GTX280

• 1.4 Billion transistors/li>
• 933 GFLOPS /li>
• 240 SPs/li>
• 10 24 SP clusters/li>
• 8 64-bit memory partitions/li>
• 512-bit memory bus/li>
• Double Precision ieee754 support/li>
• 65 nanometer process/li>

GTX260

• 576MHz core/li>
• 1292MHz Shader Processor speed/li>
• 8 24 SP clusters/li>
• 7 64-bit memory partitions/li>
• 448-bit memory bus/li>
• 65nm TSMC process/li>
• Dual Slot

Here’s a chart of the GeForce 8800GTX, the 8800GTS 512MB, the 9600GT, the GTX260 and the GTX280. Note these are the chips used as the basis for the last 5 major video chips from NVIDIA (excluding the 8600/8500/8400 and the various chips built upon the G92 including the 9800GTX, 9800GX2 and many others). As you can see, NVIDIA doubled the number of transistors going to GTX 280, doubled the ROPs compared to G92, doubled the memory bandwidth compared to G92 and nearly doubled the number of Stream Processors.

The GTX280 is based upon NVIDIA’s GT200 chipset, their 2^nd generation Unified Shader Architecture that started with the G80 in 2006. The chip itself is a bit of a beast with 1.4 billion transistors. When compared to the previous generation G80 series cards, the chip has a transistor count nearly double that of the 8800 GTX card. The number of Stream Processors has nearly been doubled from 128 to 240, accounting for a large portion of the increase in transistors.

Their GTX260, on the other hand, has the same number of transistors but 192 SPs. This card is meant to come in at the under $450 USD price tag. NVIDIA says their product stack from today on the high end will be $649 for the GTX 280, $449 on the GTX260 and $199 on the GeForce 9800GTX bringing their lineup as pretty high-end. As to how ATI’s next generation card will stack up in pricing is unknown at this moment as their HD4xxx series hasn’t launched yet.

GeForce 8800GTX was the first NVIDIA card to be based upon a Scalable Processor Array framework. The SPA is made out of a number of TPCs which is each made up of a number of Stream Multiprocessors (SMs) and in turn made up of 8 Stream Processors. THE GTX280 has ten Texture Processing Clusters, with 3 SMs per TPC and 8 SPs per SM, bringing a total of 240 SPs. The older 8800GTX had 8 TPCs with 2 SMs per TPC and 8 SPs per SM, bringing a total of 128SPs.

The GeForce GTX280 has two different architectural modes, graphics and computing. The graphics mode is similar in many respects to previous generation G80 cores. Computing mode has a hardware-based thread scheduler managing scheduling threads across the TPCs. Compute mode also includes texture caches and memory interface units.

NVIDIA decided to clock the GTX280 at a modest 602MHz. This is a combination of a large chip with 1.4 Billion transistors when compared to the previous generation 9800GTX which had clock speeds in excess of 675MHz with some overclocked versions having 740MHz clocks. The GTX280 has nearly twice the processors of the G92 chip and more ROPs (outputted pixels), allowing for higher performance with lower clock speeds.

The GTX260 has a clock speed of 576MHz a Shader clock speed of 1292MHz and a 999MHz memory clock.  The GTX260 is designed for the lower price point of $449, giving it a TDP of 182W, compared to the TDP of 236W on the GTX280. The GTX260 has a length of 10.5 inches, but two 6-pin power connectors, allowing the card to operate on a PSU without the 8-pin power connector. The chip is cooled by a dual slot Fansink.

In terms of DirectX 10.1, the GeForce GTX280 does not support 100% of the feature-set and therefore NVIDIA cannot claim compliance with it. NVIDIA says that the game developers only wanted a few things in DX 10.1 and Microsoft says 10.1 is an incremental upgrade, which it is. Of course the only game with DX10.1 support was Assassin’s Creed, but the developers have taken out the support with their first patch due to issues.

NVIDIA’s big push with the GTX280 is using CUDA to accelerate things like Folding@Home and distributed computing. The GTX280 has 240 processing cores which are massively parallel. The push of CUDA is to use the GPU to do other things than play video games. CUDA uses the standard C computer language as making math calculations and computational biology.

Transcoding, the procedure by which you change the mode of playback from a movie to another is greatly improved with the GTX280. NVIDIA tested transcoding a HD video stream to H.264 format. On a Core 2 Quad CPU with integrated graphics it took 10 hours and 33 minutes. On the same CPU with NVIDIA gtx280 the transcoding took ½ an hour, meaning a speedup of 18X.

Ageia, the company that built the PhysX accelerators was bought by NVIDIA a little over 4 months ago.  PhysX is software that allows software or hardware based physics to be run either on the CPU or the graphics card assuming it is supported by the platform. PhysX titles are available for the Xbox360, PS3, Wii, Windows PC platforms. With the release of the new driver in a couple of weeks, PhysX will be enabled on all Geforce 8 and later cards, with the GTX280 being the first to the fray. This means that games like Unreal Tournament and Gears of War which have support for PhysX will now be supported.

The GTX280 has a 512-bit memory bus with 1GB of memory. This is NVIDIA’s first foray into the 512-bit memory bus with the G80 having a 384-bit memory bus and the more recent GeForce 9800GTX having a 256-bit memory bus. The GTX280 has a memory clock speed of 1.1GHz, 2.2GHz effective. The total memory bandwidth on the GTX280 with default memory clocks is 140.8GB/second, which is a massive jump compared to the 9800GTX which had a memory bandwidth of 70.4GB/second.

The GTX260 has a 448-bit memory interface with 896MB of memory. This seems like a strange memory bus but with 192 Stream Processors makes the 512-bit memory interface a bit excessive which makes this choice clearer. The GTX260 has a memory clock of 1GHz. This means that the 260 will have a memory bandwidth of 112GB/second or more than an 8800Ultra which had 103.7GB/second on a 384-bit memory bus. Memory bandwidth comes into play in situations where anti-aliasing and high-resolutions are used and with the latest games having to use large amounts of Framebuffer memory. The GT280 chip is the first NVIDIA chip to support Double Precision 64-bit floating point computations. DP benefits high-end scientific, engineering and financial computing applications. Each Streaming Muliprocessor (SM) has a double-precision 64-bit floating math unit, meaning there are 30 DP 64-bit cores on the GT280. NVIDIA says that the DP performance is equivalent to that of an 8 core Xeon CPU, about 90 GFLOPS in DP mode.