OPEN_SOURCE ↗
REDDIT · REDDIT// 19d agoINFRASTRUCTURE
Dual NVIDIA RTX A6000s Skip Threadripper
The poster wants to run 70B-class or 24B-class coding models locally for up to five developers, so the real question is whether a mainstream AM5 build can handle dual A6000s or if Threadripper is actually worth it.
// ANALYSIS
This is a bandwidth-and-memory problem, not a "buy the biggest CPU" problem. Two A6000s already give you real VRAM headroom, so the smarter spend is a lane-clean motherboard, enough host RAM, and good cooling.
- –NVIDIA’s RTX A6000 is a 48GB PCIe 4.0 x16 card, and two of them can be NVLink-bridged into 96GB of combined GPU memory; that is useful for oversized models, but it is optional and the bridge is sold separately.
- –A mainstream AM5 chip like the Ryzen 9 9950X already exposes 24 usable PCIe lanes, and boards such as ASUS ProArt X670E-Creator WiFi support two PCIe 5.0 x16 slots in x8/x8 dual mode, which is enough lane layout for a dual-GPU inference box.
- –64GB of system RAM is the sane floor; 32GB is the "it boots" tier, not the "team box with concurrent sessions" tier.
- –Tensor-parallel sharding means the model weights can be split across both GPUs, so the whole model does not need to live in host RAM. That makes 24B-class code models comfortable and 70B-class models plausible, though long context and batching will still eat through headroom fast.
- –Threadripper only becomes worth it if you want 128-lane headroom, lots of NVMe, or a more server-like expansion plan; otherwise, spend the delta on a beefy PSU and airflow.
// TAGS
gpuinferenceself-hostedllmai-codingnvidia-rtx-a6000
DISCOVERED
19d ago
2026-03-24
PUBLISHED
19d ago
2026-03-24
RELEVANCE
8/ 10
AUTHOR
ackermann