   |
|||
|
|||
|
:
News
: Archives
|
|
State of 3D: Q4 2007 |
|
|
8800 GT, 55nm, and Other Surprises |
|||
|
|
|||
|
By Josh Walrath |
|||
|
One area that we have not been able to get a straight answer on is the ability of the SPs to handle double precision floats. The upcoming Tesla boards are supposed to be built on a next generation chip which supports double precision, but the question we have been asking is if the G92 is that chip, and just has that functionality disabled for desktop boards. I am unsure of how hard it is to integrate double precision functionality without wildly inflating the size of the SPs. My gut feeling is that integrating that functionality into the current SPs would increase the size only 5% to 15%, depending on how they decided to implement it. Back in June though, the people in charge of Tesla said that it will be a separate chip with that functionality, and the modular setup of the architecture allows them to spin off a separate chip at a minimum of cost. At this time I am going to guess that the Tesla chip will be an architecturally different chip from the G92, but the differences are going to be pretty minimal. NVIDIA did quite a bit of work internally to optimize memory access, which is important considering that the 8800 GT “only” has a 256 bit memory bus and 57 GB/sec of bandwidth available to it. The biggest improvement is in the use of the G84’s texture clusters which feature double the address units. The ROPS and MSAA units also underwent work to make them utilize what bandwidth there is more effectively. The work here has resulted in a card, that while supposedly inferior to the 8800 GTX in terms of memory bandwidth and functional units, comes very close to the GTX in performance. Only when high levels of AA and resolution are reached, does the 8800 GTX take a commanding lead. The 8800 GT also pulls a whole lot less power, and delivers less heat to the system, then its 90 nm predecessors. The card is supposed to pull around 110 watts at load, but from all indications that number is fairly conservative. For PCI-E 1.x slots, the PEG power connector is still needed, but PCI-E 2.0 slots which can supply 150 watts of power could easily handle the card without the external power connector. The chip does run fairly warm at idle (60 C), and can reach upwards of 80 C when at load. The cooler they use is fairly quiet, and does a good job of removing heat. The one area in which the card absolutely excels is the price. The 512 MB version is selling for around $259. Some of the initial boards that were offered before the NDA expired were priced upwards of $279. Expect prices to drop to recommended levels soon, and overclocked versions have already hit the stores (though all based on the initial reference design). At this time NVIDIA has released the strongest price/performance card the world has yet seen. It has all of the impressive G80 features, and adds to them with PV2 and improved texturing and efficiency. Throw in the single slot design, decent heat production, and pretty tremendous power efficiency, NVIDIA has a real winner on its hands. For at least the next three weeks. The Upcoming Battle Keen eyed readers will have noticed that AMD has not been silent about their upcoming parts, and there is a very positive buzz going about both inside and outside of company. The RV670 part from AMD looks to erase the unhappy memories of the older R600 core that was featured in the HD 2900 series of cards. The R600 architecture is not weak by any stretch of the imagination, but we can look at that particular product and see that it was seriously un-optimized. ATI was very late with the R600, and I believe that in many ways it was rushed to market to oppose the G80. ATI probably did not have all that much time to fine tune the architecture, and the biggest issues that it faced were those of under-utilizing the shader power of the chip, as well as the unacceptable performance hit when using anti-aliasing. The results of this chip were somewhat disastrous for AMD. It did have a competitive part, but they did not have the “GeForce Killer” that many were expecting from a larger chip with a 512 bit memory bus. Sometime in mid-Summer of 2007, AMD decided to stop all wafer orders for the R600 chip. The chip itself was large (420 mm square), sucked up a huge amount of power (180 watts +), and was expensive to produce on TSMC’s 80 HS line. In late Summer AMD saw that the RV670 was going to be something special for them. To get that to market ASAP, as well as clear out R600 inventory, AMD released the HD 2900 Pro cards to soak up the excess inventory. Those cards were available for about a week and a half before they were sold out. Looking online, we also see that there are few HD 2900 XT products available, and those that are on shelves are priced quite high as compared to what was offered just weeks ago. Apparently AMD’s partners are very happy about how AMD has approached the EOL of the R600, as few of them have inventory on hand anymore. The RV670 looks to be AMD’s saving grace this holiday season. It will also be the basis for many products hitting the shelves early next year. AMD took a very aggressive stance on the design of this product, and it looks as if those risks have paid off. The biggest risk they faced was the use of TSMC’s new 55 nm High Speed process. We need to look a bit into the past to see how being aggressive on process nodes has affected 3D graphics cycles.
If you have found this article interesting or a great help, please donate to this site.
Copyright 1999-2007 PenStar Systems, LLC. |
|||
