Locally Explains

Two Cheap GPUs Have a Local AI Catch Nobody Mentions

Two Budget GPUs vs One for Local AI: sources

Every figure this video puts on screen, with the primary source it comes from. Pages were checked on 30 September 2026.

The cards

GeForce RTX 5060 Ti 16GB. NVIDIA’s announcement says RTX 5060 Ti cards with 16GB or 8GB of graphics memory “will be available starting April 16 at $429 and $379, respectively”. NVIDIA’s RTX 5060 family specification table lists the RTX 5060 Ti at 16 GB or 8 GB GDDR7 and a Total Graphics Power of 180 W.

• https://nvidianews.nvidia.com/news/nvidia-blackwell-geforce-rtx-arrives-for-every-gamer-starting-at-299 (15 April 2025)
• https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5060-family/

Two cards at that launch price are 2 × $429 = $858, before tax. Two cards rated at 180 W are 2 × 180 W = 360 W for the cards alone, under the published rating.

GeForce RTX 5070 Ti and RTX 5080. NVIDIA’s CES announcement: the RTX 5080 available January 30 “at $999”, the RTX 5070 Ti available “starting in February at $749”. Both are 16 GB GDDR7 cards on NVIDIA’s own specification pages.

• https://nvidianews.nvidia.com/news/nvidia-blackwell-geforce-rtx-50-series-opens-new-world-of-ai-computer-graphics (6 January 2025)
• https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5070-family/
• https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5080/

GeForce RTX 4090. 24 GB GDDR6X, Total Graphics Power 450 W, on NVIDIA’s RTX 4090 page. Announced 20 September 2022, available 12 October 2022.

• https://www.nvidia.com/en-us/geforce/graphics-cards/40-series/rtx-4090/
• https://nvidianews.nvidia.com/news/nvidia-ada-lovelace-architecture-provides-quantum-leap-in-performance

GeForce RTX 3090. NVIDIA’s RTX 3090 family page: the 3090 Ti and 3090 carry “a staggering 24 GB of G6X memory”. Announced 1 September 2020, available 24 September 2020.

• https://www.nvidia.com/en-us/geforce/graphics-cards/30-series/rtx-3090-3090ti/
• https://nvidianews.nvidia.com/news/nvidia-delivers-greatest-ever-generational-leap-in-performance-with-geforce-rtx-30-series-gpus

AMD Radeon RX 9060 XT (16GB). AMD’s product page lists up to 16GB of GDDR6 memory.

• https://www.amd.com/en/products/graphics/desktops/radeon/9000-series/amd-radeon-rx-9060xt.html

The model

Qwen3.8-27B, Q4_K_M, 19 GB. The ggml-org GGUF repository lists the Q4_K_M file at 19 GB in its hardware compatibility table, beside Q8_0 at 28.6 GB and BF16 at 53.8 GB.

• https://huggingface.co/ggml-org/Qwen3.8-27B-GGUF

19 GB is 3 GB more than a 16 GB card holds before any context cache or runtime memory, and splits into two halves of 9.5 GB across a pair.

Splitting one model across two cards

llama.cpp’s multi GPU guide (docs/multi-gpu.md):

• layer is the default split mode: “Pipeline parallelism. Each GPU holds a contiguous slice of layers. The KV cache for layer l lives on the GPU that owns layer l.”
• “Pipeline-parallel runs different layers on different GPUs and processes tokens sequentially through the pipeline.”
• tensor is marked “EXPERIMENTAL”: tensor parallelism that splits both weights and KV across the participating GPUs; “Treat as experimental as the code is less mature than pipeline parallelism.”
• --split-mode tensor is not implemented for all architectures; the guide lists families that fail with it.
• Troubleshooting, “Performance is worse with multi-GPU than single-GPU”: “The performance is bottlenecked by GPU interconnect speed.”
• -ngl / --n-gpu-layers: “Maximum number of layers to keep in VRAM.”
• https://github.com/ggml-org/llama.cpp/blob/master/docs/multi-gpu.md

llama.cpp README. Supported backends include CUDA (NVIDIA GPU) and HIP (AMD GPU); the description lists “Custom CUDA kernels for running LLMs on NVIDIA GPUs (support for AMD GPUs via HIP …)” and “CPU+GPU hybrid inference to partially accelerate models larger than the total VRAM capacity”.

• https://github.com/ggml-org/llama.cpp

Not established

• Used RTX 3090 prices vary with condition and market; no price is given.
• Context cache and runtime memory depend on the context length, runner and settings; no single figure is given for them.