Thelio Mira AI Linux Workstation: 192 GB GPU Memory
System76’s new Thelio Mira AI Linux workstation, configurable with dual NVIDIA RTX Pro 6000 GPUs and 192 GB of VRAM, is drawing attention for configurations that can exceed $40,000–$50,000 despite a prominently advertised $3,299 base price. Commenters largely attribute the cost to Nvidia GPU pricing and high-end memory, and debate whether such a machine offers better value than alternatives like Mac Studio clusters, single H100-class cards, or cloud GPU access. Technical concerns are also raised around thermals, PCIe lane bottlenecks, and the choice of consumer CPUs in such an expensive “local AI development” rig.
Price and Value
- Base price (
$3.3k) vs maxed config ($40–50k) surprises many; several call the pricing “insane” or “obscene” for a workstation. - ~88% of a ~$42k build is just the two RTX Pro 6000 GPUs; RAM and PSU are a small fraction.
- Some compare it to buying an $80k car: for six‑figure professionals, an expensive work tool is defensible; others note that very few people actually buy cars at that price.
- Multiple comments say at that level they’d rather buy a single H100, a used server‑grade workstation, or just use cloud GPUs.
Hardware Configuration & Power
- GPUs: dual RTX Pro 6000 (96 GB each, 192 GB total VRAM) is the marquee option; cards are non‑refundable and extremely costly.
- Concerns about thermals and noise: questions about stacking workstation cards, cooling gaps, and whether all configurations are actually tested.
- CPU choice (Ryzen 9950X) is criticized as “toy” or PCIe‑lane‑starved for dual high‑end GPUs; some argue Threadripper/Xeon would be more appropriate at this price.
- Discussion of DDR5 memory speed dropping with two DIMMs per channel and the lack of a modern HEDT platform.
- High power draw implies need for strong PSUs, UPS, and possibly upgraded home/office electrical work.
Performance vs Alternatives (Macs, H100, etc.)
- Consensus that dual RTX 6000 outperforms Apple Silicon (e.g., M5 Ultra) for pure GPU compute, especially for LLM prefill/token rates.
- Others argue that for the same money, multiple Mac Studios (with large unified memory pools) might run much larger dense models, even if per‑GPU bandwidth is lower.
- Apple is seen as slower for prompt processing today, but gaining bandwidth each generation; some see Macs as quiet, high‑acceptance, good‑RAM options but still behind Nvidia in raw speed.
- Debate over whether RTX Pro 6000 vs H100 is a fair comparison: H100 is data‑center grade with more bandwidth and compute.
Use Cases, Longevity, and Purchasing Decisions
- Some buyers frame this as a company expense that pays for itself in productivity; others say model churn and rapidly rising VRAM requirements make such a purchase risky.
- Regret over not buying cheaper 4090s earlier; fear that continued price hikes could “end the personal computing era” and push users back to “big iron”/shared resources.
- Ideas floated about GPU co‑ops (shared ownership of large GPU boxes) instead of ultra‑expensive personal workstations.
OS, Ecosystem, and Vendor Perception
- System76 and Pop!_OS are praised for open hardware and long‑lived machines.
- Skepticism that enterprises dropping $40k+ would choose Pop!_OS over Ubuntu, though System76 also sells Threadripper/Xeon systems via other SKUs.
Marketing and Communication Issues
- “Affordable” tagline and $3,299 headline price are viewed as misleading given that the AI‑flagship configuration costs an order of magnitude more.
- The “192 GB GPU memory” line is noted: some argue that if GPUs only communicate over PCIe, calling it a single pool of “GPU memory” is questionable.