REDDIT · REDDIT// 25d agoINFRASTRUCTURE

Mac Pro 2019 RAM Trails M3 Ultra

A Reddit user is considering a 1.5TB RAM upgrade for their 2019 Intel Mac Pro to run Llama 3.1 405B locally. The community consensus highlights the significant bottleneck of DDR4 memory bandwidth compared to the unified memory architecture of newer Apple Silicon machines.

// ANALYSIS

–Memory bandwidth, not just capacity, is the critical factor for usable LLM inference; the M3 Ultra's over 800GB/s is about 5.7x the Mac Pro's peak theoretical 6-channel DDR4 bandwidth.
–Running an unquantized Llama 3.1 405B (820GB) on a Xeon CPU will likely yield less than 1 token per second, making it a "batch" processor rather than an interactive assistant.
–Quantizing the 405B model to 4-bit allows it to fit within 512GB of M3 Ultra unified memory, providing a "snappy" performance that is far more practical for most use cases.
–Transitioning to Apple Silicon is the definitive recommendation for high-end local AI work, as it consolidates memory and compute into a much more efficient package.

// TAGS

mac-pro-2019llmgpuapple-siliconmac-studioself-hostedinfrastructure

DISCOVERED

25d ago

2026-03-18

PUBLISHED

25d ago

2026-03-18

RELEVANCE

7/ 10

AUTHOR

Artifiko