: News
: Reviews
: Editorials

: Archives
: About
: Contact
: Advertising
: Privacy
: Links
: Forums

Creating the 512 MB 7800 GTX

Insights into the High End

by Josh Walrath

            One of the first things to be developed for the “Ultra” was the cooling system.  This had to be quite robust yet avoid the “Dustbuster” reputation that NVIDIA accrued during the disastrous FX 5800 Ultra days.  The design they came up with was the one initially spotted with the Quadro FX 4500 version using the G70.  The core of the cooler is copper (which touches the GPU die), while the fins are aluminum.  Copper is outstanding at transferring heat, but is not as effective as aluminum at dissipating heat into air.  The combination of a copper core and aluminum fins insures that the heat is more evenly spread throughout the cooler, and it maximizes heat dissipation in air.  The addition of heatpipes also helps to more evenly spread the thermal load.  The 80 mm fan does a very good job of pushing air yet staying fairly quiet.  Overall the cooler NVIDIA developed is quite impressive.

            The next thing NVIDIA had to do was spin a new PCB revision that would accomplish several goals.  The first was to supply more power cleanly to the GPU to achieve higher clock speeds.  The second was to maximize memory speed.  Now, 1.1 ns GDDR-3 was not available at that time, so NVIDIA would have to take available 1.4 ns memory to its maximum of 700+ MHz (1400+ MHz effective).  Trace lengths and power had to be tuned to achieve the maximum speed.  This made the actual PCB more expensive than what we see with the standard 7800 GTX.  Not excessively so, but still enough to make a difference to the bottom line.

            NVIDIA did not know what speed ATI would be introducing their top of the line card at, and their best guess was in the 700 MHz range if the worst rumors were true.  To be able to match the overall performance of the ATI part NVIDIA would have to clock the G70 as high as possible.  The big issue here was supplying good, clean power to the card.  While NVIDIA could tune the PCB design as much as possible, they however did not have control of the power supply that most users would have when installing this card.  The idea of an external power supply was dredged back up from the 3dfx days.  This compact power supply would provide the necessary clean power to the card to help it achieve high speeds with stability.

            Work was started on the external power supply and the PCB design to accept it, and then something interesting happened.  ATI didn’t make an order for the R520 with TSMC.  NVIDIA had placed wafer orders for the G70 for a late Spring release, but ATI did not do the same with their R520.  Better yet, it became known that ATI had done a major revision on the R520 and was waiting for first silicon on that.  Something didn’t smell right, so the plans for the Ultra were put on the back burner.  With the current schedule that ATI was on, NVIDIA would not expect to see a competitor for the 7800 series until at least Mid-Summer.  This left NVIDIA with a lot more flexibility on product placement.  Not only that, but NVIDIA heard from reliable sources that ATI couldn’t clock their R520’s past 500 MHz without pushing a lot of voltage and creating a lot of heat.  This pushed the plans for the Ultra out the back door.  NVIDIA would not need an external power supply, nor would they even have to release an Ultra in the Spring to one-up ATI.

            ATI received the second R520 revision back, and it still showed the same problems as the first revision.  This gave NVIDIA even more leverage in placing product.  By early Spring NVIDIA had its lineup solidified with what would become the 7800 GTX in late June and the 7800 GT in August.  NVIDIA knew that ATI would not be able to launch the R520 and be competitive with what NVIDIA had by late Spring or even early Summer.  This gave NVIDIA the luxury of actually building up stock of the 7800 GTX parts and offering instant availability on the day of release.  If ATI had gotten the R520 up and running as it should have been, both ATI and NVIDIA were planning on releasing their respective parts in late Spring and actually delivering parts about 1.5 months later (as they had done with all their previous parts).  With ATI falling far behind NVIDIA decided to really put the screws to ATI by launching the 7800 series later but offering instant availability.  This proved to be quite a feather in NVIDIA’s hat.

            This did not mean that the Ultra was totally shelved.  NVIDIA found that the G70 chip itself could be clocked very high with standard cooling and voltages applied.  The 430 MHz core that the GTX has for reference is quite conservative, and many manufacturers have taken advantage of that fact by delivering products clocked significantly higher than that.  Most consumer GTX cards can hit 490 MHz to 500 MHz quite easily, even with stock cooling.  Companies like BFG Tech and Leadtek have offered products with more robust cooling which should allow a bit more clock overhead.  Yields for the G70 chips were very good from the very start.  Chips that did not make the full GTX speed and pixel pipeline density were binned down to the 7800 GT line.

            By July the G70 chips had been in production for a good 5+ months, and NVIDIA had been receiving finished chips for about 2+ months.  Chips that could be clocked about 500 MHz were set aside, and NVIDIA told TSMC to start tweaking the process to achieve a better speed bin.  Yields stayed the same but NVIDIA started to get chips back in late September that could clock higher than before.  The 512 MB 7800 chip is not actually a new revision, contrary to what some sites are saying.  Again, TSMC tweaked the process to achieve greater speeds while keeping yields consistent.  With ATI about to release the X1800 series of cards, NVIDIA knew they should probably have an answer.  This time another couple of factors came into play.

Next: Why No Ultra?

Copyright 1999-2005 PenStar Systems, LLC.