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

Taking a new workstation and two professional graphics cards out for a test-drive

Posted by Alex Herrera on June 2nd 2008 | Comments Closed
Categories:
Tags:

Lenovo’s new ThinkStation S10 workstation.
(Source: Jon Peddie Research)

Table 1: ThinkStation D10 specification.
(Source: Jon Peddie Research)

A look: inside the S10’s tool-less chassis.
(Source: Jon Peddie Research)

The S10’s: front panel.
(Source: Jon Peddie Research)

The S10’s: rear panel.
(Source: Jon Peddie Research)

Table 2: Evaluation system configuration.
(Source: Jon Peddie Research)

Figure 1: Cinebench benchmark results.
(Source: Jon Peddie Research)

Figure 2: Viewperf 10.0 results for FireGL V7600 and Quadro FX 4600
(Source: Jon Peddie Research)

Cadalyst ‘08: benchmark results.
(Source: Jon Peddie Research)

Lifting the curtain on its ThinkStation S10 and D10 models back in November of ‘07, Lenovo became the first major vendor to enter the market for deskside workstations in many years. After backing into the market for mobile workstations—with its acquisition of IBM’s PC division in 2005—Lenovo has made a more conscious step forward with deskside models, mounting a new challenge to market leaders Dell and HP.

Lenovo’s ThinkStation S10 taps Intel’s Core 2 while its big brother ThinkStation D10 leverages Intel’s Xeon for dual-socket configurations. Though both of Intel’s new 45 nm processors were announced last fall, Lenovo did not manage to ramp volume until later in Q1’08. When they did finally ramp up, the company loaned an S10 to us for evaluation.

Covering the entry-level range of the workstation market, the S10 starts out at around $1,200 with specs and options as shown in Table 1.

X38 Express offers several improvements over Intel’s previous generation 975X chipset, most notably two full x16 PCI Express interfaces to support dual graphics cards (but not SLIs) The S10 takes advantage of Intel’s most advanced chipset for Core 2. The front-side bus runs at 1066 MHz, rather than the top-end speed of 1333 MHz, constrained not by the X38 Express but the 65 nm Q6600 (more on Lenovo’s processor choices ahead). And memory on the X38 is both faster and bigger, with the S10 supporting up to 8 GB (four 1067 MHz DDR3 DIMMs).

But perhaps most notably, for the first time with the X38 Express we have two true dual x16 PCI Express electrical interfaces for graphics. Accordingly, the S10 comes with two x16 slots, allowing two high-end cards to drive four high-resolution screens at once.

When selecting graphics options to launch its new deskside line, Lenovo chose to exclusively partner with Nvidia. That wasn’t a complete surprise, as Lenovo appears to emulate the old IBM mode (at least as it gets the workstation ball rolling), and IBM was one of the few major OEMs to go to a single graphics source.

The (tool-less) chassis

Why did it take so long for the industry to come up with the tool-less chassis? In retrospect, it seems both so obvious in its value and straightforward in its implementation. On paper, it doesn’t seem like a big deal to have to grab a screwdriver or work some thumbscrews manually. But in practice, it was always a pain. Now common across the workstation industry, the tool-less chassis means you can leave the screwdriver in the drawer when it comes to opening the enclosure, swapping cards or adding drives.

On the S10, a plastic cage can be released to gain access to the card slots, snapping closed with a retention bar that adds gentle pressure to (one or more of) the cards. Plastic rails attach to both sides of the drive (locked into the threaded holes on the side where the screws would otherwise go), then slides—and snaps—into place. Lenovo’s tool-less design worked very well, though the snaps and plastic slides and constraints did seem a bit more flimsy than what we’ve seen in HP’s machines (for example).

The front and rear panels

If anyone is living by the philosophy that one can never have enough USB ports, it’s Lenovo. The S10 offers ten in all, with two in the front and eight in the back. We assume it’s not costing the company much to throw them all in there, so we’re all for it.

But better than the excess of USB ports (at least we think so) are the built-in front-panel card readers, four in all: CompactFlash, SD/MMC, SM/xdD and PRO. By comparison, our recent top-of-the-line HP workstation didn’t have them (which meant we had to dig up an old USB CompactFlash reader, which of course was only USB 1.0, not nearly fast enough for downloading hundreds of 5+ MB pictures). We think front-panel readers should always come standard.

Rounding out the rear panel I/O are dual-gigabit Ethernet controllers, legacy PS/2 keyboard and mouse ports, the extra aforementioned USB I/O, five analog and one digital optical sound ports (motherboard audio), and a serial port.

The test system: ThinkStation S10, AMD FireGL V7600 and Nvidia Quadro FX 4600.

The ThinkStation S10 Lenovo supplied us with came configured with a middle-of-the-road Core 2 Quad Q6600, 2 GB of 1067 MHz DDR3 memory and 1 TB of SATA storage. For graphics, Lenovo supplied their top-end offering, the Nvidia Quadro FX 4600. Second only to Nvidia’s Quadro FX 5600 in Nvidia’s Quadro FX line (at least considering single-GPU cards), the Quadro FX 4600 is a G80-based, 768 MB high-end card introduced in February of ’07.

While Lenovo was providing us with a ThinkStation, AMD was loaning us a new FireGL V7600 card for evaluation as well. Killing two birds with one stone, we took the opportunity to run system tests with the OEM Nvidia card, and then once again with the FireGL V7600. Seeing as how Lenovo’s only shipping Nvidia products in their ThinkStations, our FireGL V7600 found itself in unfamiliar territory.

The machine Lenovo provided also came with a Creative Soundboard X-Fi PCI sound card, which we didn’t exercise.

The benchmarks

Isolating the graphics. To focus on the graphics, we ran SPEC’s Viewperf 10, which eliminates as much CPU overhead as possible to exercise the card’s rendering as fast as possible. Then, to get an idea of what the overall system performance might look like in a typical real-world workstation application, we ran two familiar and (reasonably) well accepted—though not perfect—publicly available benchmarks, one for CAD and one more relevant to digital content creation: Cadalyst’s AutoCAD 2008 benchmark and Cinebench, respectively. And finally, we attempted to run one of our favorite whole-system benchmarks—SPECapc’s application benchmark for 3ds max— but that attempt fell short as we were not able to get the benchmark running properly on both cards with our unregistered, trial copy of 3ds max.

The Quadro card presented some issues. As a (near) top-of-the-line card, the Quadro FX is AutoCAD certified (it’s on Autodesk’s list) meaning that all relevant AutoCAD functions should be turned on for maximum acceleration—functions like the Gooch shader (more appropriate for technical illustrations than the more well-known Gouraud shader) and smooth lines (specific to AutoCAD).

But for some reason, AutoCAD couldn’t find the Quadro 4600 in its own database of certified drivers. It then informed us of that fact, and proceeded to turn off a lot of accelerated functions by default. Ironically, this is precisely what Autodesk’s new certification program is supposed to avoid, but we got around the issue by manually turning on all acceleration features. What we didn’t get to check out was Nvidia’s AutoCAD-optimized Powerdraft driver, but more on that ahead.

The FireGL V7600 didn’t have any issues running AutoCAD, in Direct3D mode that is. The application recognized it as a certified card and enabled everything automatically. But there’s always some snag when it comes to these exercises. And, in the case of the V7600, it’s a good thing verification of D3D went smoothly, because Autodesk disabled OpenGL mode on the card (an act its web site says is “due to stability issues”). We did manually tune the FireGL V7600 for maximum performance (rather than quality).

The results

First, the disclaimers. No benchmark is perfect. Let’s put aside bias, specmanship and marketing spins, and instead focus on the altruistic goals of giving a user or prospective buyer a useful tool. Accepting all that, benchmarks are still far from the be-all and end-all for system evaluation, as all must select specific workloads and actions to run. And while evaluators can choose those benchmarks that best align what they do on a daily basis, the workloads—and tasks that are stressed the most—are never truly personalized.

Then there’s the issue of staying current. Because they take time to develop (usually by folks not paid to do the work, plugging away in their spare time), benchmarks often stress what applications and users were doing two or three years ago, rather than today. Take AutoCAD and Cadalyst 2008. AutoCAD has been cranking up the usage of 3D Hidden mode, and while Cadalyst has been trying to catch up, one can make a reasonable argument that 3D Hidden mode is significantly underweighted by the benchmark.

So what’s the solution? Skip benchmarks entirely? Like everything, the answer is moderation. Realize that one benchmark won’t tell the whole story and take results from that subset of impartial, third-party benchmarks most appropriate to your situation. Make sure the testing is apples-apples (as much as possible), and perhaps assess your own weighting as desired. Then view all the benchmarks as a worthy-but-not-absolute indication of the hardware’s effectiveness, as one component in an overall assessment or set of buying criteria. A benchmark should be one of many tools, not a stand-alone decision maker.

Exercising the CPU. At first glance, one might assume a 3D rendering test like Cinebench tears at the graphics, while letting the CPU off the hook. On the contrary. With software-centric rendering, it’s the CPU (or CPUs) getting most of the exercise. And accordingly, the results bear that out, with the numbers varying very little by card.

Focusing on the graphics. Viewperf 10.0 (single-thread) resulted in a tight race across the different viewsets. The more expensive Quadro FX 4600 more often beat out the FireGL V7600, but not by huge margins. And the less pricey FireGLV7600 even edged the Quadro card on two viewsets (Maya and Ensight).

We were not able to compare the antialiasing measures provided by Viewperf 10.0, as we could not get numbers for the Quadro FX 4600. What we can say is that the FireGL V7600’s antialiasing provided solid performance, with frame rates dropping off only modestly, as quality improved. In several cases, 8X AA was comparable to no AA, and worst case, performance dropped by just a little more than 50% (with lesser penalties for 2X and 4X).

Surveying the whole system. With SPECapc 3ds max not working out in the end, our one snapshot measuring whole-system performance centered on Cadalyst ’08, which attempts to provide an idea of performance running AutoCAD.

The overall score favored the Quadro FX 4600, but again, not by overwhelming margins. The Quadro FX 4600 did outscore the FireGL V7600 substantially in the 3D graphics index. Again, remember this test was not run with Nvidia’s Powerdraft driver for AutoCAD, which, in theory, should have optimized performance further, as compared to the standard (but current) driver setup.

What do we think?

Lenovo timed its push into the workstation market with Intel’s 45 nm overhaul of both its Core and Xeon based platforms. So, given Lenovo’s timing, we were surprised to find that the S10’s processor options did not (at least initially) include the latest 45 nm Penryn-based Core 2 Duos and Quads, but the older 65 nm models with 8 MB (2 x 4 MB L2). Currently (as of this writing), Lenovo’s storefront doesn’t provide any 45 nm processor options, making it the exception among its peer group (notably Dell and HP).

Besides missing out on some of Penryn’s architectural enhancements (generally modest, but including the notably bigger 2x6 MB L2), where the older 65 nm processors fall behind is in clock rate and front-side bus speed. We’ve no doubt the 45 nm parts will roll in over time. But even if most buyers were to opt for lower-priced, previous-generation parts, to compete with the big two, we think the top-end options have to be there (bear in mind, however, there is a more pricey 65 nm Core 2 Extreme option).

Regardless, the S10 is a solid performer. Though at the entry-level, differentiating by performance isn’t easy, with competing machines all based on the same core platforms from Intel. Product differentiation is often better accomplished by things like aesthetics, power consumption and ergonomics (e.g. noise, access, availability and layout of I/O), as well as many issues beyond the hardware design—things like price, customer relationships, vendor reputation, and support.

We liked the copious I/O, and the chassis appears well-designed. Noise wasn’t a noticeable factor (though no decibel measurements could be made, nor for power consumption). On the rest of the counts though, the jury is out, at least for now.

We know the company is aware of how this market works and isn’t jumping in with unrealistic expectations. And Lenovo has a big leg up on any other “new” entrant to the marketplace. With a global reach and foundation in place with both its existing PC and mobile workstation businesses, Lenovo is well positioned to shake up the marketplace. And we expect it to do just that. But whether Lenovo will present just a minor annoyance to the big two—HP and Dell—or more dramatically disrupt the status quo—remains to be seen.

With respect to graphics hardware, we were disappointed to find that despite all the best intentions by both vendors, getting a graphics card (shipped in a new OEM system) to perform optimally still required extra work and iterations: ensuring the best Windows drivers (OpenGL and DX) were installed, that applications were using all the features they can, and that any extra driver tools (e.g. Nvidia’s Powerdraft) were being employed where they should. We know both Nvidia and AMD are working to make this stuff as seamless as possible, and we expect that situation to improve—especially considering Autodesk’s decision to raise the minimum system requirements for AutoCAD.

Looking at results, we must remember this was one snapshot and not exactly a fair fight, comparing an MSRP $1,995 ($1,500-ish street price) graphics card and the MSRP $999 ($850-ish street price) FireGL V7600. But keeping in mind the disclaimers, the datapoints revealed by our benchmarking experience would seem to reinforce our expectations—that both vendors’ cards are very capable, and that AMD’s latest round of cards promise very good price/performance (not to say its high-end cards can’t match Nvidia’s all-out performance, but that’s not what this exercise addressed).

In the marketplace, AMD is a distant second to Nvidia in sales, much to Nvidia’s credit. It’s executed very well from every perspective: engineering, manufacturing and marketing. But looking at the latest round, it’s hard to justify AMD’s distant position based on the merits of its products. Its current line-up of FireGL cards are quite competitive, and had the timing of card introductions better lined up with OEMs design-in Windows (which are short and less frequent than on the PC side), we’d think AMD’s market position would be looking significantly better. And that should bode well for the next go-around of OEM products.