Recently, NVIDIA introduced its current king of the hill product in the form of two GK104 GPUs placed on the same board and called the part GeForce GTX 690. In the next couple of days, we will see the launch of GTX 670. We'll dig a bit deeper and see the consequences of recent strategy moves.
Many of you, our respected readers, industry analysts, media were all surprised at the cadence NVIDIA executed its Kepler launch. This is the first time the company launched both high-end and mainstream GPUs in its history (GeForce 4 Ti / MX launch comes to mind, but the MX part was based on GeForce 2 die).
The Birth of Kepler came from Fermi
If we want to come to the point of how GeForce GTX 690 came to life, we have to go back all the way to October 2009. While AMD gathered large number of press, analysts and partners for what it turned out to be the final nail in the ATI Technologies coffin (ATI Radeon HD 5000 Series was the last before the rebrand into AMD Radeon) – NVIDIA did a gathering of a different sort.
The quarterly global meeting was not in a good mood. The global economy was heading south, and the executive management pulled what seems to be the decision that turned the company around – instead of reducing headcount, complete executive management shifted to a $1 salary after tax and benefits. While "the $1 CEO club" usually means executives receive millions of dollars in stocks and other types of compensation, Jen-Hsun Huang and his team decided otherwise. The total salary AFTER tax and benefits was cut to $1, and no employee would get fired.
Some executives decided not to stick around and left, but the remaining majority of executives suddenly had an army of zealots who were grateful somebody else is weathering the storm for them.
Second part of the execution initiative was the broken Fermi die, i.e. GF100. The first batch of engineering silicon that was supposed to go head to head against the AMD/ATI came back from TSMC in a non-functional state. While ATI had a similar problem in 2005 with its R520 Fudo, NVIDIA's problem was not a single transistor as with R520 (X1800), it was the die as such.
At the GPU Technology Conference in 2010, Jen-Hsun Huang explained; "We found a major breakdown between the models, the tools and reality. So when we first got the first Fermi back, that piece of fabric, so imagine we're all processors, all of us seem to be working. But we can't talk to each other. It's like we're all deaf, we're all mute and deaf. And we found out it's because this connection between us is completely broken. It turns out the reason why the fabric failed isn't because it was hard, but because it sat between the responsibility of two groups. The fabric is complicated in the sense that there's an architectural component there's a logic design component and there's a physics component. My engineers who know physics and my engineers who know architecture are in two different organisations, and so you see this underlap of responsibility… 'is it my job or your job?' If you'd just simply moved it from one side to the other side they'd been more than happy to pick up the slack, but we let it sit right in the middle. 'Let's be both of our jobs'… that's a bad answer."
The result of GF100 was newly found approach for the subsequent Kepler generation. It was decided to do a die-shrink of Fermi dies such as GF100 (GTX 480), GF104 (GTX 460) etc. giving birth to GF110 (GTX 580), GF114 (GTX 560 Ti) and so on and so forth. In that time, AMD executed Evergreen architecture (Radeon HD 5000), Northern Islands (Radeon HD 6000) and now Southern Islands (Radeon HD 7000). However, NVIDIA had a new approach to launching the parts and for the first time, the company was working on three dies at the same time, what was called high-power, high-efficiency and low-power dies, i.e. entry, mainstream and high-end silicon. The high-end silicon is supposed to debut really soon, on a HPC event on May 10th.
When the company received first GK100, GK104 and GK107 dies, the performance of GK104 shocked the company executives. Our sources were telling us that there was a genuine surprise from CEO to PR/Marketing/Sales teams, since GeForce GTX 660 part was outperforming GTX 580 by a significant margin. The decision was made to rebrand the part into GTX 660 and 670 Ti, with GK100 taking the role of head honcho. Then, AMD came out with Radeon HD 7970, the very first part based on its Southern Islands architecture. The performance unveiled was the reason why the author of these lines got multiple calls from NVIDIA insiders asking "Does AMD have hidden cores in the die? What did they disable?" However, with Southern Islands being AMD's first compute intensive part (just like G80/Tesla and GF100/Fermi were compute milestones for NVIDIA), the company had to sacrifice some efficiency/performance for their first true GPU compute architecture.
The decision to rebrand GK107 into GTX 680/690 was made. Furthermore, the company decided to further increase the performance of the high-end die, giving birth to GK110. GK110 will debut in just three days, but this is the part that targets the HPC i.e. GPGPU community. We were told that the number of pre-orders for Kepler-based Tesla cards (Tesla 3000 Series) should exceed the overall number of Teslas shipped so far (over 150,000 units sold).
Getting back to GK104, the newly reorganized NVIDIA board team went on to build the GTX 680 first. Then, they followed up with Gemini (GTX 690) and the lessons learned on building the Gemini PCB were used on GTX 670. It is not excluded that you are even going to see a GTX 680 board using the PCB from the upcoming GTX 670, since it improved two steps. Seeing a GTX 680 4GB using a PCB that is smaller than GeForce GTX 550 and Radeon HD 7650 was considered impossible, but the new engineering cadence at the green company is executing on target.
GeForce GTX 690
All of this brings us to why GeForce GTX 690 product is a paramount product for the company. The performance achieved by GK107 and GK104 were beyond targets and the decision was made to go against AMD with a dual-GK104 board.
The performance results of the board show that engineers did not sacrifice anything in the design and you can run the card at GTX 680 clocks without any issues. We've been talking to our sources in the AIB board manufacturing community, and seeing a 1.2GHz clock for both GK104 dies isn't all that uncommon.
During the creation phase, GeForce GTX 690 was planned as the no-cost-sparing part, with top parts being used across the board. Beside the fan fins, there are no plastic parts on the card. The PCB is a 10-layer one with a layer of copper ("2 ounce", popular amount of copper on PCBs) for increased thermal conductivity.
The cooling solution for this card was really expensive. In fact, we've been told that the company budgeted this part for over three times as much as the most expensive cooling system they've used so far (you've guessed it right, GTX 295 and 590 cooling had a price tag of an "arm and a leg"). The base for the heatsink used to be plastic, but it is now made of aluminum with trivalent chromium plating. The bimetal approach results in extreme rigidity, which prevents PCB from bending down. This is one of reasons why NVIDIA didn't needed to use a plastic or aluminum backplate to reduce PCB flex (GTX 295, GTX 590 being prime examples) – base of the heatsink is a stabilizing element itself. Two vapor chambers did not change as much from GTX 590, but with the reduced thermal load, these chambers are more efficient since they're not pushed as hard as it was the case with previous boards. Going back to the base, you can see that it is perforated to better drive the airflow across the board.
An injection-molded magnesium alloy was used for fan housing, and it fits the aluminum base without danger of vibrations causing any bending, as was the case with plastic housing on an aluminum or even plastic base.
The decision to put polycarbonate windows has two sides – one to give you a look into aluminum fins, and another to show you if the heatsink accumulated dust on itself and to give it a good cleaning. From personal experience, decision to go all-metal (besides the plastic fan blades, but metal ones would be a potential security hazard) also reduces the amount of dust that gathers purely by statically charged plastics.
All in all, the GTX 690 is engineered like no graphics part before, and the decision to price it as high as the company did will keep it out of reach for vast majority of users. But, just like the car industry has million dollar cars such as Bugatti Veyron or Rimac Concept One, this industry has a top performing dual GPU part.
Good news for the graphics industry as such is that NVIDIA doesn't plan to stop using these materials in future boards. You can expect to see the technologies used on GTX 690 to spread on the high-end line-up, and GeForce, Tesla and Quadro parts are all planned to get the highest quality components going into the future.