than a high-end PC. Typically, a PC is built for the highest price/performance. A lot of hardware with an excellent price/performance ratio comes with drivers which do not adhere strictly to certain standards such as the PCI and AGP standards. Even if this kind of hardware might comprise stability in very rare cases, it is unacceptable for a workstation.

    Last but not least, workstations come with high-end SCSI harddisks and OpenGL videocards which are seldom found in high-end PC's. Workstations are shipped with ECC (Error Checking and Correction code) memory and can contain 2GB to 4GB memory. High-end PC's typically ship with non-ECC memory and are - in practice - limited to 512MB (i815 chipset) - 2GB (AMD760).

  8. Re:Easy by sql*kitten · · Score: 3, Insightful

    1. workstation == better processors
    2. gaming system == better graphic cards

    Not as simple as that. A games card will trade precision for speed, because precision is less important if you are updating the scene dozens of times a second anyway. If two walls don't meet perfectly for 1/60th of a second, who will even notice? A workstation card will trade speed for precision - you cannot risk a mechanical engineer missing an improperly aligned assembly because of an artifact created by the graphics card, or worse, breaking an existing design because an artifact shows a problem that doesn't exist in the underlying model.

  9. What Changes for a Linux Math Machine? by Confessed+Geek · · Score: 2, Interesting

    Quite a Nice article, and useful to me since I'm consistantly building workstations for use in physics research, but what changes would be made for a linux based system?

    The information on GPU's was great, if your running in windows and doing visualizations, but most of science doesn't use Windows. They started their projects on Big Iron Unix and are now moving to linux.

    Our current spec out looks like this:
    2 Athlon MP 2400
    Tyan Tiger MPX
    We were using Thunder, but found we didn't need the onboard SCSI so moved to tiger. After the fits I've been having w/ Gigabit cards and the AMD MP chipset though I'm considering going back to the Thunder for built in gigabit.
    2Gig Kingston ValueRam EEC RAM (its what tyan suggests)
    120GB WD Spc. Ed. 8M cache HD
    Additional Promise IDE controllers for new HD's when needed.
    Generic TNT2 or Gforce2 Video. (they are just math boxes)
    Plextor ide CDRW
    Still looking for the prefect tower.
    Extra case fans.

    The CPU's have been changing over the last year or so as the MP's get faster, And we have moved from 1 to 2G of ram.

    Biggest problem I'm still having is the system sounds like a 747 taking off and I've had official AMD CPU fans burn out on me. I would still love to get a bit more oomph out of this though if there are any suggestions.

    1. Re:What Changes for a Linux Math Machine? by Glock27 · · Score: 2, Informative
      (they are just math boxes)

      If they had higher-end NVIDIA graphics cards, they could also be very good OpenGL development/visualization stations, using Linux. Port all that SGI code with very little effort...

      Biggest problem I'm still having is the system sounds like a 747 taking off and I've had official AMD CPU fans burn out on me. I would still love to get a bit more oomph out of this though if there are any suggestions.

      I'd use aftermarket fans, I thought AMD's fans were cheesy (to use a technical term;). If you want a good product, I recommend the PC Power and Cooling Athlon CPU cooler. PCP&C generally has top-quality products (great choice for power supplies as well).

      You should probably start going for DVD/RAM drives also, lots more capacity for backups...

      One final thought on numerics - you might want to compare some of the commercial compilers with gcc. For instance, Microway resells a strong line of commercial compilers. The Portland Group compilers, in particular, look promising.

      --
      Galileo: "The Earth revolves around the Sun!"
      Score: -1 100% Flamebait
    2. Re:What Changes for a Linux Math Machine? by Brian+Stretch · · Score: 2, Informative

      Replace the AMD heatsink/fan kits with Thermalright SLK800's, YS Tech 80mm adjustable fans, and use Arctic Silver 3 thermal compound. The catch is that the pink crap AMD uses instead of proper thermal compound may be permanently attached at this point, though the right chemicals (Goof-Off cleaner followed up with rubbing alcohol) can probably remove it. I'm using SLK800's on my dual 2400+ ASUS A7M266-D board and with the fans adjusted to 2000RPM the system is very quiet, the most annoying noise is from the fan on the Ti4200 card and there's no room for one of those neato Zalman heatpipe GPU coolers. With this setup I'm getting lower CPU temps than I was with 1800+ chips and the retail box heatsink/fan kits (using AS3, scraped off the pink stuff).

      See 2CoolTek for this gear. I've been buying from them for years and highly recommend them.

      You could go with one of those Vantec fan speed adjusters (handles 4 fans) instead of variable-speed fans... might be a better choice in your case.

      Perfect tower: one of the Lian-Li aluminum cases, probably an extended length model (extra 10cm of space). See NewEgg, etc. Actually, they've got the cooling gear too.

  10. Re:Easy by clarkc3 · · Score: 2, Informative
    2. gaming system == better graphic cards

    I just cant agree with that statement - its more a 'drivers written to function better in games' than a better graphics card. The one in the article uses a Quadro FX and I know lots of other people who use a 3dlabs Wildcat series - both of those cards wipe the floor with 'gaming' cards in 3d rendering for things like cad/3d studio/maya

  11. "two sticks of RAM instead of one for Redundancy" by Zak3056 · · Score: 2, Insightful

    Did anyone else see a logical disconnect between his assertation that two sticks of RAM were better than one because if one failed, the machine could still operate while they waited for a replacement stick... and yet he chose NOT to use RAID?

    Even worse, his choice of drive was a single WD 80GB IDE drive? WTF? There's a reason the warranties on those things just dropped to a year!

    --
    What part of "shall not be infringed" is so hard to understand?
  12. high end workstation? by asv108 · · Score: 2, Interesting
    Hard Drive:
    Western Digital 80GB Caviar with 8MB Cache

    Why would you use a single IDE HD when you have SCSI built in the motherboard? In my experience storage upgrades always provided tremendous speed improvements. Disk access is always a big bottleneck. If your going to have a "high-end" workstation, you need at least SCSI, preferably SCSI RAID. If you want to go barebones, at least have IDE-RAID with a really good backup plan.

    And WTF do Quake 3 benchmarks have to do with a workstation?

    1. Re:high end workstation? by Billly+Gates · · Score: 2, Insightful
      He mentioned in the article about his budget. Have you looked at the price of scsi drives? Tiny 20-gigs $400 each! Ouch.

      Scsi is not faster or more reliable then ide unless its in Raid. So if your going to do scsi then you might as well buy not 1 but 4 drives for raid. That adds up. If your doing alot of i/o requests in parrallel then scsi is faster because it can offload the tasks and que them from the controller. A single app will not do this unless its a database or other server oriented application. I notice a bigger increase in performance from a faster processor but this is because I do not run a server. A workstation with lots of ram has the bottleneck in memory, cpu, and graphics card. A server on the other hand is different.

      More emphasis should be on the processor and video card for any workstation purchase.

      I agree with IDE-RAID if the job can not be interrupted because of a failed drive but 4 drives are expensive but still alot cheaper then scsi. Also worth mentioning is using bigger storage capacities with the ide from the amount of money saved. Keeping critical jobs is not as important as it use to be because engineers like their other white collar associates never store the finished jobs on their own drives. They rather use a network share when they are done. You would be a fool to store your work on your own drive since the file server backs it up on tape. Workstations typically run win2k today rather then unix so this means they can use NT and Novell file servers.

      --
      http://saveie6.com/
    2. Re:high end workstation? by Billly+Gates · · Score: 2, Informative
      But 320/mb is the theoritical limit. Last time I looked a typical hard drive transfers only 30/MBsec. This was over a year ago so it may be higher now. ATA can easily handle the fastest hard drives.

      A raid with 4 drives might be more usefull 4*30 = 120/MBsec which begans approaches the ATA limit in Eide. Newer drives comming out will probably hit the ata limit soon in raid and only scsi can keep up. For a single drive scsi is not worth it.

      Its strength will not show unless you run very heavy i/o bound applications. I agree that SCSI is supperior. I can't picture an engineer swaping out his hard drive while rendering a scene so swapping support is important only in the server arena.

      Your post just repeated mine in saying the emphasis on a workstation is not i/o bound and scsi is not worth it unless its in raid. Price is important in this day and age of shrinking IT budgets the scsi myth is being exposed. A sigle scsi drive is not that much faster or reliable then an IDE.

      --
      http://saveie6.com/
  13. Re:"two sticks of RAM instead of one for Redundanc by Billly+Gates · · Score: 2, Interesting
    Actually the warranty dropped from their scsi units as well. Something tells me a defect might be in there. Especially with larger capacity drives.

    Also many scsi drives are less reliable then ides. Hu?? This is because scsi drives typically spin at higher revolutions so they tend to fail more. Higher capacity drives are more prone to defectives and data corruption. The lower capacities typically are more reliable. Ask any admin how often they replace scsi drives on various raids? The fastest and biggest ones from what I read here on slashdot fail every 2-6 months! Quantums I heard fail on a weekly basis on some of the more questionable units. The newer ones seem to be the worse.

    I have been doing computers since 1991 and I have never seen a hard drive fail. I only use ide. I believe part of the reason is I use to upgrade my drives every 2 years and until recently did not run my systems 24x7 like servers do. For the last 2 years I have been running 24x7 without any problems. Like you I would still select scsi assuming its for critical level work and money isn't an issue. I would pick Ide if raid was not needed since scsi is not more reliable unless its in a raid-5 configuration. Most workstations use alot of graphics and cpu power. Server applications tend to bottleneck at the hard drive. So hard disk performance is not really a factor unless the application runs of memory and swaps to the drive. Scsi vs Ide benchmarks show that they are almost identical in speed unless lots of i/o requests go to the drive in parrellel. Most cad apps today easily stay within the 2 gigs of ram. I know exceptions exist but they are rare.

    However I would try to stay within 7200 rpm and not go above 10,000 for the drive. Your asking for trouble with the higher speeds not to mention do not really provide an increase in performance more then single percentage points in alot of benchmarks. Another benefit also with going with slower rpm drives is that they are alot more quiet.

    Scsi is nice because it offloads alot of i/o processing to the scsi card. For any database or crtical application where raid is needed its the only way. For a graphical workstation for non critical use (artist or grunt level engineer) price and huge storage might be a bigger factor as well as reliability. Scsi without raid is not more reliable. I know a few raid workstations exist but raid is almost exclusively used in servers and is expensive for a desktop. Most engineers save their work on a network share. I guess you have to take in the cost of a hard drive failure. Yes engineers are sometimes expensive but not more then any guy in sales or marketing in a big corporation. You might as well give everyone raid.

    --
    http://saveie6.com/
  14. Re:MP3 Playback? by mccalli · · Score: 2, Interesting
    Look at drivers put out by Creative Labs... they're reputed to be really buggy in a dual-processor setup.

    Confirmed in my my experiences with an AWE64 and a dual 533Mhz Celeron setup. I moved to a Santa Cruz Turtle Beach - no problems.

    And as far as "Is there something magical about MP3s?," I think he's talking about standard wave output support...

    Many card/driver combinations are supposed to be able to recognise the kind of data put through them. The Santa Cruz, for example, had a 'Hardware MP3 accelerator' option in the control panel. I really don't know how they recognise it though - by instinct I'd agree that surely the waveform has been decoded by the main CPU anyway? Be interested to hear from anyone who knows more about this point.

    Cheers,
    Ian

  15. Well written, but weak article by zaqattack911 · · Score: 3, Interesting

    I think he starts off well talking about the decision making process, the move over x86, what ECC means.

    However, he pretty much dumps his chosen hardware in our laps by the end of the article without much explanation. It feels rushed almost.

    There is way more out there than Tyan, who cares what google uses. What about dual channel DDR? What about the fact that Xeons and newer P4s have HyperThreading?

    He starts slow, then in a few paragraphs blurts out some mystery hardware he decided to go with. Then babbled about Geforce VS Quadro for the rest of the article.

    Oh well, he's a good writer. Better luck next time.