Builds / RTX 5090 + RTX 5080 + 3x RTX 5060 Ti 16GB (two hosts, llama.
RTX 5090 + RTX 5080 + 3x RTX 5060 Ti (96 GB) on two PCs via llama.cpp RPC: Qwen3.6 35B-A3B Q8_0 at 88 tok/s, gpt-oss 120B at 94, Llama 3.3 70B Q6_K at 11
Last updated:
- GPU(s)
- RTX 5090 + RTX 5080 + 3x RTX 5060 Ti 16GB (two hosts, llama.cpp RPC)
Amazon · eBay (used) · Newegg - VRAM / unified memory (GB)
- 96
- CPU
- not stated
- Motherboard
- not stated
- RAM
- not stated
- PSU
- not stated
- Other
- Host 1: RTX 5080 + 2x RTX 5060 Ti; host 2: RTX 5090 (power-limited to 450 W) + 1x RTX 5060 Ti; the two PCs are joined by a CAT6A cable at about 4.7 Gbit/s (~500 MB/s) and the model is split layer by layer with llama.cpp's RPC servers (no other optimisation); 35 C ambient, CPU 76 C, RTX 5090 peaked at 68 C. Thinking disabled
- Reported total cost
- not stated
- Power
- RTX 5090 power-limited to 450 W (as configured)
Measured speed (generation tokens/sec)
| Model | Quant | Engine | tok/s | Prompt processing (tok/s) | Details |
|---|---|---|---|---|---|
| Qwen3.6 35B-A3B | Q8_0 (GGUF, ~42 GB) | llama.cpp RPC (5 GPUs, 2 hosts) | 88 | not stated | context: short context, thinking disabled; 128K maximum usable context measured📏 Decode tok/s (short ctx) in the author's overall-comparison table, whole number as published; single runs with no voting or averaging |
| gpt-oss 120B | MXFP4 (~65 GB) | llama.cpp RPC (5 GPUs, 2 hosts) | 94 | not stated | context: short context, reasoning set to low (cannot be disabled); 128K maximum usable context measured📏 Decode tok/s (short ctx) in the author's overall-comparison table, whole number as published; single run |
| Llama 3.3 70B legacy | Q6_K (GGUF, ~73 GB) | llama.cpp RPC (5 GPUs, 2 hosts) | 11 | not stated | context: short context; 98K maximum usable context measured📏 Decode tok/s (short ctx) in the author's overall-comparison table, whole number as published; single run. The author attributes the low speed to a dense model paying the cost of splitting 80 layers over 5 GPUs plus one network hop |
dev.to post of 22 Oct 2026 about a two-PC consumer GPU cluster built on a budget. Speeds are whole numbers in the author's table, from single runs. Not listed: GLM-4.5-Air Q4_K_M (53 tok/s), Qwen2.5-72B Q6_K (10) and Qwen3-235B-A22B-2507 Q2_K_XL (31, below Q4), which are not current RigBase models or are sub-Q4. The author notes the Q8_0 35B-A3B model fits on one machine, so the cluster mainly shows what RPC splitting costs.
Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.
Copy into a README or blog. Links back to this build. See badge docs (CC BY 4.0 attribution). The SVG does not invent speeds — numbers stay on the build page.
Markdown
[](https://getrigbase.com/builds/rtx-5090-5080-3x-5060-ti-96gb-llamacpp-rpc-2-hosts/)
HTML
<a href="https://getrigbase.com/builds/rtx-5090-5080-3x-5060-ti-96gb-llamacpp-rpc-2-hosts/" rel="noopener"><img src="https://getrigbase.com/badge/rtx-5090-5080-3x-5060-ti-96gb-llamacpp-rpc-2-hosts.svg" alt="Measured on RigBase" height="20"/></a>
Preview: 
· How to use