My small department of questionable decisions
A fanless laptop, three rather warm 3090s, and two 3060s that still have work to do. This is the hardware behind my experiments.
I link a machine to a test when the original record supports it. Each post explains what ran where, how many cards it used, and which conditions matter. Owning three GPUs doesn’t make a one-GPU benchmark three times faster. The arithmetic is annoyingly firm on that.
The portable one
Metal, MLX and on-device model work
My fanless MacBook Air. The CPU and GPU share one 24 GB memory pool, so model size and the rest of the desktop have to get along.
Unified memory is shared by the CPU and GPU; it is not a separate 24 GB VRAM allocation.
The big rig
CUDA inference, quantization and multi-GPU experiments
Three RTX 3090s give me room to test larger models and split workloads. A single-card result still means one card, not the whole tower taking a victory lap.
72 GB installed across three GPUs. Each card has its own 24 GB pool; pooling it requires a workload that supports splitting.
The old faithful
Local language models on a 12 GB budget
An i7-4790 with a 12 GB RTX 3060. Older CPU, useful GPU, and a very practical way to find out whether a model earns its memory bill.
The inbox benchmark shared this GPU with a resident embedder. That occupied VRAM is part of the recorded setup.
The Windows one
Windows CUDA builds and local-model comparisons
An i7-7700 and another 12 GB RTX 3060, this time on Windows. Same GPU family, different runtime conditions: a useful reminder to write the setup down.
48 GB installed across four DIMMs. The historical inbox run recorded no competing GPU load on this machine.