User compares modded RTX 4090, Radeon AI Pro R9700, and Arc Pro B70 for local coding LLMs

A user is seeking advice on selecting a GPU for running local coding large language models, weighing three specific hardware options. The comparison focuses on a modded RTX 4090 with 48GB VRAM, dual Radeon AI Pro R9700 cards, and dual Intel Arc Pro B70 cards.

• Modded RTX 4090 48GB: Offers full AD102 architecture, ~1TB/s bandwidth, and CUDA support for $3,500, but carries risks regarding third-party firmware, reliability, and lack of warranty.
• 2x Radeon AI Pro R9700 32GB: Features RDNA4 architecture, 640 GB/s bandwidth, PCIe 5.0, and official warranty for ~$1,300, though ROCm compatibility is noted as less mature than CUDA.
• 2x Intel Arc Pro B70 32GB: Priced at ~$1,080 with Battlemage architecture, 608 GB/s bandwidth, and 367 TOPS INT8, but lacks FP4 support and faces driver maturity questions for oneAPI/OpenVINO.

The user aims to achieve 30-40 tokens per second (tps) for inference and light fine-tuning, noting that their current DGX Spark setup is limited to 20 tps.