A 2026 build log documents 80+ tokens per second on a 27 billion parameter model by pairing a current generation GeForce with a refurbished RTX 3090, and the hard part is the firmware, not the money.
When the RTX 5080 landed in 2026, the obvious move was to slot it in and run the largest local language model a single card could justify. A small but growing corner of the enthusiast community found a different path: pair the new GeForce with a refurbished RTX 3090, then spend more time configuring the motherboard's firmware than the budget would suggest.
The reason is arithmetic. Qwen 3.6 27B, an open-weight 27-billion-parameter model in the Alibaba Qwen family, no longer fits comfortably in 16 GB of VRAM, the dedicated video memory on each graphics card. To run it at usable speed, you need roughly 24 GB of framebuffer across two cards. The cheapest way to get there in early 2026 was to combine a current-generation card with a previous-generation refurb, and a 2026 build log of an RTX 5080 and RTX 3090 setup documents what that actually takes: 80+ tokens per second on the Q8 quantization, where tokens are the word-fragments the model reads and writes, and "Q8" means the model's weights are stored at roughly eight-bit precision, trading some output quality for memory and speed.
Tokens per second is, in practical terms, the speed at which text appears on screen. Eighty tokens per second is roughly real-time reading speed, fast enough that the model can finish a paragraph while you are still reading its first sentence. The Q8 quantization, or "quant," is the lever that makes the speed possible: by compressing the model's numerical weights from 16-bit to 8-bit, the rig avoids running out of VRAM, at a documented cost to output coherence.
The build itself is a study in second-hand constraints. The author's Asus Prime X570-Pro motherboard is the load-bearing component, and the "Pro" SKU matters: only that variant splits its 16 PCIe lanes into two 8-lane groups so both GPUs can run at full bandwidth. Non-Pro X570-Pro boards reportedly lack that bifurcation. The 5080 sits on a PCIe 4.0 riser cable in the second slot to keep the primary slot at full bandwidth. The author used existing DDR4 memory and storage, buying only the motherboard and riser to enable the dual-GPU path.
Then the firmware fight. Compatibility Support Module (CSM), a legacy BIOS compatibility layer, must be disabled in the BIOS, the disk must be partitioned with a GPT (GUID Partition Table) layout, and the operating system must boot in UEFI (Unified Extensible Firmware Interface) mode rather than legacy BIOS. Skip any of these and the second card disappears from the system, or worse, both cards become unusable until kernel parameter workarounds are applied. The build log describes this as a hard blocker for the dual-GPU path on this specific board; readers on other boards may face different constraints.
The driver story is the quieter cost. NVIDIA's open-gpu-kernel-modules, internally codenamed "nova," work only when both GPUs are the same model, which rules out cross-generation pairs. Mixed-card builds live on nvidia-open, the older driver tree, with the performance and feature limits that implies. Peer-to-peer (P2P) GPU access, the ability for one card to read directly from another's memory, is uneven across the topology: the build log's topology matrix shows GPU0-to-GPU1 transfers working, but symmetry is not guaranteed for every workload.
The Q4 baseline on the 3090 alone ran at roughly 30 tokens per second, climbing to 50 to 60 tokens per second when Multi-Token Prediction (MTP), a speculative decoding trick that lets the model guess several tokens ahead and verify them in parallel, was enabled. The 80+ tokens per second headline figure comes from the Q8 quant across the two cards. The build log is upfront about the bottom line: the 5080 alone often trails the combo by a smaller margin than the spec sheets suggest, and the Q8 quantization is doing real work on output quality. The combo wins on memory headroom, not raw per-card throughput.
The pattern underneath the post is the actual story. In 2026, the default consumer path to running 27-billion-parameter models locally is no longer a single flagship card. It is a current-generation GeForce plus a refurb of the previous generation, with a firmware and driver tax on top. The Qwen family itself, moving from Qwen 3.5 through Gemma back to Qwen 3.6, has been the visible pressure: each release has pushed parameter counts past the 16 GB VRAM line that defined the 2023 enthusiast build.
What to watch next is the refurb market. If the Qwen 3.6 cycle or its successor keeps the floor above 24 GB, the 3090-tier refurb becomes a structural input to local LLM rigs, not a one-off workaround. If the next release fits back under 16 GB, the cross-generation combo goes back to being a curiosity. The hardware is settled. The model release cadence will decide whether this two-card pattern is a 2026 phenomenon or a template.