Eternal Battle Day 3: Stretching Your Dollar June 28, 2005

Introduction

Page:: ( 1 / 7 )

Every good article needs a gimmick. Since we have our roots in gaming, we decided to take a cue from the classic console RPG where after the hero has finally defeated the big-bad boss, he discovers that he has only battled the “first form” of the boss, and the second, “true form” of the final boss is revealed which is even more amazing.

We’ve built two of the best systems, but they weren’t realistic projects since we had no real concept of budgets – it was just about getting the best “when it made sense.” But did we really need that 850W power supply? Couldn’t I have done just as well with slightly lower-end stuff? What if we took a step back using lower-priced, best-bang-for-the-buck components, how close would we get? When would we choose ATI or Intel over NVIDIA and AMD? What alternatives to PC Power and Cooling do I have?

In short, this article tells you about your options.


Now before you get any ideas, those super systems would truly reflect systems I would build if I wanted a PC with “no-budget, but don’t-waste-money” approach. Flagship CPUs and flagship RAM are indeed going to give you the best performance. It’s the same way that a Ford GT is going to give you better performance than a Mustang GT. But just as we all can’t afford a supercar, we all can’t afford a super computer. Let’s take a look at what we can do to reduce the price and see what’s possible with a more reasonable budget.

Desktop CPUs

We spent $1000 on our Athlon FX-57 and $1740 on our pair of Opteron 252s. As you all know, going with the flagship CPU is almost never the best bang for the buck. For some people, getting the faster CPU is a luxury. One of the nice things about CPUs is that the transitions are always more predictable. That is to say, the previous generation Athlon FX-55 is still going to be quite good in comparison to Athlon FX-57, whereas the previous generation GPU usually seems completely outclassed by current generation products. So in that way, it’s not unreasonable to invest in a high-end CPU when you’re interested in the top-end performance. For example, you can buy an Athlon FX-57 at 2.8GHz today for $1000, or get an Athlon 64 3500+ (2.2GHz Venice E3 core) for $275 today, and then wait a year for the 2.8 GHz single-core Athlon64 to drop down to $725. The question of course is 1) will you be able to find a use for that 3500+ in a second system? 2) If it took a year to wait for the Athlon FX-57 to drop in price to $650, would it have been a better idea to have invested the $1/day it took to get 365 days of extra performance?

There’s no right answer on whether it’s better to buy an FX-57 or an Athlon 3500+, but without question the 3500+ is going to give you a higher price-performance ratio. Of note, it’s critical to pick up one of the new “San Diego” or “Venice” core CPUs. In addition to minor improvements in power consumption, these two Athlon64 variants feature SSE3; the difference between the two is simply the amount of L2 cache (1MB vs 512kB). These new chips are also excellent when it comes to overclocking. If you go with AMD, I would only recommend E3 or newer steppings; it’s worth the price. http://www.amdcompare.com/us-en/desktop/

I would expect the Athlon64 X2 to represent the best bang-for-the-buck when it comes to a dual-use machine that needs to play both games and do work.

Workstation CPUs

Page:: ( 2 / 7 )

Initially, for the workstation my thought was to drop down to the Opteron 246’s since the 2GHz CPUs are a relative bargain at $333 each.

However, the Athlon64 X2 changes things up a bit. An Athlon64 X2 4200+ running dual 2.2GHz CPUs with 512 kb L2 cache is probably on par with the performance of a pair of Opteron 246’s when it comes to traditional media and digital image applications (where the L2 cache will play a smaller role). The “disadvantage” of the X2 is that it will not support things like ECC Chip Kill – but this also reflects a price advantage since conventional DDR-RAM and single-socket nForce4 Socket 939 motherboards can be used.

When would you use Intel?

With all of our systems in this article running AMD, some of you might think that we’re biased against Intel. That’s not true, I’m typing this article on a Pentium M right now. Let’s take a look at what has changed. Two years ago, in our system building article we had dual Opterons as the flagship performing systems, but we turned to the Pentium 4 3.0GHz for our storage server. Back then, we saw the Tyan Trinity motherboard running the i875P as a true reflection of Intel’s superiority over AMD when it came to infrastructure. Intel had the CSA bus for their GigE and integrated Serial ATA whereas the Via K8T800 and AMD-9000 series products weren’t nearly as robust, and NVIDIA nForce3 wasn’t readily available. In 2005, things have changed. Intel needed to turn to NVIDIA for SLI support. The platform division from NVIDIA that had its roots in the pre-Riva 128 NV2 era is now the chipset design team to beat. They’re not the best at everything, but they make the best decisions about compromises.

With the infrastructure issue moot (we’d be going with “nForce4 SLI for Intel”), the battle between Intel and AMD is now strictly at the CPU level. Intel’s Netburst architecture is starting or has already faltered depending on how pessimistic you are, whereas AMD’s K8 architecture is just only beginning to blossom. The AMD platform provides better performance at the high-end, better performance at the low-end, has integrated memory controllers, and at this snapshot in time, it’s hard not to recommend the AMD as the CPU platform of choice.

But present performance won’t predict future performance. AMD’s early success with the K7 architecture, with the first to 1 GHz win proved to be a short lived victory. When the Athlon XP platform stopped progressing and Athlon64 platform was not yet out, the clear gaming choice would have been the Pentium 4. The industry doesn’t stop – you can be sure that Intel has spent years on the drawing board, drafting up the successor to NetBurst. We’ve already gotten a taste of that with the Pentium M. Likewise AMD has its own team on the K9 architecture. A great battle between AMD and Intel lies ahead, but for 2005 it’s AMD who’s enjoying their spot as number one.

Motherboards

Page:: ( 3 / 7 )

The MSI K8N Platinum SLI is a cost effective alternative to the DFI LanParty. Although the overall performance may be marginally lower, the MSI is still a great motherboard. Moreover by omitting several of the more frivolous benefits such as the PC carrying strap, or Front X ports, the K8N saves you about $50. Likewise, there are non-overlapping talents to these two boards – the MSI has more SATA-II ports than the competition, a PCI-e based 2nd GigE port, and the integrated SB Live 24-bit.

Going to a non-SLI nForce4 might save you a few extra bucks also, but I probably would stay away from that since it limits your upgrade options. You might not be interested in SLI now, but what happens in the next generation?


For an SLI-based dual Opteron workstation, we like the Tyan Thunder K8WE if you need PCI-X. It is expensive, but that’s the admission price for a dual-socket 940 system with dual PCIe and a full complement of PCI-X slots. The other motherboard we were very tempted to use was the Supermicro H8DCE workstation board. Like the K8WE, it is one of the few Dual Opteron boards with dual PCI-express 16x slots for SLI performance. However, by forgoing the PCI-X support, Supermicro is able to enable eight SATA-II ports. The best part is that it’s available for only $360. We haven’t taken a look at this board in our labs yet, but given Supermicro’s history, we would anticipate it to be a superb performer. The big problem of course is that Supermicro does not officially acknowledge the H8DCE as a retail board. Like the original “white box” K7 boards from ASUS, Supermicro is afraid to disappoint long-time partner Intel. Politics aside, the real problem is that this makes it very difficult to find BIOS updates for the H8DCE, and technical support will be handled by the reseller rather than the manufacturer. Right now, Supermicro simply has 3 FAQ items for the H8DCE and they don’t show up unless you search for it.

The available options for a dual Opteron workstation with a single PCIe slot include ASUS’s K8N-DL for just $300. If you did not need SLI (which you’ll be able to figure out after seeing tomorrow’s articles), the ASUS is a great bargain.


SIDEBAR: Pentium M’s are even better when you by undervolting them (think of it like overclocking to improve battery life).

Memory

Page:: ( 4 / 7 )

While everyone is aware of the importance of quality power supplies, it’s still common to see people tempted to completely skimp on the memory. Don’t do it. Like power supplies, bad memory isn’t something that’s obviously noticeable at first. You’ll just have lingering reliability problems that you might simply attribute to Microsoft. Now, while it’s not necessary to always go with flagship manufacturers Corsair and OCZ, I do recommend brand-name memory be it Kingston, Crucial, or Mushkin. Generic RAM with brand-name components (Infineon, Samsung, Nanya, Micron) are acceptable, although my experience has been hit and miss – they’ll usually be great, but I will see more frequent errors. Please avoid the purely generic RAM, where the actual chips themselves as pure black without any branding at all.

All that said, the price premium for RAM from companies like Corsair or OCZ is not that high if you avoid the lowest latency products. By going with a more conservative Corsair’s Twin-X 1024-3200C2 on our desktop for example, you save over $100 in comparison to the 2-2-2-5 timing modules. These C2 modules are still quite good with better-than-most 2-3-3-6 timing and preserve the reliability you associate with Corsair modules.


For the workstation, there’s not that much leeway. JEDEC-spec RAM runs at 3-3-3-8 timing and so there’s no way you can go any slower. Dropping to PC2700 instead of PC3200 is a foolish idea, as measurements of bandwidth show nearly a 100% difference in performance. We used Corsair RAM in yesterday’s article and admittedly use the same modules in our web servers, but any server module meeting JEDEC specifications will likely be good. An important distinction to remember is that produces from a member of JEDEC does not mean that every product from meets JEDEC specifications.

Case and Power Supply

The PC Power and Cooling PSUs and even though we all agree they are the best, we can make do with cheaper PSUs. I still like to use PC Power and Cooling PSUs any time I build a system for someone for the simple fact that I’ll know that it’s a great product. Most of today’s brand-name PSUs do a great job with most systems, as opposed to the state of PSUs back in 2000. You can get great Antec, Enermax, Fortron, and OCZ PSUs for much less money. Still, the PC P&C PSUs are still the best and always overkill. The flagship PSUs from companies like OCZ, Silverstone, and 1Fortron are like a Mitsubishi Lancer Evolution VIII tuned by Vishnu Performance Systems -- amazing in performance and price. By that measure, however, the PC Power & Cooling 850SSI would be equivalent to the Audi R8 Le Mans race car. It’s still in a different class. If I was on life support and my life depended on a Windows XP computer, I would want a PC Power and Cooling PSU in there. Unfortunately, they’re so expensive that even though I know they’re the best, and everyone on the Internet can agree that they’re the best, they rarely make it into my own personal systems.

We brought in a few “PC Power & Cooling PSUs” into our labs including the OCZ PowerSteam line (manufactured by TopPower) and Silverstone Zeus (manufactured by Enhance). We had a PowerStream 600 running in a production environment. Even under significant power demand, the OCZ had no trouble handling equipment at their stock voltage and clockspeed. The Silverstone Zeus is an SLI certified with PSU with significant cross-loading precautions (not a good thing). That is to say, that while the Zeus can handle 650W total power, the available current on the +12V rails is dependent on the load on the +5/+3.3V rails. Moreover, it’s a paradoxical relationship – the HIGHER the draw on the +5V rail, the HIGHER the available current on the +12V rail. If you had nothing connected to the +5V lead, the +12V rail would be tenuous.

Power (cont’d)

Page:: ( 5 / 7 )

It’s about Clean Power not Raw power

There was nothing special or new there. Most PSU reviewers move onto peak power or try to load the PSU with a lot of devices and see how reliable it is. The team of Digit-Life will do a complex test to show cross-loading performance which is also great approach. However, recall that I said that most people are buying more watts than they need. Unless you’re running tons of SATA drives in a RAID array, a “perfect” 400W PSU should be more than enough. What I wanted to test was clean power, and going to traditional line-noise or oscilloscope test wasn’t going to tell me what I needed to know. So, I devised a devious, dastardly, insane test.

I took our Athlon FX-57, the fastest gaming CPU on the market today, the CPU that pushes the limits of AMD’s current manufacturing line and likely reflects the bin with the smallest yields. We took this flagship FX-57 CPU, and undervolted it to 1.26V, approaching Pentium M range. We then ran our 3D stress test with the CPU running at a full 2.8 GHz. The rest of the setup was “normal” with a single optical drive and a single SATA HDD, and SLI’d GeForce 6800 Ultras. By any measurement, the PC Power & Cooling 510-SLI, the Silverstone Zeus ST65F, and OCZ PowerStream 600 should have enough watts. If any PSU was going to fail our test, our guess would have been the Silverstone since it had the worst voltage stability, and has unusual cross-loading requirements meaning that in our lightly populated system with minimal +5 and +3.3V demand but exceptionally high +12V draw the Silverstone would be at the greatest risk. The OCZ is what failed.

I ran these tests several times, back to back, different times of the day, different wall sockets, etc. to really make sure I was seeing a real phenomenon. I was. This led me to get an OCZ PowerStream 520 for evaluation. 520W should be enough raw watts and because efficiency increased as the power draw approached the rated spec, there was a possibility that the 520W unit would somehow handle this system better. It did. It didn’t crash.

In all fairness to OCZ, the PowerStream 600 handled the standard, normally-volted system deftly. It was only when pushed to the extreme that I was able to draw out a software-crash. However, this is really a testament to the quality of the Silverstone Zeus ST65F and additional evidence to further support PC Power & Cooling’s reputation as manufacturers of the best PC power supplies on the planet. I may just have to rethink my plans about how much PSU is overkill. Even with a light load, the Silverstone ST65F remained a quality PSU despite the measured variations in rail voltage. It’s a tough decision between the two as SilverStone is a bit cheaper, but PC Power and Cooling has the proven track record.

The Other Perspective

No one wants an unstable system, but what if I offered you $100 cash right now, as long as you let me sneak into your room and press the reset button on your computer once at a random point in time once or twice a year, but maybe not even at all? Would you be willing to take that deal? I think it sure would be tempting. Or what if it was like you had bought a PC Power and Cooling PSU and then had it stolen, except the thief left $100 in cash and a “pretty good” replacement PSU for you?

That’s the deal on the table when it comes to choosing between high-end power supplies. Being evil helped me figure out a way to place extraordinary stress on the quality of a PSU, but would that really reflect the real-world use? Likewise, is one or two extra crashes a year that bad, when you’re already trusting your desktop to run non-ECC memory without ChipKill and only a small fraction of you have an uninterrupted power supply and an even smaller fraction of you add RAID-0? We all accept risks in computing. Where you draw the line is going to be personal, but for many of you the OCZ PowerStream 520 is going to be a better value for you.

Chassis Recommendations

A high-end chassis is about dealing with noise. You can improve cooling through brute force (adding fans) even if inefficient and so I’d rather have a $30 case with a $250 power supply than a $250 case with a $30 PSU. In our last Ultimate Workstation build from two years ago, we selected the SuperMicro 742 One of the questions worth answering is how the new Lian Li case compares to the SuperMicro 742i that we used in our previous “Ultimate Workstation.” In general, the SuperMicro cases are among the best value for true server-grade equipment. The 650W model with 7 hot-swap SATA bays (including the backplane) is available for a relatively modest $400 with a 550W being only $300. The five fan design (two in the PSU, 120mm rear exhaust, and dual 90mm HDD fans) moves air very well, although it is louder. For a full-sized server/workstation chassis, it’s hard to find something with better value than the SuperMicro stuff.

Hard Drives

There isn’t as big of a difference in pricing between different HDD manufacturers. It’s all about trade-offs. I have HDDs from Western Digital, Hitachi, Seagate, and Maxtor in active use, all of which perform well. You probably won’t save as much money by going with different branded drives.

GPU	Page:: ( 6 / 7 )

The 6800 Ultra SLI represents a poor price/performance ratio, however it clearly is a faster product than the competition. At the time of writing, the upcoming NVIDIA G70 was not available yet, but it’s obviously going to be a good decision for someone who wants maximum performance and will likely offer a better price/performance ratio than SLI’d 6800 Ultra’s.

In the sub $300 price range, the Radeon X800 XL is probably one of the best deals out there. My concern is that although it gives you the best bang-for-the-buck, it still doesn’t come anywhere close to the performance of an SLI setup. It’s not the same price, so that’s understandable, but when you think about things like a PC Power & Cooling 510 PSU compared to the OCZ PowerStream 520, or 2-2-2-5 DDR400 RAM versus 2-3-3-6 RAM, where you can get most of the performance for a fraction of the cost it’s still disappointing to know that you need to spend quite a bit of cash on your GPU if you want maximum performance. This is especially true since GPU prices were trending down just a few generations ago (the pricing cycle is heading back up).

Sadly, unlike the PSU or RAM where it can be considered an investment, a GPU goes obsolete very quickly. At least it’s cheaper than bottled water. ATI CrossFire looks interesting, but ATI’s shaky experience with their motherboard chipsets is a bit of a concern. We haven’t been able to bring in Crossfire test platforms into our own labs yet, so I’ll have to reserve judgment.


In terms of alternatives to the Quadro, it depends on your application. Going with a workstation board has never been about price/performance ratio – you go with the faster NVIDIA card because you work with datasets that require that kind of performance. The FireGL V5000 has been considered to be one of the best-bang-for-the-buck when it comes to workstation applications; however we’re concerned that the sub-pixel precision of ATI’s cards simply isn’t up to the level of the Quadro FX line. Sub-pixel precision is key when dealing with CAD/CAE environments since this determines how accurate the polygons are placed. The FireGL V5100 and V5000 is ATI’s most precise workstation card with 8-bit sub-pixel precision, but all of their other cards, even the X850 and FireGL 3100 offer a paltry 4-bit sub-pixel precision. 3DLabs and NVIDIA feature 12-bit sub-pixel precision which is the highest in the industry. All this extra precision does come at a cost to performance, but it’s pretty clear that NVIDIA’s engineers really believe in going the whole nine yards when it comes to precision especially with the 7800GTX going all-out with 64-bit texture filtering.

Between the 3DLabs and NVIDIA workstation boards, it again is a question about your needs. 3DLabs has unique features such as 36-bit precise vertex processing, and SDI support on PCI-e (NVIDIA’s SDI support is AGP only at the moment) but the Quadro will likely be faster with shader operations and clearly will also serve as a better gaming GPU when the workday is over.

Can a GeForce substitute for the Quadro?

Page:: ( 7 / 7 )

In the original Quadro design from the GeForce 256 era, the two chips were very similar and in fact, NVIDIA went on record to say that the original Quadro would outperform an equivalently spec’d GeForce 256.. Over the successive six generations, however, the GeForce and Quadro line-up continued to diverge and the chips are more different today than they used to be. Although both still share many of the same design principles (with high-precision rendering originally engineered for the Quadro trickling down into the GeForce line-up and the GeForce’s shader technology powering the Quadro FX), NVIDIA is now publicly cautious about comments about gaming on the Quadro in particular because the price premium on the product would be confusing.

For a given core (i.e. Quadro FX4400 versus GeForce 6800 Ultra), the Quadro will obviously be better but if you’re asking the question about the GeForce in a workstation environment, you’re probably asking if you can make do with a GeForce 6800 Ultra instead of a Quadro FX 1300 (GeForce 5950 class) or a Quadro FX 1000 (GeForce 5800 class).

There’s two areas where the GeForce does a great job substituting for the Quadro. Working standard textures or complex shaders. Being an NV30 class GPU, the Quadro FX1000 won’t have anywhere near the raw shader performance of the GeForce 6800. Likewise, many of the Quadro-specific features such as dual sided lighting, clipping planes, and line anti-aliasing aren’t important for conventional texture mapping. The Quadro really shines when it comes to exceptionally high polygon counts or when working in wireframe or mixed shading applications. If your application needs a Quadro, your best bang for the buck will be a used Quadro FX on eBay – the pro guys upgrade as frequently as gamers do.

It was interesting to note that the Quadro FX still excels in analog graphics quality, although this is much less of an issue given the rise of high-quality and aggressively priced DVI LCD panels.

DVD-R

Plextor does represent top-quality optical drives, however our DVD+-RW of choice is the Pioneer DVR-109. At the dirt-cheap price of around $55, you get 16x DVD-R and DVD+R burning, 8x DVD+RW, and 6x DVD-RW and dual layer burning. Although the Plextor will still occasionally eek out a few extra kb/sec when burning and will handle audio CDs and scratched CDs a little better, the Pioneer is still a remarkable drive. As long as you use quality media such as that from Taiyo Yuden, the Pioneer DVR-109 will serve you fine. All that said, if you’re really into ripping your music CDs with maximum quality and burning audiophile grade discs (I mean, truly audiophile grade where the jitter between a pressed CD and burned CD can be heard), the extra $70 may be worth it. For everyone else, the Pioneer DVR-109 will remain our budget burner of choice.

Keyboards

If you need to save money, but want a quality keyboard, the IBM model M is a good choice. Although the springs are exceptionally stiff (which you would expect to be a poor choice for ergonomics), they’re great because the positive feedback means that you won’t have as many mis-typed keystrokes. Without question, the newer ergonomic keyboards are better – but if you needed a cheap keyboard, you can find IBM Model M’s at local surplus shops or even thrift shops for $5. Sometimes if you’re lucky, you can find Omnikey Ultra’s or equivalent Alps-based keyboards at swapmeets too.

Conclusion

As you can see, there’s plenty of room to reduce the price of the systems we’ve built by being more judicious about the components selected. Tomorrow is the main event though, the benchmarks showing you just how fast our super-built systems can be. Trust me, you won't want to miss this one -- it's the best article of the series...