Locally Explains

RX 9060 XT Runs 27B AI For $270 Less Than Nvidia

Every figure the video puts on screen, with the page it came from. Checked on 15 October 2026. Prices move daily; the dated listings are in PRICES.md.

Prices (15 October 2026)

• RX 9060 XT 16GB, $519.99 new (Sapphire PULSE, the cheapest 16GB listing at Newegg; the same price at Best Buy). https://www.newegg.com/sapphire-tech-pulse-11350-03-20g-radeon-rx-9060-xt-16gb-graphics-card-double-fans/p/N82E16814202457
• RTX 5060 Ti 16GB, $789.99 new (MSI Ventus 2X OC, the cheapest 16GB listing at Newegg). https://www.newegg.com/p/pl?d=rtx+5060+ti+16gb&Order=1
• The gap: $270 (789.99 − 519.99). Per gigabyte: $32.50 for the Radeon, $49.37 for the Nvidia.
• Intel Arc A770 16GB, used, $310 market average (40 listings; asking range $299 to $321). https://resaleprices.com/gpu/intel-arc-a770-16gb
• RTX 4060 Ti 16GB, used, $639 market average (17 listings; range $620 to $680). https://resaleprices.com/gpu/nvidia-rtx-4060-ti-16gb
• RTX 3090, used, $1,649 going rate (median asking price; 77 active listings). https://rigprice.com/gpu/rtx-3090/

Memory and bandwidth

• Qwen 3.8 27B at Q4_K Small is 16.68 GiB (27.32 billion parameters), already over 16 GB before any cache; Q4_K_M is larger still. https://www.hardware-corner.net/qwen3-8-27b-hardware-tests/
• Qwen3.8-27B-UD-IQ4_XS.gguf is 14.25 GB (UD-IQ3_XXS 10.9 GB). https://github.com/L4TN/qwen3.8-27B-MTP-rtx5060ti/blob/master/docs/tutorial.md
• Memory bandwidth: RX 9060 XT up to 320 GB/s (AMD specifications); RTX 5060 Ti 448 GB/s and RTX 4060 Ti 16GB 288 GB/s (128 bit GDDR7 at 28 Gbps and GDDR6 at 18 Gbps; NVIDIA’s spec pages give the bus width, the figure is the product); Arc A770 16GB 560 GB/s, 256 bit (https://www.intel.com/content/www/us/en/products/sku/229151/intel-arc-a770-graphics-16gb/specifications.html); RTX 3090 936 GB/s, 24 GB.

Published Qwen 3.8 27B speeds

• RTX 5060 Ti, 29.2 tok/s generation, UD-IQ3_XXS, q4_0 KV cache, 4,096 context; no prompt processing figure on that row. https://llamaperf.com/gpu/rtx-5060-ti-16gb
• RTX 5060 Ti tuned: 92.73 prompt tok/s and 52.68 generation tok/s at 32K context, UD-IQ4_XS with multi token prediction (draft), Q8_0 KV cache, on its chat workload. https://github.com/L4TN/qwen3.8-27B-MTP-rtx5060ti/blob/master/docs/tutorial.md
• RX 9060 XT: about 21.4 tok/s generation without MTP, prompt processing ~74.5 to ~148.9 tok/s, UD-IQ3_XXS, llama.cpp with Vulkan. A second user in the same thread reports 23 tok/s on the same file, about 32 with MTP. https://www.reddit.com/r/LocalLLM/comments/1welwcx/qwen3827b_on_rx_9060_xt_16gb/
• RX 9060 XT, Qwen 2.5 Coder 14B Q4_K_M: Vulkan 33.51 tok/s, ROCm 30.79 tok/s generation; prompt (pp512) 1,210 and 1,211 tok/s. https://github.com/parsapp/rx9060xt-llm-benchmarks
• RTX 4060 Ti 16GB: 10.7 and 17.6 tok/s, UD-Q3_K_XL at 32K context, the faster with draft MTP; prefill 289 to 291 tok/s. https://llamabench.ai/browse/rtx4060ti16gb/qwen3-8-27b?quant=Q3
• Arc A770: about 9 tok/s, IQ4_XS on Vulkan; the poster reports a SYCL backend build error and gives no prompt processing figure (they later mention 10 to 11 tok/s after changing settings). https://www.reddit.com/r/IntelArc/comments/1wdlyt6/i_want_to_use_qwen38_27b_a_bit_faster_on_the_a770/
• RTX 3090: 39.7 to 40.0 tok/s at Q4_K_M. https://chinkeong.github.io/qwen-27b/index.html

Corrections made to the supplied script before recording

The supplied script said $789 “at Best Buy” from an October 8 snapshot, an A770 bandwidth of 512 GB/s (the 8GB card’s figure), a tuned 5060 Ti run of 36 prompt and 70 generation tok/s (those are the test’s token counts, not its speeds), Radeon prompt processing of 300 to 500 tok/s (another user’s MTP build), a used RTX 4060 Ti at $330 (a mixed 8GB/16GB figure) and a used RTX 3090 at $1,700. Each was corrected to the figures above with the fewest changed words.

Not checked

• The video compares results from different quantisations, contexts, backends and MTP settings; no single controlled test ranks the four cards.
• Used prices are trackers’ averages of asking prices, not completed sales.
• The RTX 5060 Ti and RTX 4060 Ti bandwidth figures are derived from bus width and memory speed; NVIDIA’s own pages list only the bus width.