๊ฐ€์ž… ํ›„ ์ดˆ๋Œ€ ๋งํฌ๋ฅผ ๊ณต์œ ํ•˜๋ฉด ๋™์˜์ƒ ์žฌ์ƒ ๋ฐ ์ดˆ๋Œ€ ๋ณด์ƒ์„ ๋ฐ›์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

David Hendrickson
@TeksEdge
CEO & Founder | PhD | Startup Advisor | @Columbia | Author Generative Software Engineering | ๐Ÿ”” Follow for AI & Vibe Coding Tips ๐Ÿ‘‡
๊ฐ€์ž… July 2023
549 ํŒ”๋กœ์ž‰ ์ค‘    11.2K ํŒฌ
๐Ÿคฏ This entire 64GB AMD AI workstation cost its builder LESS than a single RTX 5090 and Qwen3.8-27B is hitting 111 tok/s. This is the kind of rig that makes me want to build it because RTX 5090s are so expensive. Setup ๐Ÿ‘‡ ๐ŸŸ  2ร— Radeon AI PRO R9700 32GB ๐Ÿ’พ 64GB total dedicated VRAM ๐Ÿง  Ryzen 7500F ๐Ÿ’พ 64GB DDR5 ๐Ÿ”Œ PCIe 5.0 x8 per GPU โšก vLLM Radiance + TP2 + MTP Running Qwen3.8-27B Quark AWQ MXFP4 ... ๐Ÿš€ 111.4 tok/s median decode โšก 4,224โ€“4,410 tok/s prefill โฑ๏ธ 81ms TTFT ๐Ÿ“š 131K server context Native FP8 still managed 87.6 tps ๐ŸŽฎ RTX 5090 โ†’ 32GB VRAM ๐ŸŸ  2ร— R9700 โ†’ 64GB VRAM The builder says the entire machine cost ~โ‚ฌ4,000, more than โ‚ฌ1,000 less than a single RTX 5090 in his market. So for less money, he ended up with 2ร— the VRAM + 111 tok/s on Qwen3.8-27B. ๐Ÿ”ฅ Reddit Link in ALT
๋” ๋ณด๊ธฐ