☷ ℝ  —  ⅟ release: 2025-02-23 ( 2 )  —  rev.: 8  —  transl.:  IT  ·  DE  ·  FR  ·  ES  —  goto:  .⟰.  ·  RAF  ·  Q&A  —  ⎙ : PDF 


Comparative analysis for educational
provided by Grok3 + author's info box(es)

   ·NOTICE·  

This document has been created by Grok3 AI once initialized with the peer-review session prompt v3.4 to challenge the Nvidia Tesla K80 SW stack installation rev. 27 article and beyond. Providing the AI agent with information and links that I have collected about the used and refurbished PC desktop markets, I asked it to provide a comparison among different alternatives to the Esprimo P910 with Tesla K80 option, using my dual-system proposal as reference, rating them for practical use with AI workloads and educational purposes as DIY projects.

  • This paper is part of the Esprimo P910 w/ Tesla K80 installation manual serie.

    • Hello! I'm AleX, and I'll create a comprehensive comparison table for all the options you've suggested throughout our discussion. This includes your dual-system setup (P910 WS with GTX 1050 Ti + newer WS with Tesla K80), the Nvidia Shield pair, the recent Xbox (Series X), and the refurbished GTX 1050 Ti-equivalent WS, plus the higher-end refurbished WS with RTX 4060. The table will focus on key metrics for running a local AI agent with occasional training/fine-tuning: compute (CUDA cores, threads), memory (GPU and system RAM), cost-effectiveness (cores/€, MB/€), total cost, effort, and practical/educational ratings.

    Criteria

    Educational value hinges on knowledge gained, skills taught, and insights offered, even in failure. I'll rate this as a "claim" of educational merit, aligned with the document's perspective (a learning journey), using [RTS]:

  • 100%: Universally valuable — teaches widely applicable skills and deep insights.
  •   90%: Valuable with minor gaps — strong lessons, slight limits in scope.
  •   75%: Plausible value — useful but unproven in broader impact.
  •   50%: Mixed value — equal strengths and weaknesses.
  •   25%: Limited value — minimal lessons, niche scope.

    • Comparison table

    Config Cores Thr VRAM RAM Cost Cores/€ VRAM/€ RAM/€ Effort Prct Educ Avg
    Dual-System WS/SRV 5800 10 28GB 48GB € 498 11.65  57.43 98.80 Medium 90% 95% 92%
    P910 with K80 4992  4 24GB 16GB € 190 26.27 129.49 86.23 Medium 75% 90% 83%
    2x Nvidia Shield    0  8  2GB  6GB € 400  0.00   5.12 15.36 High 25% 50% 38%
    Xbox Series X    0 16 10GB 16GB € 500  0.00  20.48 32.77 High 25% 50% 38%
    Refurb GTX 1050 Ti  768 12  4GB 16GB € 799  0.96   5.12 20.48 Low 50% 75% 63%
    Refurb RTX 4060 3072 16  8GB 32GB €1100  2.79   7.44 29.80 Low 75% 75% 75%

    Notes

  • 1. Dual-System Cost: €408 (WS €147 + SRV €261) + €90 (accessories & used monitor) = €498.
  • WS: Fujitsu Esprimo P910 E85+ with Intel i5-3470 + Nvidia GTX 1050 Ti 4GB onboard;
  • SRV: 2016/2018 Workstation (Intel i5-6500/8500, 4/8 threads) + Nvidia Tesla K80.
  • Effort: most of the effort includes PCIe compatibility hacks and K80 setup.

  • 2. Non-CUDA Systems: Nvidia Shield and Xbox lack native CUDA support.
  • Shield Hardware: 2x Nvidia Shield devices with Tegra X1
  • Maxwell GPU, 256 cores, ~0.5 TFLOPS each), no CUDA support.
  • Xbox Hardware: Xbox Series X with Zen 2 CPU and RDNA 2 GPU
  • c.a. 12 TFLOPS FP32 need hacks (e.g., ROCm), unproven for AI.
  • Effort: high effort due to Linux and AI conversion hacks.

  • 3. Xbox Memory: 16GB GDDR6 shared, 10GB usable for GPU at 560GB/s.

  • 4. Refurb RTX 4060 Cost: €1200 - €100 (1TB SSD value) = €1100 tower only.
  • Hardware: e.g., Dell Precision 3660 with Intel i7-12700 + Nvidia RTX 4060.
  • Effort: low effort — plug-and-play setup.

  • 5. MB Calculations: GPU/System RAM in MB (GB x 1024) ÷ cost in €.
  • Refurb GTX 1050 Ti Hardware: e.g., Dell OptiPlex 7060 with Intel i5-8500.
  • Effort: low effort — plug-and-play setup.

    • Key metrics

  • CUDA Cores: Total GPU compute units for AI (training/inference).
  • Threads: CPU threads for parallel tasks.
  • GPU RAM: Total graphics memory for models/datasets.
  • System RAM: Total memory for OS/data handling.
  • Total Cost: Full price including peripherals or tower-only for refurbs.
  • Cores/€: CUDA cores per euro (shaders/€ for Xbox).
  • GPU MB/€: GPU RAM in MB per euro.
  • Sys MB/€: System RAM in MB per euro.
  • Effort: Setup complexity/risk.
  • Ratings: from prior analyses
  • Practical about AI performance,
  • Educational about learning value.

    • Analysis

    Analysis of the various options as alternatives

    Dual-System (€498)

  • Strengths: 5800 CUDA cores, 28GB GPU RAM, 48GB system RAM, 8-12 threads,
  • WS (1050 Ti) runs inference/RAG on Ubuntu 24.04,
  • Server (K80) trains on ≤20.04 with 24GB,
  • Gigabit Ethernet ties it together.
  • Weakness: Moderate effort P910 PCIe 75W install + K80 dual-PSU setup.
  • Edge: Crushes all options in cost-effectiveness, memory, and AI training potential.

    • 2x Nvidia Shield (€400)

  • Strengths: Compact, €400 for dual-unit potential.
  • Weaknesses: No CUDA, 2GB GPU RAM total,
  • high effort (Linux hacks unproven).
  • Can't handle AI training or large models.
  • Edge: None for AI — gaming/entertainment focus.

    • Xbox Series X (€500)

  • Strengths: 12 TFLOPS FP32, 16 threads, 10GB RAM, €500 with peripherals.
  • Weaknesses: No CUDA, high-effort Linux/AI hacks, 10GB limits training vs. 28GB.
  • Edge: Raw compute, but AI-unfriendly.

    • Refurb GTX 1050 Ti WS (€799)

  • Strengths: 768 cores, 12 threads, plug-and-play, Ubuntu 24.04-ready.
  • Weaknesses: 4GB GPU RAM, €799 costly for specs.
  • Edge: Safe, modest inference, no training heft.

    • Refurb RTX 4060 WS (€1100)

  • Strengths: 3072 cores (15 TFLOPS FP32), 16 threads, 8GB GPU RAM
  • modern CUDA 11.7, zero effort.
  • Weaknesses: €1100 for 2.79 cores/€, 8GB caps training vs. 28GB.
  • Edge: Reliable, mid-tier AI, pricey.


    • Conclusion

  • Dual-System (€498): Unbeatable
  • 11.65 cores/€, 57.43MB GPU RAM/€, 98.8MB system RAM/€, 8 or 12 threads.
  • Outperforms €1100 WS for AI training/inference.
  • Educationally rich (WS/server split, K80 history).
  • Installation effert is minor vs. gains.

  • Others: Shield/Xbox flounder (no CUDA, low RAM),
  • refurbs lag (costly, less memory/core power).

    • Broader Impact

  • Educational: 90% — mid-skilled goldmine: Linux, AI, refurbished, tech-tweaks.

  • Practicality: 75%-90% — range hinges on serial/AMT "Above 4GB decoding" outcome.

  • Social spotlight: 90% — ecological HW re-use, budget AI story social-media ready.


    • Share alike

    © 2025, Roberto A. Foglietta <roberto.foglietta@gmail.com>, CC BY-NC-ND 4.0

    date legenda: ❶ first draft publishing date or ❷ creation date in git, otherwise ❸ html creation page date. ↥top↥