MEMORY
Bandwidth is the pipe
Peak memory bandwidth comes from the memory clock, the bus width and the data rate. GPU work is memory-intensive, so this pipe often sets the pace.
Follow the dataHOW TO READ A GPU SPEC SHEET
Read the numbers,
not the hype.
Every graphics card comes with a wall of numbers.
A few of them do most of the work: how much memory the chip has, how fast it can move data to and from that memory, how many calculations it can attempt, and how much heat it is designed to shed.
Follow the data Read the short answer
01 Bandwidth feeds the cores02 TFLOPS is an estimate03 TDP sizes the coolingSpecs are a starting point
THE SHORT ANSWER
Three numbers
tell most of the story.
Memory, compute and power.
Each has a simple idea behind it.
COMPUTE
TFLOPS is a ceiling
GPU performance is usually quoted in teraflops. It is an estimated measure, worked out from cores, clock speed and operations per cycle, and other factors affect real performance.
Work it out
POWER
TDP is for the cooler
Thermal design power is the heat a cooling system is designed to remove in normal use. GPUs are known for peak power well above their TDP.
See the numbersA QUICK DATA-PATH CHECK
Where does data
move fastest?
Pick a route that data takes in a computer with a GPU. The value is a peak figure from the sources; the bar shows how fast the route is.
Theoretical peak figures. Real transfers are slower, and newer hardware can be faster.
TWO NUMBERS WORTH KNOWING
A wide road on the card,
a narrow one to it.
Both figures come from one worked example
in a GPU programming guide.
ON THE CARD · CUDA BEST PRACTICES GUIDE
Peak bandwidth, 4096-bit HBM2 at 877 MHz
898<small> GB/s</small>
Memory clock in hertz × bus width in bytes (4096 ÷ 8) × 2 for double data rate, divided by 10⁹, gives 898 GB/s for the GPU in the guide's example.
TO THE CARD · CUDA BEST PRACTICES GUIDE
Peak bandwidth over PCIe x16 Gen3
16<small> GB/s</small>
Moving data between the computer's main memory and the GPU is far slower than the card's own memory, so the guide advises keeping transfers to a minimum.
The same guide separates theoretical bandwidth, worked out from the spec sheet, from effective bandwidth, measured by timing how many bytes a program actually reads and writes.
FOUR STEPS WITH ANY SPEC SHEET
Make sense of
the numbers.
You do not need to know the chip design. These four checks work for any graphics card's published specifications.
- Find the memory type and bus widthNote the memory clock, the interface width in bits and whether the memory is double data rate.
- Work out the bandwidthClock in Hz × (bus width ÷ 8) × 2 for double data rate, then divide by 10⁹ for GB/s.
- Treat TFLOPS as a ceilingIt is cores × clock × operations per cycle: a theoretical maximum, not a measured speed.
- Check TDP against your coolingTDP tells you how much heat the cooler must handle. Expect peaks above it.
WORKING OUT TFLOPS
Cores, clocks
and operations.
FLOPS, floating point operations per second, is a standard measure of computing performance. For a single processor it simplifies to cores × cycles per second × floating point operations per cycle.
Modern GPUs contain hundreds or thousands of calculation units working in parallel, which is why their totals reach teraflops (10¹² operations per second). But clock speeds, memory and software all shape what you actually get.
QUESTIONS PEOPLE ASK
Frequently asked
What is VRAM?
Video random access memory is the dedicated memory on a discrete graphics card. Integrated graphics have little or none and use a share of the computer's system RAM instead.
How do I calculate memory bandwidth?
Multiply the memory clock in hertz by the bus width in bytes (bits ÷ 8), multiply by 2 if the memory is double data rate, then divide by 10⁹ to get GB/s.
Are TFLOPS figures accurate?
They are theoretical. GPU performance quoted in TFLOPS is an estimated measure, and other factors affect actual performance.
What does TDP mean on a GPU?
Thermal design power is the maximum heat a component can generate that its cooling system is designed to dissipate in normal operation. GPUs are known to have large gaps between peak power and TDP.
Why does PCIe speed matter less than VRAM bandwidth?
Data moves far faster between the GPU and its own memory than across the PCIe link from the computer's main memory, so well-written programs keep data on the card and minimise transfers.
SOURCES
Where the facts come from
Each figure on this page comes from the sources below. Check them for the full detail and context.