Nvidia developed a new unified PCB design for the G92-based graphics card series. The PCB is shorter than the one used for the GeForce 8800 GTX which solves the problem of installation into small system cases like the Antec NSK1380, making it possible to build very compact yet very advanced gaming stations that are as fast as systems built in full-size ATX cases. The PCB is so compact because the developer has abandoned the 384-bit memory bus that made the wiring so sophisticated and has transitioned to a thinner tech process which makes the G92 more economical than the G80.
However, Nvidia seems to have had a fit of gigantism when designing its GeForce 9800 GTX. The new card is the same size as the GeForce 8800 GTX. This cannot be explained easily. We know of many models of GeForce 8800 GTS 512MB with the same specs that have a short PCB. Gainward has also released a version with 1024MB of graphics memory that is superior to the GeForce 9800 GTX in every parameter except for the memory frequency (and the difference is as small as 50 (100) MHz). All of those cards were perfectly stable delivering highest performance under harsh conditions of our tests. There was no real need to return to the era of GeForce 8800 GTX, but the GeForce 9800 GTX brings us back to it.
The continuity of generations is obvious: the new card look as impressive as its predecessor with the same suffix. The PCB is as long as 27 centimeters (without the mounting bracket) as opposed to 23 centimeters of the GeForce 8800 GT/GTS 512MB. This makes it impossible to install the GeForce 9800 GTX into some small system cases. For example, it will press against the HDD cage in Antec’s Sonata Plus. You may suppose that the PCB is larger to reduce the number of its layers, but that’s not true. It has 12 layers whereas the PCB of the GeForce 8800 GT/GTS 512MB had 10 layers and a shorter length.
The Gainward Bliss 8800 GTS 1024MB was quite satisfied with a three-phase power circuit, but the GeForce 9800 GTX uses a four-phase one despite the lower frequency of the GPU, which is the main consumer. The power circuit is based on the Primarion PX3544 controller we are familiar with by the whole series of Nvidia’s G92-based cards. The memory has an independent two-phase power circuit controlled by an Anpec APW7066 chip we have not met before.
What looks like another inexplicable caprice of the developer, the card has two 6-pin PCI Express 1.0 connectors. The GeForce 9800 GTX is unlikely to consume more than 110W, so there is no actual need for two connectors. The combined load capacity of one such connector and the power section of the PCI Express x16 slot is 150W, which is far above the requirement of any G92-based card as we could see by the example of XFX GeForce 8800 GTS 512MB XXX and Gainward Bliss 8800 GTS 1024MB. So, that’s a questionable solution that doesn’t make it any easier for the user. The power connectors are placed like on the GeForce 8800 GTX.
The left part of the PCB is designed in a classic way. There is nothing extraordinary here. The card doesn’t have a seat for a DisplayPort chip that used to be present on every GeForce 8800 GT/GTS 512MB. The memory chips are placed around the GPU.
The card carries a total of eight GDDR3 memory chips (Samsung K4J52324QE-BJ08, 512Mbit, 16Mbitx32, 1.9V voltage, 0.83ns access time, rated frequency of 1200 (2400) MHz). Such chips were installed on the Gainward Bliss 8800 GTS 1024MB GS GLH. The GeForce 9800 GTX has a memory frequency of 1100 (2200) MHz. With a 256-bit memory bus, the memory bandwidth is 70.4GBps. This is lower than the GeForce 8800 GTX’s 86.5GBps, but the G92 has a more advanced memory controller that allows it to compete successfully with the G80 across most applications. Our tests of the GeForce 8800 GTS 512MB prove this point. The additional bandwidth growth only ensures that the new card is not slower than the previous GTX-indexed flagship even at high resolutions. We’ll see shortly, in our tests, if it’s really so in practice.
The smaller amount of memory (512MB as opposed to the GeForce 8800 GTX’s 768MB) may be a factor in memory-sensitive applications. We don’t know why Nvidia didn’t equip its new flagship with 1024 megabytes of memory although this is not something extraordinary from a technical point of view as the Gainward Bliss 8800 GTS 1024MB GS GLH proves. Nvidia must have been limited by economical reasons: fast memory is rather expensive and using two times more chips would make the new card much costlier.
The GPU is marked as G92-420-A2. We already saw variations in the marking of the G92 chip when we tested the GeForce 9800 GX2 whose cored were marked as G92-450-A2. To remind you, the GeForce 8800 GT/GTS 512MB carry G92-400-A2 chips. So, all of them have the same revision number but differ in the middle number. We can only guess that it’s the way Nvidia denotes the frequency potential of the chip. This sample of the G92 is dated the 7th week of this year, i.e. February 17-23.
Of course, there is nothing new about the GPU configuration: the graphics core contains 128 unified streamed processors, 32 (64) texture processors and 16 raster operators grouped into 4 sections with four 64-bit memory controllers. The main GPU domain, including TMUs, texture caches, memory controller and raster operators, is clocked at 675MHz. The shader processor domain also called Lumenex Engine is clocked at 1688MHz. This is not much higher than the frequencies of the GeForce 8800 GTS 512MB and resembles the GeForce 8800 Ultra not being far faster than the GeForce 8800 GTX.
A significant difference of the GeForce 9800 GTX from the GeForce 8800 GTS 512MB is that it has two MIO connectors to enable triple-core 3-way SLI configurations. The 2-pin S/PDIF connector has been transferred to the right part of the card as the consequence. Besides that, the card has two DVI-I ports supporting dual-link mode and a 7-pin mini-DIN connector. Besides serving as video output in RCA, S-Video and YPbPr formats, the latter seems to do double duty as an external S/PDIF output as implemented in some models of GeForce 8800 GT.