Your GPU reports 24GB. Only 7.9GB of it loads a model, and half of that is already gone

On an RTX 5060 laptop sold with 8GB of VRAM, dxdiag reports the graphics card as having 24,144 MB of display memory ^[1][2]. Nothing is broken: the figure is arithmetically honest, and it is also not the number that decides whether a model loads ^[3][5].

Four lines down the same Display tab, the panel breaks it out: 7,899 MB dedicated, 16,245 MB shared, 24,144 MB total ^[4]. The shared pool is a slice of system RAM the driver treats as overflow, half the machine's 31.7GB, which is the standard Windows allocation ^[5]. Only the dedicated figure governs whether a model fits in fast memory ^[3]. Taking the headline number at face value overstates usable capacity by roughly 3.1x on this machine ^[4].

It gets worse if you add adapters up. Every display device on the box claims the same 16,245 MB: integrated Intel graphics, the RTX 5060, and two DisplayLink devices ^[6]. Sum what the machine appears to have and you get 64,980 MB of shared memory that exists exactly once ^[1]. The dedicated figure is not perfectly stable either. A model-fitting tool on the same box reported 7.96GB against dxdiag's 7,899 MB ^[7], a gap of about 61 MB ^[5], close enough to be the same answer and far enough apart that a tight fit calculation will disagree with itself depending on which tool it asked ^[7].

The two pools behave differently under load, which is the part that costs you. Spilling into shared memory does not fail, it slows down ^[8]. A 9B model at 32K context sits entirely in VRAM at 5.9GB and generates 51.5 tokens/sec; push it to 64K, it needs 7.5GB, splits 18/82, and drops to 22.6 tokens/sec ^[9][10]. Same model, same machine, same quantisation, one setting, and throughput falls to about 44% of what it was ^[2]. A failure tells you something is wrong. A slowdown lets you keep going, wondering why everything feels off ^[8].

Then subtract the desktop. Under ordinary use, browser and editor open, between 2 and 3.5GB of the dedicated 7.9GB is spoken for before a model loads, and not in a way you reclaim by closing one app ^[11]. That leaves an effective budget of roughly 4.4 to 5.9GB ^[3]. The spread is the problem: measured with a browser, a notes app, a spreadsheet and around thirty other things touching the GPU it sat at almost exactly 2GB, and on a quiet afternoon it was under 1GB ^[12][13].

That is why the published benchmark in this account carries a footnote rather than a headline. A 36B mixture-of-experts model was measured three ways at 32K context, same day, flash attention on, KV cache quantised to q8_0, with ambient GPU load between 0.3 and 0.8GB against a normal desktop load of around 3.5GB ^[14][15]. The fastest row leaves about 800 MB of headroom and does not fit at all under normal desktop load ^[16]. The best number in the table cannot be reproduced while actually using the computer ^[16].

The rule that came out of it for dense models: about 7GB or smaller at 4-bit, and it has to run entirely on the GPU, because what spilling costs is more than a bigger model would have gained ^[17].

Worth watching: measure your own ambient GPU load on a busy day, not a quiet one, before trusting any fit calculator, and treat any published headroom figure that does not state the ambient load it was measured under as unreproducible.