AMD Ryzen AI 9 HX 370 (Strix Point, dual-channel LPDDR5X)
amd Zen 5 + RDNA 3.5 (Strix Point APU) 2024
Fields
id
accelerators/amd-ryzen-ai-9-hx-370Vendor
amd
Architecture
Zen 5 + RDNA 3.5 (Strix Point APU)
Release year
2024
Form factors
laptop-apumini-pc
Memory type
LPDDR5X-7500 dual-channel unified
Memory bus (bit)
128
Memory bandwidth (GB/s)
120
Bandwidth basis
LPDDR5X-7500 x 128-bit = ~120 GB/s theoretical peak; ~90-100 GB/s effective
TDP (W)
54
Unified memory
yes
Consumer part
yes
notes
The desktop-CPU trap for local LLM inference: a fast modern APU on a dual-channel LPDDR5X bus delivers only ~120 GB/s - ~4x less than 12-channel EPYC Genoa (460.8) and ~7x less than Mac Studio Ultra (800-819). Measured 3.5 tok/s on Qwen2.5-32B Q4 (llama.cpp discussion #19480). Soldered LPDDR5X caps capacity at 32-96 GB. Fine for 7-13B; do not buy for 32B+ dense decode - more cores do not help because decode is bandwidth-bound.
Notes
The desktop-CPU trap for local LLM inference: a fast modern APU on a dual-channel LPDDR5X bus delivers only ~120 GB/s - ~4x less than 12-channel EPYC Genoa (460.8) and ~7x less than Mac Studio Ultra (800-819). Measured 3.5 tok/s on Qwen2.5-32B Q4 (llama.cpp discussion #19480). Soldered LPDDR5X caps capacity at 32-96 GB. Fine for 7-13B; do not buy for 32B+ dense decode - more cores do not help because decode is bandwidth-bound.
Cross-references
This record references no other records — it is a leaf. It is cited by 10.