Jon Peddie Research - a technically oriented marketing, research, and management consulting firm in the graphics industry.

The payoff for the E5540 is in power and noise. It’s rated at 80 Watts, versus the W5590’s 130 Watts. And reduced power consumption means it will—all else being equal—run cooler and quieter. Last but not least, the E series machines are far less expensive than the W series, making them a more economical choice on two levels (initial purchase and then operation).

Bottom line: the E5540 we had was far more representative of what most customers would buy, but the W5590 would be the choice for those who are willing to give up power consumption advantages and pay more for some varying degree of performance
improvement.

Different motherboard topology with Nehalem

Migrating to Intel’s Nehalem architecture means a different motherboard layout than FSB-based CPUs have. Therefore, the memory DIMMs need to be closer to each of the CPUs (see photo), rather than the single North Bridge.

Exploring ways to differentiate

Workstation OEMs today don’t have a whole lot of room to differentiate on components. They’ve all got access to the same components from Intel, Nvidia, and AMD, as well as disk and storage components. It’s not easy for one machine at one price point to argue performance advantages over another similarly priced system based on the same hardware components. So to a large degree, differentiation must come from other avenues, such as:

• Price.
• Aesthetics.
• Ergonomics.
• Power consumption.
• Noise level.
• Expandability.
• Reliability.
• Bundles and turnkey solutions.
• Service and support.

Aesthetics? In the past that was either a low priority or a non-priority, aesthetics is carrying more and more weight, varying by segment. The trend lately has been leaning toward the beefy, industrial look, showing grill work and—cost allowing—metal rather than plastic (for example, brushed aluminum enclosures). And not just on the outside either, as a look inside Apple’s Mac Pro attests.

Rather than polished bare metal all around, D20 designers went with a black finish, combining a metal grill and enclosure with plastic trim. One component of the trim is a beefy front handle, which, when paired to a metal tab on the backside, makes lugging the system less of a hassle.

Touting a tool-less chassis in and of itself isn’t much of a differentiator anymore; tool-less is a checkmark item these days. However, vendors’ approaches to tool-less aren’t the same. In the D20’s case, the panel removes easily, revealing several tool-less devices: a plastic retention bracket that supports add-in cards when the panel is closed, pull tabs on drive trays and press-to-release tabs for optical. Everything snaps closed and releases fairly reliably.

Benchmarking the Lenovo D20 ... and an opportunity to compare Nvidia’s Quadro FX 4800 and AMD’s FirePro V8750

To get a feel for the D20’s general system-level capability running typical workstation loads, we ran three SPECapc benchmarks: 3ds max, SolidWorks and Lightwave. And to focus in on the graphics, we opted for the tried-and-true SPEC Viewperf.

Comparisons are always difficult from a number of perspectives. First off, the systems or AIBs being compared need to represent a reasonable apples-to-apples comparison. For our definition, “reasonable” means a similar configuration and/or similar price point and similar target usage. That is, if they’re built up similarly, or if they sell for around the same amount (street price) and they’re intended for the same class of user, then they’re
comparable.

Second—and we stress this often—benchmarks always need to be taken with a grain of salt. Used improperly, benchmark results can be misleading. And even under the best circumstances, benchmarks only tell one part of the story.

Lining up a Quadro FX 4800 and AMD FirePro V8750

With the D20 review, we tested an Nvidia Quadro FX 4800 (provided in-system) and an AMD FirePro V8750. The two are based on comparable GPU generations (the Quadro FX 4800 derived from the GT200 and the FirePro V8750 from the RV870), and the two sell in a similar range (as of this writing) on the street. Results from an on-line price scan showed both cards selling in a range of around $1,200 to $1,400, with the Quadro FX 4800 perhaps just a nudge less expensive. One notable difference is that the FirePro V8750 comes outfitted with 2 GB of GDDR5 memory, where the Quadro FX 4800 is equipped with 1.5 GB of GDDR3.

No apples and apples

We couldn’t compare apples-to-apples by configuration, but we could by price ... so what’s better to buy: more modest-speed cores or fewer fastest-possible cores?

We didn’t have any dual-socket E5540 Xeon systems to compare to the D20, however, we could compare by price. The 2.53 GHz E5540 has a more power efficient CPU and it’s far cheaper. So much so that two 2.53 GHz E5540’s can be had for about the same price as one 3.33 GHz W5590 (Dell’s site had one W5590 only $200 cheaper than two E5540’s when configuring a Precision T7500).

Now all else being equal, two 2.53 GHz processors will offer more theoretical performance than one 3.33 GHz processor. But in the reality of the multi-core age, how well they compare depends on how much the application can take advantage of more cores. Specifically, with today’s quad-core Nehalem-EP processors, the question is how much better can eight physical cores (16 logical with HyperThreading) perform compared to four (or eight, respectively). Running SPECapc on our dual E5540 D20 and comparing to benchmark results from SPEC-submitted single W5590 systems yielded some interesting results.

For the SPECapc 3ds max 9 benchmark, we compared the D20 to both a Dell Precision T7500 and an HP 7800. The W5580 and W5590 systems outperformed our dual E5540 machine on the Hardware shader and Graphics components. So the superior clock rate of the single W5580 and W5590 trumped the dual E5540’s. But for the CPU render component the more plentiful (albeit slower) Nehalem CPU cores performed better.

For SPECapc Lightwave, all components scaled poorly with the number of cores, and again the single W5590-based Dell Precision T7500 and HP Z800 bested our dual E5540-based D20 on all counts. On the Render and Interactive scores, the Dell and HP systems beat the D20 by 23% to 30%, a range that’s no surprise, considering their processors are 31% faster. But on the Multi-task score, which we’d expect to scale better by core count, the D20 was much closer.

We ran SPECapc for SolidWorks, comparing the D20 against the Dell Precision T7500 and a Fujitsu Celsius 670. The benchmark again rewarded the faster, fewer-cored competitors, with the exception of I/O, which as one might surmise is not particularly CPU-sensitive.

Somewhat counter-intuitively, the D20 performed worse on the graphics component, despite having the pricier Quadro FX 4800. But remember, this is SPECapc, where the CPU is still busy with application processing, even if the component name suggests it’s testing just the graphics. And when considering the GPU on-board, the FX 3800 is essentially equal that of the FX 4800, and the result isn’t completely out of whack (much of the premium for the FX 4800 is for memory size).

How well the application scales by core count makes all the difference ... in this case, the SPECapc applications preferred the one faster CPU over the two slower ones.

Specifically in the context of this subset of these SPECapc benchmarks then, we can conclude that allocating dollars upgrading to a faster single CPU pays off more than forking out dollars on a second slower CPU. Of course, that conclusion doesn’t reflect on the performance of D20 compared to the Z800, Precision T7500 or Fujitsu Celsius 670. Rather, it reflects on how well—or in this case, not so much —the application tasks that SPECapc dispatches take advantage of more available cores.

SPECapc results vary by graphics card

As expected, the D20 system did yield modestly different SPECapc results, depending on which graphics card was installed. But while we expected the scores to vary, we were surprised that the card had relatively little impact on scores. And for 3ds max and Lightwave, the D20-with-FirePro V8750 modestly edged out the D20-with-Quadro FX 4800 on 3ds max, while the reverse situation held for Lightwave, where the Nvidia-enabled system finished slightly ahead.

It’s worth noting that for the Quadro FX 4800, we ran with the company’s Graphics Performance Driver, not the specific 3ds max Performance Driver. The latter was not available for Windows 7 at the time (only beta), but when available should provide some (we assume noticeable but modest) performance improvement on the 3ds max test.

However, on SPECapc for SolidWorks the difference was significantly (and surprisingly) more substantial, with the Quadro FX 4800-configured D20 beating the FirePro V8750-configured D20 by between 10% and 20%. The reason for it wasn’t clear, as even the I/O component was superior, and the edge was most pronounced on the CPU sub-score; in neither case would we assume the graphics card should have much impact.

Viewperf 10.0 results

With the exception of the SolidWorks test, SPECapc results didn’t change dramatically depending on the graphics card installed. And where the differences were modest (3ds max and Lightwave), each vendor scored one win. However, next up for testing was Viewperf, which effectively isolates the load on the graphics subsystem (driver and card). Would we see big differences by card? Would one card perform better on some tests and worse on others, or would one card sweep the testing? As it turned out, while both cards did win some corners, one card took top honors more consistently.

We ran Viewperf 10.0 three times per card and across three corners: single-thread, dual-thread and quad-thread. If any one result was significantly out of line, we would have discarded that result; however, that didn’t turn out to be necessary. The three results per corner were then averaged to yield the final tallies.

The numbers were generally close, with the edge in most—but not all—instances going to the Quadro FX 4800. In general, as the number of threads increased, the FirePro card narrowed the gap to the Quadro card. But that was not a rule without exceptions. As the number of threads increase, the stress on the cards’ transformation and lighting capabilities should increase linearly, while the stress on fill and texture rate should change less (as the windows are smaller). That is, the load balance shifts, which opens the door for different architectures (not to mention different driver optimizations) to perform differently.

In the single-thread corner, the Quadro FX 4800 won out modestly, but the FirePro V8750 came out on top with the Maya (maya-02) viewset. As the number of threads increased to two, the edge got more pronounced, but then moving on to four, the Quadro turned the tables, finishing higher.

Where’s the money

So, in testing the two cards, we’re seeing slightly different slants on where dollars are spent.

The Viewperf results were in-line with our general expectations. We didn’t expect either card to post results to shame the other, we didn’t expect one card to sweep all the tests, and we didn’t expect the difference to be consistent from viewset to viewset. And we weren’t surprised to see the Quadro FX 4800 edge out the FirePro V8750 more of the time. That’s no indictment of the latter card, but rather an indication of the subtly different approach the card vendors had.

It’s not a case of the two cards being apples and oranges. They’re both apples, but they’re apples of somewhat different shapes. Nvidia allocated a bigger percentage of its build cost to the GPU and AMD a bit more to memory.

The heart of the Quadro FX 4800 is a full-blown GT200, where FirePro V8750 sports an RV770. Both (as of this writing) represent the current top of the line GPU for each company in the context of their professional line (AMD’s recently introduced Evergreen, and Nvidia will in coming months launch Fermi).

In terms of raw horsepower, the GT200 has held an edge over the RV770, modestly but consistently, as echoed by Nvidia’s GTX 200 series holding a performance edge over AMD’s Radeon HD 4800 series. Nor should the GT200’s edge over the RV770 surprise anyone, as the two companies had different-enough goals in mind when shaping the devices; AMD was looking for superior price/performance, while Nvidia aimed for the single-GPU performance crown.

So both are solid performers with similar street prices. But if you’ve got $1,200 or so to spend and want maximum GPU horsepower, you’d probably lean toward the Quadro FX 4800. Conversely, if you’ve got an application sensitive to video memory size (big, deep quad-buffers perhaps?) and/or memory speed, you might lean to AMD’s 33% larger frame buffer with GDDR5 speeds.