Why memory bandwidth beats core count for Malairte GPU mining

A breakdown of why the headline shader count on a GPU spec sheet matters far less than its memory bandwidth for the kind of work Malairte mining actually does on the card.

If you only read the front of a GPU box, you would expect the card with the most CUDA cores or stream processors to mine best. In practice, that is rarely how it shakes out. Malairte mining is far more sensitive to memory bandwidth and capacity than to raw shader horsepower.

What the workload actually does

Mining is a memory-heavy task. The hash work involves shuffling data through GPU memory at high rates, and the speed at which that data moves between the memory chips and the GPU core ends up gating overall performance. Once the cores are fed, adding more cores does not help if memory cannot keep up.

The bandwidth number that matters

Look for memory bandwidth in GB/s on the GPU spec sheet. It is the product of memory speed and bus width. A 192-bit bus running GDDR6 at 15Gbps gives you roughly 360 GB/s. A 256-bit bus at 18Gbps gives roughly 576 GB/s. That second number is closer to what wins on memory-bound work.

Why the RTX 3060 12GB punches up

The RTX 3060 12GB has a 192-bit bus and GDDR6 memory. On paper it should be soundly beaten by cards with more cores. In memory-heavy mining it is competitive far above its class, because its memory subsystem is generous relative to its core count, and the cores rarely sit idle waiting for data.

Where wider buses appear

RTX 3070 Ti: 256-bit, GDDR6X
RTX 3080: 320-bit, GDDR6X
RX 6800: 256-bit, GDDR6 with Infinity Cache
RX 6700 XT: 192-bit, GDDR6 with Infinity Cache

The Infinity Cache twist

AMD's RDNA2 cards (RX 6000 series) include a large on-die cache called Infinity Cache, which acts as a memory bandwidth multiplier for workloads that fit within it. For mining, this can make a 192-bit RX 6700 XT punch closer to a 256-bit NVIDIA card than its raw GDDR6 bandwidth suggests.

What to take away

When you are comparing two GPUs for Malairte mining and one has more shaders but the other has more memory bandwidth, lean toward bandwidth. Add VRAM capacity and board power into the picture too. Core count is rarely the deciding factor on cards in the same generation, and it is almost never the right thing to optimise for on a home mining build.