NVIDIA GeForce GTX 280 review
- The GPU
Author: Luka Rakamaric
Date: 18 Jun 2008
Unlike the series 8 and 9 cards, which shared the same GPU design, the new GT 200 GPU has some very big distinctions. First of all, the GPU is almost twice as big, with 1.4 billion transistors. As a result, the new GPU has 240 processor cores, which is almost double the 128 of G80 and G92 GPUs. The architecture of the GPU is based on the same principles, but this time, every Texture Processing Cluster has 3 Shader Multiprocessors instead of 2, consisting of 8 Shader Processors each. Also the maximum number of TPCs per GPU was increased from 8 to 10. In G80 and G92, you had 8x2x8=128 Shader Processors, and the GT 200 GPU has 8x3x10=240. Since games today are relying heavily on shaders, the shader to texture processing ratio has been increased by 50%.
One of the main reasons that the G80 GPU was still superior to the G92 was the reduced number of ROPs, 24 vs. 16, and the reduced memory bus width, 384 vs. 256-bit. The new GPU has 32 ROPs and a 512-bit bus, which enables it to achieve up to 141.7 GB/s with GDDR3 memory. The 32 ROPs were additionally enhanced, so that the GTX280 can blend up to 32 pixels per clock, while the last generation could manage only 12.
Memory clock is 1107 MHz, or 2214 MHz transfer rate because of Double Data Rate. This is extremely fast, on the level of some other GDDR4 cards which have memory that should easily tolerate higher clocks. GDDR3 and GDDR4 should not be confused with DDR3 (or 4) system memory since they are based on the DDR2 standard with some improvements to fit the purpose of frame buffer memory.
Take a look at these cards in single and Tri-SLI mode:
 
Thermal dissipation of the new GPU is much greater than that of the previous ones, since it is almost double in size but still built using 65 nm TSMC production process. However, the new cooler design keeps the card stable, although not by any means cool. Good case airflow is a must if you are aiming for one (or more) of these cards. Our card heated up to 86 degrees, which is tolerable, but you have to consider that our benchmarks are done on a testbed and not in a closed system case.
Power requirements are rated at 235W, and the card uses one 6 pin and one 8 pin power connector. It also supports normal and 3-way SLI mode, if the motherboard supports it. If used in a 3-way SLI configuration, we would recommend at least 1KW power supply. The rear bracket has 2 DVI-I connectors and a 7 pin S-video output. The card also supports HDMI or DisplayPort connectors. It is also the first card that can use 10 bit color out through Display Port, while older cards could use it internally there was no output available.
The VP2 engine is capable of offloading all video processing to the GPU, for both H.264 and VC-1 codecs used in High Definition content. HybridPower technology on supported motherboards can completely shut down the card when working in 2D mode, but for other motherboards idle power consumption is only 25W which is less than the previous generations.
NVIDIA also emphasizes its CUDA technology, which turns the GPU into a big parallel microprocessor. It can be used to recode video, theoretical simulations on an industry level and similar CPU intensive procedures. The GT200 is designed that more than 80% of the die is dedicated to processing, while in CPU world most of the CPU die is occupied by cache, while only 20% is used for computational power. The GPU processing employed by CUDA technology relies heavily on computation power and throughput, while CPUs are more suited to reduce latency and filling the pipelines with as few cache misses as possible. This is because the CPU has to work with serial application for the most time, and when we input heavy parallel tasks it doesn’t perform as well as a GPU could. This includes video editing and playback, scientific computing, physics simulation, and of course 3D graphics. Also, the GTX 280 has 8 on die memory controllers, which can access 141 GB of data per second, which is far more than what system memory on a PC is capable of. We can say that CUDA enables the GPU to do tasks that the user requires that are not very well suited for the CPU. When we compare our testbed’s CPU, Intel’s Core 2 Quad X9650 to the GT200 GPU on the GTX280 we can see the difference in processing capabilities. But we will get down to CUDA in a separate article in our series due tomorrow.
|
Processor
|
Intel Core 2 Extreme QX9650
|
NVIDIA GeForce GTX 280
|
|
Transistors
|
820 million
|
1.4 billion
|
|
Processor clock
|
3 GHz
|
1296 MHz
|
|
Cores
|
4
|
240
|
|
Cache / Shared Memory
|
6 MB x 2
|
16 KB x 30
|
|
Threads executed per clock
|
4
|
240
|
|
Hardware threads in flight
|
4
|
30,720
|
|
Peak gigaflops
|
96 gigaflops
|
933 gigaflops
|
|
Memory controllers
|
Off-die
|
8 x 64-bit
|
|
Memory Bandwidth
|
12.8 GBps
|
141.7 GBps
|
This does not mean that we can do without the CPU or that it’s not that important, it only means that some tasks other than 3D graphics, that are more suited for GPU architecture will with CUDA be run on the GPU.
|
