Few weeks ago, NVIDIA unveiled the Kepler-powered Quadro graphics cards for the mobile workstation, while the first day of Siggraph was used to unveil the second generation Maximus platform, which combines the Quadro and Tesla cards for a unified digital pipeline.

Second Generation Maximus Platform brings two Kepler cards together... but only as of December 2012.
Second Generation Maximus Platform brings two Kepler cards together… but only as of December 2012.

The second generation Maximus platform is a combination between GK104-powered Quadro K5000 (the same chip powering the GTX 660/670/680/690 and Tesla K10) and the GK110-powered Tesla K20. The Quadro K5000 will be available in October, while the Tesla K20 will only come in December 2012.

Quadro K5000 finally addresses the holy triangle of content creativity - three display support, two of which can be 4K panels
Quadro K5000 finally addresses the holy triangle of content creativity – three display support, two of which can be 4K panels

NVIDIA is touting the display capabilities of the card, since we didn’t know that the GK104 chip is designed to run two 4K panels at 60Hz (or a 3D 4K 120Hz projector from SONY Professional and the upcoming one from RED Digital Camera) and an additional 1080p panel. With the K5000, NVIDIA finally answered to demands from the broadcast/movie/FX market, whose workflow included three displays for years and years of time.

The K5000 or How NVIDIA Creates Highest Margin Products

When Nvidia was creating the Kepler architecture, the goal for the architecture was to launch in 2011. However, with the decision coming to launch in 2012, the extra time the company had resulted with a number of different PCB designs which further optimized the design for the GK104 chips. The original Kepler appeared on market as the GeForce GTX 680, followed by the GTX 670 and GTX 690. The GTX 690 design is also serving as the Tesla K10 8GB for those that need ultimate single-precision performance, as the double-precision performance is reserved for the upcoming GK110 GPU (Tesla K20 XXGB GDDR5).

NVIDIA Quadro K5000 4GB uses a very short PCB which made its debut with the GeForce GTX 670. The board is pictured inside the Dell's latest workstation.
NVIDIA Quadro K5000 4GB uses a very short PCB which made its debut with the GeForce GTX 670. The board is pictured inside the Dell’s latest workstation.

In order to create the K5000, NVIDIA utilized the GK104 chip and surprisingly, the GTX 670 PCB design. This enabled company to reduce the costs to an absolute minimum, with the BoM ending probably in the lower $200 range ($180-280 is our guesstimate based on $399 retail price plus additional 2GB of memory), for a card which retails at $2249. Naturally, there is an offset cost in terms of driver development and certification, which is astoundingly expensive when compared to the regular consumer drivers. Still, there is no doubt that the company is earning a lot of money on this card, especially if you decide to pair it with a Tesla C2075 or the upcoming K20. As you can see on the slides, pairing a K5000 with a K20 will set you back for at least $5,548. However, the change in workflow can yield thousands of dollars in savings in terms of reducing the time needed to complete the task at hand.

The K5000 customers will be getting a second generation PCB design which is much more optimized than the first gen (GTX 680) one. This also enabled companies such as Dell and HP to create new enclosures

Not all is Great – Is the K5000 a drop from Previous Generation?

During the Limelight LA party, which NVIDIA organized in cooperation with Dell and LucasFilm we spoke with a lot of industry insiders and learned their perspectives on the Kepler-powered Quadro cards. First off, we learned that some of key members of the digital creating industry were holding off on purchasing the new hardware due to slower-than-expected ramp-up of Intel’s Romley workstation platform, as well as unavailability of Kepler-based Quadro cards. Naturally, when Intel shipped the Xeon E5 and associated Romley platform, the company needed to satisfy the demands for the server market first, leaving the workstation market waiting.

Even though, customers that already had a chance to sample the K5000 4GB and the Intel Romley-based workstations were not as enthusiastic. Quadro 6000 in a 1GPU or 2GPU mode simply excelled with its 6GB of onboard GDDR5 memory, while the K5000 features "only 4GB". There are customers in the automotive, airline and advertising industry that already max out the 6GB and the initial tests showed that they had to reduce the image quality to fit the 4GB onboard frame buffer.

Good beginning, but the industry wants more

Our experience from the two days spent on Siggraph, talking about Kepler-based Quadros ended on a positive note. Just like on the enthusiast segment, NVIDIA will probably sell as much GK104-based Quadro cards (K5000, K5000M, K4000M, K3000M) as they can make, and selling cards with disabled SMX clusters should significantly improve the bottom line (NV pays per wafer, thus every working chip makes money for the company).

Given that the line up consists out of a single desktop and six mobile parts, there is a feeling that NVIDIA is choosing its battles. There aren’t many GK104 chips to begin with, and from the looks of it, GK104 is without a doubt the most widespread chip in company’s history (GTX 660, GTX 660 Ti, GTX 670, GTX 680, GTX 690, Quadro K5000, K5000M, K4000M, K3000M and Tesla K10). However, it is also spread too thin to address all customers, as we probably won’t see the highest-end Quadro chip before the end of the calendar year. Should TSMC fail to improve the 28nm yields, we won’t be seeing GK110-based Quadro (K6000) for quite some time.

One of ways to leverage that problem is the work the company is doing with Fusion-io and their latest ioFX card, which can serve as a 1.4GB/s fast memory, which going into the future – GPUs should be able to address directly.
All in all, Quadro K5000 is a good step forward, but that’s about it. The industry wants more and they’re ready to pay for it.

The real question is, can AMD step in with their FirePro cards and seal the gap?