Memory Bandwidth Calculator

How is memory bandwidth calculated?

Memory bandwidth is how many bytes per second memory can deliver. Each transfer moves one bit over every wire of the memory bus, so the peak is the transfer rate times the bus width, divided by 8 to get bytes.

DDR5-6000 makes 6,000 million transfers per second (MT/s). A desktop with two 64 bit channels has a 128 bit bus, so it peaks at 6,000 × 128 ÷ 8 = 96,000 MB/s, or 96 GB/s. GDDR and HBM speeds are usually given in Gbps per pin instead, and 1 Gbps per pin is 1,000 MT/s.

It matters most for jobs that read a lot of data. A dense local AI model reads all of its weights for every token it writes, so bandwidth sets a ceiling on how fast it answers.

Tool description

Pick an Example configuration, from DDR4-3200 dual channel to the GeForce RTX 5090 and Instinct MI300X, or a Memory type from DDR3 to HBM3E, and the fields fill in. You can change the Data rate (in MT/s or Gbps per pin), the bus width per channel, chip or stack, and the number of channels, memory chips or stacks. The tool computes:

peak (bytes/s) = data rate (MT/s) × 1,000,000 × bus width (bits) × count ÷ 8
effective = peak × efficiency ÷ 100
time to read = data to read ÷ effective

You get the Peak bandwidth and Effective bandwidth in GB/s, the peak in GiB/s and Gbit/s, the Total bus width, the bandwidth per channel, chip or stack, and the Formula with your numbers. Enter Data to read in GB, such as a model size, to see the Time to read data once and Full reads per second. GB here means 1,000,000,000 bytes.

Examples

At 100% efficiency, with 16 GB to read:

Features

• 14 memory types and 12 real example configurations
• Effective bandwidth with your own efficiency figure
• Time to read a model or dataset once