PC for Machine Learning Engineer: Training, Fine-tuning, Research
A machine learning engineer has different needs from a user who just wants to chat with a model. Training, fine-tuning, experimenting, processing large datasets: these tasks impose hardware constraints very different from simple inference. Buying the wrong machine means wasting hours on every experimentation cycle.
This guide starts from the real workflows of an ML engineer, deduces what the hardware must provide, and offers the workstations suited to each type of workload, from the development workstation to the multi-GPU training station.
What an ML engineer really does, and what it requires
Model training
Requires: VRAM, compute, long-term stability
Training from scratch or continuing pre-training fully utilizes the GPU, sometimes for days. VRAM limits the size of the model and the batch. 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 requires 24 GB or more depending on the size of the base model.
Inference and evaluation
Requires: VRAM, bandwidth
Testing your 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 step, often underestimated, is limited by the CPU, RAM, and storage speed, not by the GPU.
The components that really matter
- VRAM, above all. It sets 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. On a multi-day training run, a silent memory error can corrupt an entire run. ECC VRAM (RTX PRO 6000 Blackwell cards) protects critical calculations.
- CPU and RAM for the data pipeline. A powerful GPU is wasted if it is starved by slow data loading. Plenty of cores and ample RAM ensure your GPU runs at full capacity without bottlenecks.
- Fast NVMe storage. Large datasets and checkpoints require high throughput. A Gen 4 or Gen 5 NVMe ensures that your storage does not become a bottleneck.
- Multi-GPU for scaling. Dual cards allow for parallel training, handling larger models, or running multiple experiments simultaneously.
Which VRAM for which ML task?
| Task | Recommended VRAM | Card type | Comment |
|---|---|---|---|
| Learning, prototyping, small models | 16 GB | RTX 5070 Ti | Ideal for getting started and development |
| LoRA/QLoRA fine-tuning (up to 14B) | 24 to 32 GB | RTX 5090 32 GB | The professional balance point |
| Multi-experiment training, medium models | 2 × 32 GB | 2 × RTX 5090 | Parallelism, several runs |
| Heavy fine-tuning, large models, long runs | 96 GB ECC | RTX PRO 6000 Blackwell | ECC for reliability |
| Research, full fine-tuning, R&D | 192 GB ECC | 2 × RTX PRO 6000 Blackwell | Top-tier in dual card |
| Very large model prototyping | 128 GB unified | NVIDIA GB10 | Unified memory, compact format |
A local environment ready to code
Our workstations are delivered with the standard ML ecosystem preconfigured on request, so you code from unboxing rather than spending hours fixing CUDA versions.
# 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 # LoRA fine-tuning ready to use with PEFT + Transformers
Our workstations by ML workload type
All our machines are assembled by hand in Auriol (13390), except the GB10 assembled by ASUS, tested before shipping, and delivered throughout the European Union. Fully configurable, including the GPU.
CoreIA RTX 5090
32 GB of VRAM and record bandwidth: the reference machine for LoRA and QLoRA fine-tuning.
The workstation that covers the essentials of the job. 32 GB to fine-tune models up to 14 billion parameters, 64 GB of RAM for the data pipeline, a 12-core CPU. The reference choice for an individual ML engineer.
Complete ML stack preconfigured on request
Configure this workstation
CoreIA 2x RTX 5090 (64 GB)
Two GPUs to train in parallel, launch several experiments, or distribute a large model.
To scale. Two RTX 5090s allow distributed training, processing larger models, or launching several experimentation runs simultaneously. 128 GB of RAM to feed both cards without bottlenecks.
Distributed training, on-site installation possible
Configure this rack
CoreIA Threadripper PRO 4 GPU
96 GB of ECC VRAM and up to 2 TB of RAM: for heavy fine-tuning and research.
The platform of ML engineers who push the limits. 96 GB of ECC VRAM for large models and long runs without risk of corruption, a Threadripper PRO CPU and RAM expandable to 2 TB for the most demanding data pipelines.
Custom-made, personalized quote, on-site installation
Request a quote
CoreIA 2x RTX PRO 6000 (192 GB ECC)
192 GB of ECC VRAM for full fine-tuning, large models, and continuous R&D.
The top of our range in dual card for local ML. 192 GB of ECC VRAM allows 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 quote
CoreIA GB10 (ASUS Ascent GX10)
128 GB of unified memory to prototype very large models, in a desktop format.
A different approach: the 128 GB CPU-GPU unified memory allows loading models that even an RTX 5090 cannot accommodate, in a compact and quiet format, with the CUDA and Jupyter environment ready from the start.
DGX OS, ready-to-use ML environment
Discover this serverWhy a local workstation rather than the cloud
GPU cloud has its place, but for an ML engineer who works daily, a local workstation presents concrete advantages.
- Controlled cost. GPU hours in the cloud add up quickly. A local workstation is a one-time investment, paid off in one to two years of continuous operation.
- Instant iteration. No provisioning, no waiting for an instance, no data transfer. You launch your experiments immediately.
- Private data. Your datasets and your proprietary models stay with you, without going through a third-party provider.
- Total availability. No GPU quotas, no instance shortages, no outages. Your machine is always there.
- Stable environment. Your software stack doesn't change from one session to the next.
In short
How much VRAM does an ML engineer need?
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 runs lasting several days, yes: ECC protects against silent memory errors that can corrupt a run. RTX PRO 6000 Blackwell cards come equipped with it.
Does the CPU matter for ML?
Yes, for data preparation. A powerful GPU that is poorly fed runs idle. Many cores and generous RAM prevent this bottleneck.
Does fine-tuning require more than inference?
By far. 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 demand: PyTorch, CUDA, Hugging Face libraries, and Jupyter preconfigured (CUDA), so you can start coding right out of the box.
Can the configuration be customized?
Yes, entirely. The graphics card, CPU, RAM, storage, power supply, cooling, and case are configurable on every product page via the online configurator (excluding the GB10, which has fixed components). For a specific need or custom configuration, contact us at contact@radiancesystems.eu or via the quote form: we adapt the machine to your exact workload.
