
Llama 3.1 8B @ q8, 32,768 ctx, concurrency 1

  weights   8.03B params x 8 bits / 8           =    7.5 GB
  kv cache  131,072 B/tok x 32,768 tok x 1   =    4.0 GB   (GQA)
  overhead  8% of weights + 1.5 GB floor      =    2.1 GB
  --------------------------------------------------------------
  required                                            13.6 GB
  usable                                              96.0 GB
  headroom                                            82.4 GB

  decode is bandwidth-bound, and reads both:
    weights    7.5 GB/token  (8.03B active)
    kv cache   4.0 GB/token  (the whole cache, every token)
    total     11.5 GB/token at 72% of peak
  ~32 tok/s single-stream  (roofline estimate, not measured)

  verdict: FITS

