My role

Requirements gathering, component research, system design, sourcing, assembly, configuration, testing, user guidance, and post-build support

Project scope

Multiple purpose-built workstations designed for personal use and other users

Hardware strategy

Repurposed RTX 3090 GPUs combined with supporting components selected around each system’s intended use

Core environment

Windows 11, updated NVIDIA drivers, CUDA support, local AI software, and workload-specific applications

Primary workloads

Local language models, image generation, coding, data analysis, document workflows, graphic design, productivity, and gaming

Primary challenge

Translating each user’s goals, budget, preferences, and future plans into a balanced system

Final outcome

Built, configured, tested, and delivered multiple practical workstations, including a separate system retained for ongoing personal use

The original goal was to build a personal workstation that was better suited to local AI experimentation than the six-GPU mining platform.

Rather than relying on a large riser-based system, I wanted a workstation that could support local models and everyday technical work while being easier to cool, maintain, configure, and operate.

The first workstation was built for my own use. As others became interested, I began building additional systems designed around different combinations of:

• Local AI

• Local language-model inference

• Image generation

• Graphic design

• Coding

• Data analysis

• Document-assisted workflows

• General productivity

• High-performance gaming

What began as a personal system eventually developed into a broader workstation-design project involving multiple builds and users.

Unlike the six-GPU platform, the workstations did not use one fixed configuration.

Each system was designed around its intended workload, budget, user preferences, and future upgrade needs.

Common elements included:

• NVIDIA RTX 3090 GPUs repurposed from the earlier multi-GPU platform

• Motherboards selected for the intended CPU and expansion requirements

• CPU, memory, and storage configurations matched to the expected workload

• Power supplies sized around the complete system

• Cooling selected for the case, processor, GPU, and operating environment

• Windows 11

• Updated NVIDIA drivers

• CUDA support

• Local AI applications

• Workload-specific software

I handled nearly the entire workstation-development process, including:

• Discussing user goals and intended workloads

• Gathering system requirements

• Researching components

• Comparing hardware options

• Making hardware recommendations

• Establishing and working within a budget

• Planning future upgrade paths

• Sourcing components

• Verifying component compatibility

• Assembling the hardware

• Managing cable routing and internal layout

• Installing and configuring Windows

• Installing NVIDIA drivers

• Configuring CUDA support

• Installing selected applications

• Testing the completed system

• Providing setup and usage guidance

• Offering post-build support

The process required more than assembling a collection of compatible parts. Each build involved translating a general idea of what the user wanted into a practical system configuration.

The primary challenge was determining what each user actually needed from the system.

Users often described their goals in broad terms, such as wanting a computer for AI, gaming, graphic work, coding, or general high-performance use. Those goals had to be converted into specific hardware, software, cooling, power, and budget requirements.

Factors Considered

• Intended applications

• Expected workload

• Local AI requirements

• GPU memory

• CPU performance

• System memory

• Storage capacity

• Power consumption

• Cooling

• Physical size

• Appearance

• Budget

• Upgradeability

• Ease of use

• Long-term reliability

Design decision: I avoided using one standard configuration for every workstation. Each system was treated as a separate design problem based on the user’s actual priorities.

In one case, the user had already selected the computer case they wanted. I designed the remaining system around that preference while accounting for the available space, cooling requirements, budget, workload, and future upgrade plans.

The NVIDIA RTX 3090 was the central performance component in several builds, but the GPU alone did not determine whether the workstation would be successful.

The surrounding components had to support the intended workload without creating unnecessary cost, heat, complexity, or limitations.

Design Priorities

• Matching the CPU to the workload

• Providing enough system memory

• Selecting appropriate storage

• Sizing the power supply correctly

• Maintaining sufficient airflow

• Avoiding unnecessary bottlenecks

• Preserving a reasonable upgrade path

• Balancing performance against cost

• Accounting for appearance and case preferences

• Keeping the system maintainable

Design decision: The objective was not to use the most expensive component in every category. The objective was to create a balanced workstation in which the components supported the intended use case.

This represented a different approach from the six-GPU platform. Instead of maximizing the amount of installed GPU hardware, I focused on matching one complete system to a defined set of requirements.

The RTX 3090 GPUs were repurposed from the original multi-GPU platform.

This allowed the GPUs to be redirected into systems that were more practical for individual use while preserving their 24 GB of VRAM for workloads such as local language models, image generation, graphics, and other GPU-accelerated applications.

The supporting components were selected separately for each workstation rather than retaining the mining platform’s original architecture.

This approach made it possible to design around:

• Direct motherboard GPU connections

• Appropriate processor performance

• Sufficient system memory

• Faster storage

• Workstation-specific cooling

• Proper power delivery

• Individual software requirements

• Future component upgrades

The result was a collection of independent systems rather than one large shared platform.

Software varied according to the date of the build and the user’s intended workload.

Depending on the system, configuration included:

• Windows 11

• Current NVIDIA drivers

• CUDA support

• GPT4All

• Locally available Llama-family models

• Locally available DeepSeek-family models

• Image-generation software

• Coding and development tools

• Data-analysis applications

• Document and productivity software

• Graphics applications

• Gaming software

Some build-specific software and configuration details are intentionally omitted from the public portfolio.

Configuration principle: The software environment was selected around the intended use of the workstation rather than installing the same collection of applications on every system.

Each workstation was configured and tested before it was placed into service or handed over to another user.

The process included confirming that:

• Windows installed and operated correctly

• Installed hardware was recognized

• NVIDIA drivers functioned correctly

• CUDA support was available where required

• Selected applications launched and operated as intended

• The system was prepared for its expected workloads

• Users understood the basic setup and operation of the system

The builds did not encounter the same type of recurring operational failures as the earlier six-GPU platform.

The greater challenge was ensuring that each completed workstation matched the expectations established during the planning process.

After completion, I provided setup guidance and post-build support when needed.

A workstation was considered ready for service when:

• All installed components were recognized

• Windows operated without unresolved installation or driver issues

• NVIDIA drivers functioned correctly

• Storage, memory, graphics, and connected devices were available

• Required applications launched successfully

• The system was prepared for its intended workload

• Basic setup and operating guidance had been provided

• Known limitations or future upgrade considerations had been communicated

The project evolved from personal experimentation into the development of practical workstations for multiple users.

Each completed system was:

• Designed around a defined use case

• Built using components selected for that system

• Configured with an appropriate software environment

• Tested before use or handoff

• Prepared for local AI or other demanding workloads

• Accompanied by setup guidance and support

The project also resulted in:

• Productive reuse of hardware from the multi-GPU platform

• Smaller and more manageable systems

• Reduced dependence on the original riser-based architecture

• Better alignment between hardware and actual workloads

• Practical experience translating user goals into system specifications

• Continued development of workstation design and support skills

Some systems were retained, sold, or gifted, while a separate workstation was built for my ongoing personal use.

To my knowledge, the delivered systems remained operational after handoff. Their later software configurations, maintenance, and usage were outside my direct control.

The purpose-built workstation approach addressed many of the limitations encountered with the six-GPU platform.

Compared with operating all six GPUs in one system, the workstations were:

• Easier to cool

• Easier to maintain

• Simpler to configure

• Better suited to individual users

• More practical for mixed workloads

• Less dependent on riser-based GPU connections

• Easier to upgrade over time

• Better aligned with real-world budgets and use cases

The goal was no longer to combine as much GPU hardware as possible into one environment.

The goal became delivering the right combination of hardware and software for the person using the system.

I am most proud that I was able to turn lessons from the original multi-GPU platform into systems that were more practical, efficient, and appropriate for real users.

The work required more than assembling components. It involved understanding goals, identifying constraints, recommending an appropriate architecture, configuring the hardware and software environment, and helping users begin working with their systems confidently.

The project demonstrated that a technical limitation in one system can become the starting point for a better design approach.

The Custom AI Workstations project remains ongoing.

The personal workstation continues to support local AI experimentation, coding, data analysis, graphic work, gaming, document workflows, and other technical projects.

Future workstation decisions continue to be guided by the lessons learned across the complete project sequence:

• Build and operate the multi-GPU platform

• Test its suitability for local AI

• Identify the architectural limitations

• Redirect the hardware into purpose-built systems

• Match each new system to a practical workload

A successful workstation is not defined by the most powerful hardware available.

It is defined by how well the complete system matches the user’s actual needs, including:

• Intended workload

• Budget

• Performance requirements

• Power and cooling

• Software compatibility

• Reliability

• Usability

• Maintainability

• Appearance

• Future upgrade plans

The project reinforced that good system design begins with understanding the use case—not with selecting components.