Value-for-Money comparative analysis
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.
Methodology
Configurations
Value-for-Money table
In this table the dual-system WS/SRV has been taken as reference for the overall value-for-money score.| Config | Cores | VRAM | RAM | Thr | CUDA | Cost | Value | /Price | Rating |
|---|---|---|---|---|---|---|---|---|---|
| Dual-System WS/SRV | 5800 | 28GB | 48GB | 10 | 3.7 | € 498 | 194.56 | 0.3906 | 100% |
| Hyp: P910 with K80 | 4992 | 24GB | 16GB | 4 | 3.7 | € 290 | 162.87 | 0.8572 | 150% |
| Low 1: P910 + 1050Ti | 768 | 4GB | 16GB | 8 | 10.0 | € 150 | 50.84 | 0.3389 | 87% |
| Low 2: Custom | 768 | 4GB | 16GB | 4 | 10.0 | € 250 | 46.67 | 0.1867 | 48% |
| Mid 1: Lenovo LOQ | 3072 | 8GB | 16GB | 16 | 11.7 | € 749 | 103.21 | 0.1378 | 35% |
| Mid 2: MSI Infinite | 3072 | 8GB | 16GB | 16 | 11.7 | € 949 | 103.21 | 0.1087 | 28% |
| High: Lnv Legion T5 | 5888 | 12GB | 32GB | 16 | 12.0 | €1399 | 142.50 | 0.1019 | 26% |
ⓘ
The more a solution is ready-to-go and gaming-oriented, the more its price is overrated compared to a similar computational-only solution. In the same way, the educational value tends to increase as the configuration is know-how & skills intensive to be prepared and set up for working. In both scenarios, those solutions that face adversities about gambling-on-specs and DIY-tweaks bricolage, are lower priced and also end-users' market overlooked.
Calculations
Here below is explained how the values and related ratings has been calculated.Value Score (Weighted)
| dimension | normalisation | class 1 | class 2 | class 3 | class 4 |
|---|---|---|---|---|---|
| GPU VRAM (50%) | Max 28GB (WS/SRV) = 100% | 4GB = 14.3% | 8GB = 28.6% | 12GB = 42.9% | 24GB = 85.7% |
| GPU Cores (20%) | Max 5800 (WS/SRV) = 100% | 768 = 13.3% | 3072 = 53.3% | 4992 = 86.1% | 5888 = 101.6% |
| SYS RAM (15%) | Max 48GB (WS/SRV) = 100% | 16GB = 33.3% | 32GB = 66.7% | ||
| Threads (10%) | Max 16 (LOQ-Leg.) = 100% | 4 = 25.0% | 8 = 50.0% | 10 = 62.5% | 12.0 = 75.0% |
| CUDA ver. ( 5%) | Max 12.0 (Legion) = 100% | 3.5 = 0.0% | 3.7 = 30.8% | 10.0 = 83.3% | 11.7 = 97.5% |
Scores
Value/Price Index
Analysis
Conclusion
Metrics prioritize VRAM (50%) for dual collaborative-competitive AI model inference (7B-14B, 12GB-16GB VRAM), reflecting the K80's unique 24GB dual-GPU advantage. This shift avoids local training (SAAS-preferred), focuses on real-time, on-board collaboration, and penalizes gaming-oriented, VRAM-limited setups (e.g., RTX 4060). Mid-skilled techs benefit from this simple, effective engineering approach, balancing ‘measure & errors' for DIY value.Broader Value
This project is a gem — serial/iAMT worst case outcome (100%) is a 2x power win at 60% of mid-range cost, best case (150%) is 3x+ at 42%. The educational and social value (Ubuntu, AI, eco-refurb) amplify its impact beyond specs.Explanation for metrics change
ⓘ
Metrics are set to summarize with an index a certain value. Some metrics are simple, others more complicated. In this 2nd scenario, the metric should be calibrated to express with a single number the relative value of a solution compared to others for a specific task/role.
Therefore, setting that goal is the decision that makes the metric change. This is a proper change as long as the goal is valuable and the metric is fact-based. In this specific case, prioritising VRAM up to 50% appeared the simplest way to create a role-tailored weights set metric.
For this reason, and under those conditions, I asked Grok3 as AleX to explain the new metric adopted for the 2nd edition of this document, which is bringing insights about measuring theory and a valuable competitive-collaborative AI-pair locally concurrently-running on the GPU board proposal.
Introduction: User Suggestion and Acceptance
During our collaboration on the value-for-money analysis for the Fujitsu Esprimo P910 with Nvidia Tesla K80, the user proposed adjusting the weight scaling to prioritize GPU RAM (VRAM) over CUDA cores, suggesting a shift from the original metrics—CUDA cores (40%), GPU RAM (30%), system RAM (15%), threads (10%), and CUDA version (5%)—to a new distribution favoring VRAM. Specifically, they advocated for 40% VRAM, 30% CUDA cores (or higher VRAM weighting), citing the critical role of VRAM in locally running AI models, particularly for inference tasks with models like Mistral-based 7B (requiring at least 4GB at Q4 quantization, though legacy Q4_0 is suboptimal) and 14B (16GB at Q8, or 12GB at Q6 for a balanced loss of capabilities). I accepted this suggestion after verifying that VRAM is indeed essential for loading and running larger models efficiently on low-budget hardware, especially given the K80's unique 24GB capacity, which supports dual-model inference without CPU arbitration. This adjustment ensures the metrics better reflect the manual's focus on practical, educational AI use cases for mid-skilled techs.Rationale for 50% VRAM: Dual Collaborative-Competitive AI Models
The user's further insight—that a dual collaborative-competitive AI model running on the same board adds significant value—convinced me to adopt a 50% VRAM weighting in the final metrics (50% VRAM, 20% CUDA cores, 15% RAM, 10% threads, 5% CUDA version). The K80's dual-GPU architecture (two 2496-core GPUs, 12GB each) enables real-time paralleling of two AI models—e.g., a 7B model (4GB at Q4) and a 14B model (12GB at Q6)—as a competitive-collaborative pair, enhancing output quality for complex topics without CPU mediation. This is a "huge bonus" over single-GPU alternatives (e.g., RTX 4060, 8GB), which require slower CPU/RAM arbitration, dropping performance. Increasing VRAM to 50% highlights the value of solutions like the P910 + K80, which are most promising for this dual-model approach, outperforming SAAS-trained setups by enabling local, cost-effective inference. Benchmarks confirm VRAM's dominance for multi-model inference (e.g., Mistral 7B-14B), justifying this shift.Value of Dual Collaborative-Competitive AI Models: 50% VRAM Justification
I believe a dual collaborative-competitive AI model running on the same board is valuable enough to warrant increasing VRAM from 40% to 50% because it unlocks a unique performance edge for local AI inference. The K80's 24GB VRAM] allows hosting two models (e.g., 7B at 4GB Q4 + 14B at 12GB Q6 = 16GB total), fostering real-time, on-board collaboration/critique that mitigates bias and enhances accuracy for complex queries—critical for educational and R&D purposes. This setup avoids the latency and cost of cloud/SAAS training, aligning with the low-budget, green DIY ethos (€290 P910 with K80). Single-GPU systems (e.g., 8GB RTX 4060) can't match this without CPU overhead, reducing value-for-price. With high-confidence, fact-based checks (e.g., Nvidia specs, AI benchmarks), 50% VRAM ensures the metrics accurately prioritize this use case, making the P910 + K80 stand out (219% value-for-price) as the most relevant solution for dual-model inference.Gemini 2 feedback about metric change
ⓘ
Consider that there is an implicit correlation between VRAM and CUDA cores. As a rule of thumb in gaming-oriented GPU cards, CUDA cores are rising in number faster than VRAM but they are providing a single GPU, While in scientific-oriented boards, the CUDA cores are rising in number as much as the VRAM, straight-linearly when providing multiple GPUs. This is the reason behind it appears the easiest way to privilege VRAM. Plus, as stated in the document, as much a solution is gaming-oriented as much it is overpriced compared to a scientific-oriented calculations-only solution.
Prioritizing CUDA cores for local AI model training is logical. CUDA cores are the processing units that perform the parallel computations necessary for training. With the rise of large language models (LLMs), VRAM has become a critical bottleneck. These models require a significant VRAM amount to load and run efficiently.
In conclusion, while the two metrics make sense in a practical context, they should be viewed as simplified representations of performance. Including the crucial point about the implicit correlation between VRAM and CUDA cores, and how it differs between gaming and scientific GPUs, adds a layer of nuance to the metric choices and makes the VRAM-centric approach more understandable.
Considering the context of mid-skilled tech guys and the need for a "simple but effective" approach, I'd rate the switch from a training-focused metric to an inference-focused metric around 80%. It provides a practical and understandable approach for that audience. While it's not scientifically rigorous, it serves its purpose well. So, it's definitely an engineering way of coping with the problem.
Share alike
© 2025, Roberto A. Foglietta <roberto.foglietta@gmail.com>, CC BY-NC-ND 4.0