Every time there is a new graphics card release, a rice, n rice heated debate is very detailed content, in addition to video card game performance, power, temperature and other key indicators, memory bit width is one of the controversial points, high-end graphics card with 384bit bit width or even 512bit bit width no problem, Again a bit of the video card with 256bit bit width is not a problem, but more than thousand of graphics card if the use of 128bit bit width will often be spit groove. Today we'll talk about the graphics card's memory bandwidth and how the memory bandwidth is calculated.
Images from the Web
If memory bandwidth is the key to the processor and memory exchange data, then the video memory bandwidth is also the same--GPU core responsible for the operation, video memory is responsible for data storage, the two need to exchange data frequently, it depends on the memory bandwidth, higher bandwidth can make the graphics card in the processing of high resolution, high image quality more handy.
The role of memory bandwidth, we can understand--like Shun Fung Biaoju Logistics transport, Shun Fung to make more money, orders, send pieces of the process will be shortened. Shun Fung that end equivalent to the GPU core, the receipt/delivery of buyers/sellers equivalent to video memory (on the other hand), shun Fung to improve efficiency or to make The courier faster (more frequent), or send a large number of courier to collect or send pieces (bit wider), the best case of course is two ways to operate simultaneously, Logistics transportation efficiency can be maximized.
Increase the speed of transport, can increase the speed of the vehicle or even use aircraft, but also can be multi-way
The determining factor of memory bandwidth: Memory type
Memory bandwidth from a large aspect is the memory frequency and memory bit width to determine, but the actual bandwidth depends on the specific situation, the current mainstream graphics card bit width is 128bit, 256bi, 384bit and 512bit, more can determine the bandwidth or memory type, they determine the limit of memory bandwidth. At present, the most mainstream video memory is of course GDDR5, before there is a fleeting GDDR4, and now the low-end market is still gDDR3 memory remnants, and AMD recently on the R9 Fury series graphics card using HBM memory, compared to GDDR5 memory more powerful, bandwidth greatly increased, Next year's HBM2 bandwidth will continue to double.
Here alone GDDR5 memory and HBM video memory, before the Fury X related analysis and evaluation in fact we talk about a lot of HBM's advantages, the current GDDR5 memory frequency has reached 7Gbps or even 8Gbps, and then the increase in frequency will bring many side effects, so AMD, SK Hynix and other companies to re-start, no longer pursue the frequency to increase the internal memory width to improve the total bandwidth.
The difference between HBM's memory and GDDR5 memory
In both case, the GDDR5 internal I/O bandwidth is 32bit, the current NVIDIA graphics card GDDR5 memory frequency can reach 1750MHz (a, in 1500MHz), it is 4 times times the rate mechanism, the data frequency is 7Gbps, the bandwidth of a single chip is 28gb/s. The current HBM video memory frequency only 500mhz,2 multiplier rate mechanism, the data frequency is 1Gbps, but its high I/O bandwidth, to compensate for the lack of frequency.
In particular, the current HBM video memory is stack-designed, each stack has 8 channels, each channel provides 128bit bit width, so the total I/O bit width is 128bit*8=1024bit. This is still a single stack, Fury x graphics is currently 4 stacks, so the equivalent bit width is four times times 4096bit, Nvidia next year's Pascal architecture GPU will also use the 4-hi stack, but it should be HBM2 memory, the frequency is higher.
In summary, the previous, present, and future memory is simply summed up as follows:
The main specifications and features of different memory
Said so much, in fact, the current gDDR3 video memory is basically NVIDIA and AMD part of the low-end graphics card in use, this memory can be said to be the store JS favorite, if it is 600 yuan in the graphics card, JS with 4GB super-large memory to attract you, then you have to be careful.
GDDR5 There is nothing to say, this is the absolute mainstream, the capacity of a single chip gradually increased from the previous 2Gb to 4Gb, before the United States of light production of 8Gb (1GB) particles, high-end graphics card also as long as 4-8 chips can realize 4-8GB capacity memory, This will further drive the advent of high-capacity graphics.
HBM is a rising star, currently the only AMD family, they use the first generation of HBM, the stack of memory core capacity of 2Gb (a stack is 4 cores), data frequency 1Gbps, bit width 1024bit.
Memory Bandwidth calculation method
After understanding the characteristics of different memory types, it is now easy to calculate the bandwidth of memory, the general formula can be summed up as
Memory bandwidth (GB/s) = Video memory actual frequency (MHZ) * Memory data magnification (1, 2, 4) * Memory equivalent bit width (64-512bit unequal)/8
Because the graphics factory is more accustomed to using digital larger and better-looking data frequency to mark product specifications, the above companies can actually be more simple, directly into:
Memory bandwidth (GB/s) = Memory data Frequency (Gbps) * Memory equivalent bit width (bit)/8
Find a few more specific examples to calculate will be more direct, first find an NVIDIA GeForce GT 720 graphics card to cushion back, this card bit width is only 64bit, while supporting gDDR3 and GDDR5 memory, the former typical frequency 900MHz, the latter typical frequency is 1250MHz, The bandwidth of the two configurations is:
GDDR3:GT 720 graphics card has a bandwidth of 900mhz*2*64bit/8=14.4gb/s, or 1.8gbps*64bit/8=14.4gb/s.
GDDR5:GT 720 graphics card has a bandwidth of 1250mhz*4*64bit/8=40gb/s, or 5gbps*64bit/8=40gb/s.
HBM memory is currently available only with AMD's Fury series graphics card, with fury X graphics as an example, with bandwidth calculated as follows:
Hbm:fury x Graphics Bandwidth is 500mhz*2*1024bit*4/8=512gb/s, or 1gbps*4096bit/8=512gb/s.
The above is the main content of today's super class, it looks very complex, after mastering the law is very simple.
In addition, this article mainly introduces the memory width of the calculation method, in fact, the topic of video memory is very large, such as said here gDDR3 memory, in fact, there are DDR3, gDDR3 and GDDR3 a variety of statements, they are not the case or the abbreviation so simple, GDDR3 and GDDR3 are actually different types of memory, and later, we can continue to interpret the type of video memory.