Video memory bandwidth

Calculation formula

DDR3 video memory bandwidth = video memory frequency × video memory bit width/8, DDR5 video memory bandwidth = video memory frequency × video memory bit width/8. Most low-end graphics cards can provide 6.4GB/s, 8.0GB/s of video memory bandwidth, while mainstream graphics products provide more than 75GB/s of video memory bandwidth. When conditions permit, buy a graphics card with a large memory bandwidth as much as possible, which is the key to choice.

Constraints

Because the video memory bandwidth refers to the exchange speed between the graphics processing chip and the video memory, the wider the number of bits of the video memory interface bus, the higher the exchange rate. , And the faster the video memory, of course, the higher the bandwidth. For the bus, although the display chip has now developed to 256bit, it only uses a 128bit video memory bus. The speed of the video memory cannot keep up with the speed of the display chip, which will cause a bottleneck problem. Like GeForce 256, because its 166MHz 128bit bus can only provide a maximum bandwidth of 2.6GB, its performance under 32bit color cannot be fully utilized. As for the video memory, like the main memory of the system, its development speed will lag behind that of the processor. Although CPUs above 700MHz have been developed, the speed of the memory is still below 133MHz. This obviously has a great impact on the system. Video memory also has the same problem. Its development speed has been significantly behind the development of graphics chips. The use of ordinary SDRAM has increasingly affected the performance of the chip. Of course, this is mainly due to the price of video memory. If the GeForce is equipped with 200MHz video memory, the price will be about the same as the DDR version.

The impact of parameters on bandwidth, first of all, for high-performance graphics cards, 128bit video memory bus must be used. Savage 4 using 64bit bus is obviously behind other fourth-generation cards in performance. Very good proof. In the case of the same bus speed, the faster the video memory speed, the higher the bandwidth. For DDR-RAM, since two data transmissions can be completed in one cycle, it will generate twice as much in a certain period of time. The bandwidth of the same frequency SDRAM. However, it should be noted that this does not mean that the speed of DDR video memory is twice that of SDRAM, because the performance of video memory is not only affected by data transmission. When data is transferred from the graphics chip to the video memory, its execution speed is the same as that of SDRAM, so We cannot say that DDR-RAM is twice as fast as SDRAM. After all, DDR-RAM is a low-end solution that solves the transmission bottleneck problem.

Video memory impact

The G400MAX with a video memory speed of 200MHz and a 128bit bus has a memory bandwidth of 3.2GB/s, while the TNT2 Ultra, which also uses a 128bit bus, uses only 183MHz. The memory bandwidth is smaller than that of G400 MAX, only 2.9GB/s. What impact does the memory bandwidth have on the accelerator card? Let’s take a look at the following example, after the graphics chip receives instructions from the CPU , Calculate the pixels that need to be filled, and then pass the pixels through the video memory and other channels, and finally complete the digital-to-analog conversion display. Therefore, if the transfer rate of the graphics chip and the video memory channel is not enough, the pixels processed per unit time are only limited Video memory bandwidth. GeForce 256 with a bandwidth of 2.6GB can reach a fill rate of up to 480MT/S, which is exactly the same as the design. When there is a very important problem, when 32bit processing is performed, it will be more than usual. A write operation, so that GeForce with only 2.6GB bandwidth can only achieve a fill rate of 280MT/S. If you observe carefully, you will find that for a 3D card with 32bit rendering capability, its memory speed is faster than the chip speed. Less, this is to reserve the bandwidth for 32bit rendering. The 150MHz TNT2 can provide a fill rate of up to 300MT/s when performing 16bit rendering, but it can only maintain 200MT/s when performing 32bit rendering. The fill rate is higher, but its loss is much smaller than that of GeForce. Because GeForce 256 uses a video memory with a speed of only 166MHz, the bandwidth requirements are very tight even at 16bit, not to mention 32bit rendering. So, nVidia is in Some approaches to the fill rate of GeForce 256 have failed. Although we have never got a graphics card with the same speed as 16bit and 32bit, the excessive shrinkage of the GeForce 256 memory has caused the speed loss in 32bit colors to sometimes reach 16bit. This is obviously unacceptable. After all, playing games with 32bit is one of the important features of GeForce.