This machine is built to do several things at once. The point of two cards is not double speed on one model but two roles in parallel: one card runs the language model for the team, the other handles transcription, search across company data and invoice extraction. The numbers below are measured, not estimated.

First card, the language model. For everyday work — translation, summaries, e-mail triage — Gemma 4 26B runs in 4-bit quantisation: 19.9 GB of memory at a 32k-token context, roughly fifty pages of text, at 100 tokens per second. When accuracy matters more than speed, Qwen3.8-27B takes over at 20.4 GB: 33 tokens per second per request and 93 with eight concurrent users. Both fit comfortably on a single card.

Second card, everything else. Speech-to-text runs on Whisper large-v3 (faster-whisper, fp16) together with pyannote for speaker diarisation — just 4 to 5 GB of memory between them, and a twenty-minute recording is done in two minutes. The rest of the card stays free for search embeddings, reranking and invoice extraction. That is exactly why the second card is there: so transcription and search never wait for the language model to finish.

When you need the largest model. Qwen3-Coder-Next (80B MoE, 51 GB) spreads across both cards and serves as an auditor and tester for in-house development — finding bugs in code and across whole systems, from CRM to warehouse management to an e-shop. In our comparison it found all 8 planted bugs where the paid cloud model found 7. Across both cards it goes from 56 to 105 tokens per second per request and from 155 to 296 with eight concurrent users. It does take 23 GB on each card, so the rest of the workload has to wait or fall back to the CPU — you run it outside peak hours.

Where the ceiling is. A 70B-class model in 4-bit does not fit on one card at all: it needs 40 to 43 GB. It only runs across both cards, and even then just a few gigabytes are left for context, answers are slower, and the second card is unavailable for anything else meanwhile. We have not measured one ourselves; the only verified data point is a 51 GB model that did not fit into 48 GB and had to spill into system memory. Treat 70B as the ceiling, not as an operating mode.

Quality against the paid cloud. On real business tasks the local Qwen3.8-27B was more accurate than both DeepSeek and Gemini: e-mail triage 73% against 60%, product recommendations 67% against 60%, invoice extraction 8 out of 8 for both. And it answers faster — 1.2 seconds per e-mail against three to five. Models up to 32 billion parameters on a single card therefore replace paid cloud services outright, with no fees and no data leaving the building.

The ASUS ProArt X870E-Creator WiFi board is built for two graphics cards: both slots run at PCIe 5.0 x8/x8, full speed to each card. That is the difference from ordinary boards where the second card only gets four lanes — here tensor-parallel operation through vLLM makes sense, not just splitting the model layer by layer. An NVLink bridge is not needed for it; over PCIe two cards already deliver close to double. The board also offers 10Gb networking for moving models and datasets across the company network, four M.2 slots, two USB4 ports and Wi-Fi 7.

Cooling was measured, not estimated. Ten minutes of full load on both cards: 67 and 62 °C at most, fans at 69 and 60 percent, no thermal throttling. Both cards are Turbo versions with a radial blower that pushes hot air straight out of the case — with two cards stacked that is essential, ordinary triple-fan models would suffocate one another. The Fractal Design Torrent case is built precisely for maximum airflow. The whole server drew 686 W in that test, so the 1300 W Gold power supply still has room for the short spikes the 3090 is known for.

A twelve-core Ryzen 9 9900X and 64 GB of DDR5 handle data preparation, document indexing for RAG and serving several users at once without holding the graphics cards back. Two 2 TB SN850X drives keep models separate from company data. A PiKVM V4 Mini is included — remote management at keyboard and screen level, so you can reach the machine even when the operating system is down. For a machine in a server room or at another branch that saves a trip.

Good for: a language model for the whole team, an assistant over company documents, search across your own data, invoice extraction, transcribing recordings with speaker diarisation, code auditing and LoRA fine-tuning of smaller models — all of it concurrently, which is the whole point of two cards. Not for: training models from scratch, which calls for server cards with 80 GB of memory.