Our GPU pool: VRAM utilization and headroom on 4x RTX PRO 6000

Our GPU pool: VRAM utilization and headroom on 4x RTX PRO 6000 reports operational numbers measured directly on our ai-server (Hax) stack, and — instead of dumping figures — explains what each number actually means for a real decision. Even if you are new to running local AI, this single post should let you grasp 'what do I decide when I see this number' in about five minutes.

What these numbers mean#

카드당 95.6GB 중 최대 84.8GB가 상주해 사용률 88.7%, 최저 여유는 10.2GB다. 여유 10.2GB는 추가 LoRA나 소형 모델 1개를 더 로드할 수 있는 실제 헤드룸을 의미한다 — 풀 증설 전에 이 여유부터 소진하는 게 비용 효율적이다.

How we measured it (reproducible conditions)#

These are not vendor specs or marketing figures; they are values we measured ourselves under the conditions below. We list the conditions because when the conditions change, the numbers change too — cold start versus warmed up, batch size, and the exact hardware all shift the result for the same model. So we state reproducible conditions (measured 2026-07-04):

• bench_harness.probe_comfy_gpus (bc_comfy_gpus 실측)

Rather than memorizing a single number, understand it together with these conditions, so you can diagnose for yourself why your own environment produces a different value.

How to use this in practice#

The derived judgment above translates straight into an operating decision. The point is not to memorize raw figures but to read the relationships between them — a ratio of two values, a utilization rate, a cross-check — because those relationships tell you what to scale up and what to conserve. We use this to check existing headroom before buying new hardware, and to split workflows into a fast path and a quality path. The same logic applies directly to your own local AI setup.

Why this beats vendor specs#

These are values measured in our own operating environment, not vendor sheets or someone else's benchmark. Every number above is measured (not estimated), with date and source (Hax /data). Unlike generic AI-written prose, this derived judgment cannot be produced without the measurement. Only our own measured values are used; no private tokens or internal paths are exposed.

Note: the values above are our own stack measurements as of 2026-07-04 and are refreshed when conditions change (measured values only, no estimates).

Related reading: comfy-pool Z-Image Turbo 응답 지연 p50·p95 실측, Gemma 4 MoE 가정용 GPU 업그레이드 시점: 지연과 처리량 관측을 통해 판단하는 방법

Full guide: 노트북에서 돌리는 AI 모델, 흔한 함정과 해결법

References#

• Hax data