We've established that the large number of parameters in an LLM needs to fit into the GPU's memory (VRAM). However, just having enough VRAM capacity isn't the whole story. The speed at which data can move between the VRAM and the GPU's processing cores is also extremely important. This speed is called memory bandwidth.

Think of VRAM as a large warehouse (its capacity measured in Gigabytes, GB) and memory bandwidth as the width of the road leading to it (measured in Gigabytes per second, GB/s). If you have a huge warehouse but only a narrow, single-lane road, you can't move goods in and out very quickly, even if your workers inside (the GPU compute cores) are very fast. Similarly, if the memory bandwidth is low, the GPU cores might spend a lot of time waiting for parameters and other data to arrive from VRAM, slowing down the entire process of generating text.

Why is Bandwidth Significant for LLMs?

Running an LLM, especially for inference (generating text), involves a constant back-and-forth of data:

The GPU needs to fetch the model's parameters (weights) from VRAM. Remember, there can be billions of these.
It performs calculations using these parameters and the input data.
It often needs to write intermediate results (like activations, which represent the state of neurons during computation) back to VRAM and read them again for subsequent steps in generating the output.

Large models mean a lot of data needs to be moved constantly. Modern GPUs have incredibly powerful processing cores capable of performing trillions of calculations per second (FLOPS). But these powerful cores are ineffective if they are starved for data.

If the memory bandwidth is low (the road is narrow), the GPU cores can't get the parameters or intermediate data fast enough. They end up idle, waiting for the data transfer to complete. This means the overall speed at which the LLM generates text (often measured in tokens per second) is limited not by the raw calculation power of the GPU, but by how quickly data can be fed to it. This situation is often described as the process being memory-bound.

Bandwidth as a Potential Bottleneck

Consider two hypothetical GPUs:

GPU A: Has 16 GB of VRAM and very high compute power, but relatively low memory bandwidth (e.g., 400 GB/s).
GPU B: Also has 16 GB of VRAM and slightly lower compute power, but much higher memory bandwidth (e.g., 800 GB/s).

For running a large LLM, which requires accessing billions of parameters frequently, GPU B might actually generate text faster than GPU A. This happens because its high bandwidth keeps the processing cores supplied with data more effectively, minimizing idle time, even though its peak calculation speed might be lower.

Low memory bandwidth can create a bottleneck, slowing down LLM inference even on GPUs with high computational power. Higher bandwidth allows faster data transfer between VRAM and compute units, enabling more efficient processing and faster output generation.

Different types of GPU memory technologies contribute to these differences in bandwidth. For example, consumer GPUs often use GDDR6 memory, while high-end data center GPUs frequently use HBM (High Bandwidth Memory). HBM is specifically designed to offer significantly higher bandwidth, which is one reason these GPUs are preferred (and more expensive) for training and running the largest AI models.

Summary: Capacity and Speed Both Matter

When evaluating hardware for running LLMs, VRAM size (capacity) tells you if a model can fit, but memory bandwidth (speed) strongly influences how fast it will run. For large language models that constantly shuttle vast amounts of parameter data, higher memory bandwidth often translates directly to better performance, measured in faster response times or more tokens generated per second. Both factors are significant considerations when selecting a GPU for your LLM needs.