Ryzen AI Max+ 395 Llama-3.3-70B dense Q4_K_M token-gen
5.0 tok/s decode (Vulkan AMDVLK)
Fields
id
benchmarks/comm-strix-halo-llama33-70b-dense-q4-tgMetric
decode_tok_s
Value
5
Unit
tok/s decode (Vulkan AMDVLK)
Model
Llama-3.3-70B
Methodology
lhl/strix-halo-testing llm-bench, dense 70B Q4_K_M, pp512 26.4 (Framework Desktop-class)
Measured by
third_party
Thermal state
unknown
notes
The ~250 GB/s wall is the wall - dense 70B is ~5 t/s. Do NOT buy Strix Halo for dense 70B; it is a MoE machine.
Notes
The ~250 GB/s wall is the wall - dense 70B is ~5 t/s. Do NOT buy Strix Halo for dense 70B; it is a MoE machine.