What is peak memory bandwidth

[lenkiatleong]

Hi,

I am wondering why some high end graphic card offer peak memory bandwidth as high as 2x256 GB/s !!

I am confuse because most motherboard memory bus speed is much much below that, e.g., my Intel DP35DP could only achieve max 8GB/s.

I am not sure what is the highest memory speed today but i believe its not over 30GB/s. So, the question is, do one need to pay top dollar for such cards?

Another thing is, i saw many reviews that run several games and showed the frame per second capabilities of the tested card. High end cards could achieve over 100fps !! But our PC monitor mostly refresh at 75Hz or below. So, once again, do one need to pay top dollar for such cards?

I am posting this because i am not sure what i am writing above so please be kind to me.

Thanks.

 

[BFG10K]

Peak memory bandwidth is simply the maximum theoretical bandwidth the memory system can manage. It’s seldom hit in reality (if at all) because of other bottlenecks in the system.

As for frames per second, there are lots of reasons, including running without vsync, getting a tighter input response, 120 Hz LCDs, and the fact that excess frames can be converted into higher image quality by raising the detail level.

 

[nomagic]

The peak video RAM memory bandwidth is the maximum speed at which data can be transferred to the GPU for drawing images. Suppose that we have a 256-bit double-data-rate memory bus working at 2GHz. Then the peak bandwidth would be 256bits*2*2GHz, which is about 128GB/s.

This much bandwidth is implemented because modern GPUs demands so. In fact, there are cases, in which GPUs are bandwidth starved. If the memory is not fast enough to keep the GPU fed, then the GPU would idle and have nothing to do. Consequently, the performance would suffer.

Those that enjoy their game with everything turned to high at the resolution of 2560 x 1600 would appreciate the extra horsepower provided by their top-of-the-line graphics accelerators.

 

[Seero]

Hi,

I am wondering why some high end graphic card offer peak memory bandwidth as high as 2x256 GB/s !!

I am confuse because most motherboard memory bus speed is much much below that, e.g., my Intel DP35DP could only achieve max 8GB/s.

I am not sure what is the highest memory speed today but i believe its not over 30GB/s. So, the question is, do one need to pay top dollar for such cards?

If you have notice, the relationship between CPU and RAM is pretty different from GPU and VRAM. When you buy RAM, say DDR3, there are weird numbers like PC8500 or PC16000. This number roughly refers to the peak bandwidth, PC8500 = 8.5Gb/sec and PC16000 = 16Gb/sec.

The exact formula is

Peak Bandwidth = (Bus Width) x (Bus Speed x Operations/Clock Cycle)
or
Peak Bandwidth = (Bus Width) x (Data Rate)

Now VRAM runs at a much data rate and much wider bus width. For example, HD5850, 256 bit, 4000Mhz memory clockrate. That is

(256bit/8) X 4Ghz = 32 byte X 4Ghz = 128GB/sec

GT300 was said to have 512 bit interface, that means 512bit bus width, so in theory:
(512bit/8) X 4Ghz = 64 byte X 4Ghz = 256GB/sec

If 2 VRAM running in parallel, then it becomes
2x256GB/sec

However, the real spec of fermi is not out yet.

Another thing is, i saw many reviews that run several games and showed the frame per second capabilities of the tested card. High end cards could achieve over 100fps !! But our PC monitor mostly refresh at 75Hz or below. So, once again, do one need to pay top dollar for such cards?

I am posting this because i am not sure what i am writing above so please be kind to me.

Thanks.

Most LCD support up to 60Hz because at 60FPS, human eye will see the picture as continuous, meaning you will not feel shuttering. However, that doesn't mean you can't detect differences between 60FPS from 120FPS if the LCD actually support 120Hz. Now of course, any games above 50FPS is very smooth already and everything above 60FPS+ is only meaningful for benchmarking as most LCD can show it anyways. Yet no games can maintain a steady FPS throughout the game, as sometimes FPS will dip to say 20 FPS and you will feel shuttering, or lagging.

However, 3D scene with shuttering glasses divides FPS by half, one picture to one eye per cycle(Hz), that means FPS is cut by half. In other words, 60 FPS = 30 FPS in 3D, and people feel dizzy due to the shuttering with normal LCD, and 120Hz LCD is needed for 3D.

Now new LCD is coming out supporting up to 120Hz, allowing games to produce 120FPS. At that rate, each eye will still see 60FPS, producing smooth scene for each eye, or smooth scene in 3D. That means, 100FPS vs 60FPS is no longer a benchmark difference, but a difference that you can actually see.

As to whether or not you should spend big bucks on heavy guns depend on you. Newer games are more demanding, and high-end video cards can still play newer games with decent FPS 2-3 years down the road with most things max out. If 30-40 FPS is okay for you, then buy mid-high range video card and replace it with a new one when you can no longer maintain those FPS with new games. If you have a high resolution 120" LCD and 3D is your cup of tea, then high-end video card is what you are looking for.

 

[Schmide]

GT300 was said to have 512 bit interface, that means 512bit bus width, so in theory:
(512bit/8) X 4Ghz = 64 byte X 4Ghz = 256GB/sec

Caveat. GT300 has an ecc memory controller, so approximately 20% of the address lines would be allocated if they use hamming correction. Rounding off to the nearest byte boundary.

(392bit/8) X 4Ghz = 49 byte X 4Ghz = 196GB/sec (still cache like speed)

 

[formulav8]

Like others have said it basically the absolute best number you can get without any overhead or other influence involved. But with everything there is always some level of overhead that limits it peak potential.

Looks at sata2 for example. They have a peak bandwidth of 3gbps. But even majorly fast drives like ssd's cannot even hit that 3mbps mark. They max out to around 250bmps at best. So you will most likely never hit peak bandwidth for any device.


Jason

 

[lenkiatleong]

Thank you gentlemen for sharing your knowledge.

Regarding peak memory/bus bandwidth, we have the below affecting the performance:
- cpu to ram (limited by ram spec)
- cpu to pci express (8GB/s, in my case)
- gpu to vram (2x256GB/s)
I understand that application could direct the cpu to transfer data from ram to vram via pci card and the bottleneck is the ram and pci bus speed whichever is lower. Its clear from your explaination that one would not reach the peak bandwidth. So, in order to achieve >120fps for gaming, how do i know how much bandwidth is needed for the ram and pci which are the bottleneck? Is there a way to calculate this? I suspect that one do not need super high bandwidth like 2x256GB/s right? I reckon that the gpu is waiting for data to arrive from cpu >80% of the time. So what is the "hype" about this high bandwidth as advertised by high end graphic card?

Regarding the monitor refresh rate, i now understand why graphic card offers >100fps capability. If one want to enjoy 3D gaming, one should buy monitor with at least 120Hz refresh rate and graphic card that could render at least 120fps for eyes comfort. However for 3D movies like Avatar, i am not sure how many fps the bluray disc could render. I suspect that its much below gaming software. So, i reckon its not a problem.