PC for machine learning engineer: training, fine-tuning, research
A machine learning engineer does not have the same needs as a user who just wants to chat with a model. Training, fine-tuning, experimenting, processing large datasets: these tasks impose very different hardware constraints than simple inference. Buying the wrong machine means losing hours at every experimentation cycle.
This guide starts from the real workflows of an ML engineer, deduces what hardware must provide, and proposes stations suited to each type of workload, from development workstations to multi-GPU training stations.
What an ML engineer really does, and what it requires
Model training
Requires: VRAM, computation, long-term stability
Training from scratch or continuing pre-training fully loads the GPU, sometimes for days. VRAM limits model and batch size. Stability becomes critical on long runs.
Fine-tuning (LoRA, QLoRA, full)
Requires: VRAM, iteration speed
The most common case in practice. LoRA and QLoRA reduce requirements, but serious fine-tuning demands 24 GB and more depending on the base model size.
Inference and evaluation
Requires: VRAM, bandwidth
Testing models, comparing variants, serving a local API. Less demanding than training, but VRAM remains the model size factor.
Data preparation
Requires: CPU, RAM, fast storage
Cleaning, tokenization, augmentation, loading. This often underestimated step is limited by the CPU, RAM, and storage speed, not the GPU.
The components that really matter
- VRAM, above all. It determines the size of the models you can train and fine-tune. 24 GB is a comfort threshold, 32 GB opens up serious models, 96 GB ECC targets research and large models.
- ECC memory for long runs. During training that lasts several days, a silent memory error can corrupt an entire run. ECC VRAM (RTX 6000 Blackwell cards) protects critical computations.
- CPU and RAM for the data pipeline. A powerful GPU starved by slow data loading idles. Many cores and ample RAM feed the GPU without bottlenecks.
- Fast NVMe storage. Large datasets and checkpoints require high throughput. A Gen 4 or Gen 5 NVMe prevents the disk from becoming a bottleneck.
- Multi-GPU for scaling up. Two cards allow parallel training, handling larger models, or running multiple experiments simultaneously.
Which VRAM for which ML task?
| Task | Recommended VRAM | Typical card | Comment |
|---|---|---|---|
| Learning, prototyping, small models | 16 GB | RTX 5070 Ti | Ideal for starting and developing |
| LoRA/QLoRA fine-tuning (up to 14B) | 24 to 32 GB | RTX 5090 32 GB | The industry sweet spot |
| Multi-experiment training, medium models | 2 × 32 Go | 2 × RTX 5090 | Parallelism, multiple runs |
| Heavy fine-tuning, large models, long runs | 96 GB ECC | RTX 6000 Blackwell | ECC for reliability |
| Research, full fine-tuning, R&D | 192 GB ECC | 2 × RTX 6000 Blackwell | Top-tier local setup |
| Prototyping very large models | 128 GB unified | NVIDIA GB10 | Unified memory, compact form factor |
A local environment ready to code
Our workstations come with the standard ML ecosystem preconfigured on request, so you can start coding right out of the box instead of spending hours fixing CUDA version issues.
# Typical environment, preconfigured on request # PyTorch with CUDA 12.8 (Blackwell RTX 50xx / RTX 6000 cards) pip install torch torchvision torchaudio \ --index-url https://download.pytorch.org/whl/cu128 # Common ML tools pip install transformers datasets accelerate peft bitsandbytes pip install jupyterlab scikit-learn pandas # Ready-to-use LoRA fine-tuning with PEFT + Transformers
Our workstations by ML workload type
All our machines are hand-assembled in Auriol (13390), tested before shipping, and delivered throughout the European Union. Fully configurable, including the GPU.
Radiance PC CoreAI 32 — RTX 5070 Ti 16 GB
Ideal for learning, developing, inference, and light LoRA fine-tuning.
Entry-level ML development workstation. 12 cores for the data pipeline, 16 GB VRAM for inference and prototyping. The right starting point before serious training.
PyTorch + CUDA stack preconfigured on request
Set up this workstationRadiance PC CoreAI 64 — RTX 5090 32 GB
32 GB VRAM and record bandwidth: the reference machine for LoRA and QLoRA fine-tuning.
The workstation covering the essentials of the job. 32 GB to fine-tune models up to 14 billion parameters, 64 GB RAM for the data pipeline, a 16-core CPU with 3D cache. The best capacity/price ratio for an individual ML engineer.
Full ML stack preconfigured on request
Set up this workstationRadiance CoreAI Rack — 2 × RTX 5090 (64 GB)
Two GPUs for parallel training, running multiple experiments, or splitting a large model.
For scaling up. Two RTX 5090s enable distributed training, handling larger models, or running multiple experiments simultaneously. 128 GB of RAM to feed both cards without bottlenecks.
Distributed training, on-site installation available
Configure this rackRadiance PC Pro AI Ultra — Threadripper PRO
96 GB of ECC VRAM and up to 2 TB of RAM: for heavy fine-tuning and research.
The platform for ML engineers pushing boundaries. 96 GB ECC VRAM for large models and long runs without corruption risk, a Threadripper PRO CPU, and RAM expandable up to 2 TB for the most demanding data pipelines.
Custom, personalized quote, on-site installation
Request a quoteCoreAI 128 Rack — 2 × RTX 6000 Blackwell (192 GB ECC)
192 GB ECC VRAM for full fine-tuning, large models, and continuous R&D.
The pinnacle of our range for local ML. 192 GB ECC VRAM enables full fine-tuning of substantial models, distributed training on two professional cards, and datacenter-level reliability for continuous workloads.
R&D, full fine-tuning, on-site installation
Request a quoteNVIDIA GB10 Mini AI Server — ASUS Ascent GX10
128 GB unified memory to prototype very large models, in a desktop format.
A different approach: 128 GB unified CPU-GPU memory allows loading models that even an RTX 5090 cannot handle, in a compact and silent format, with CUDA and Jupyter environment ready at startup.
DGX OS, ready-to-use ML environment
Discover this serverWhy choose a local workstation over the cloud
GPU cloud has its place, but for an ML engineer working daily, a local workstation offers concrete advantages.
- Controlled cost. GPU hours in the cloud add up quickly. A local workstation is a one-time investment, paid off in a few months of intensive use.
- Instant iteration. No provisioning, no instance waiting, no data transfer. You start your experiments immediately.
- Private data. Your datasets and proprietary models stay with you, without passing through a third-party provider.
- Full availability. No GPU quotas, no instance shortages, no downtime. Your machine is always there.
- Stable environment. Your software stack doesn’t change between sessions.
In short
What VRAM is needed for an ML engineer?
16 GB for development and prototyping, 24 to 32 GB for LoRA/QLoRA fine-tuning, 96 GB ECC and more for full fine-tuning and research.
Is ECC memory necessary?
For long training sessions lasting several days, yes: ECC protects against silent memory errors that can corrupt a run. RTX 6000 Blackwell cards are equipped with it.
Does the CPU matter for ML?
Yes, for data preparation. A powerful GPU poorly fed runs idle. Many cores and generous RAM avoid this bottleneck.
Does fine-tuning require more than inference?
Definitely. Due to gradients and optimizer states, fine-tuning can require two to three times the VRAM of inference for the same model.
Local or cloud?
Local is more economical and faster for daily use. A hybrid approach, local plus occasional cloud, is often the most relevant.
Are the machines ready to code?
Yes, on request: PyTorch, CUDA, Hugging Face libraries, and Jupyter preconfigured, so you can start coding right out of the box.
Can the configuration be customized?
Yes, completely. The graphics card, CPU, RAM, storage, power supply, cooling, and case are configurable on each product page via the online configurator. For a specific need or custom configuration, contact us at contact@radiancesystems.eu or via the quote form: we tailor the machine to your exact workload.
