Data Science Workstations Comparison | Buyers Guide | SCAN UK

AI Development Hardware Buyers Guide

Development is the initial stage of any AI journey, prior to training and inferencing. As such it is seen as the most critical phase, as your development foundations set the scene for all future model training, fine-tuning and scaling of your projects. Mistakes made at the development phase can be very costly further down the line in terms of both expenditure and time, thus hardware selection to carry out development tasks is crucial to get right.

This guide takes you through the various AI development hardware options, explaining their differences and suitability for projects including small language models (SLMs), large language models (LLMs), generative AI, agentic AI and physical AI models.

Starting your AI Project

Although model development is the first of the three classic stages of an AI project, this phase is always preceded by a number of equally critical steps, including high-level planning, project scope setting and data preparation, as illustrated below.

Problem Statement

Project scope and high level ROI

Data Preparation

Classification, cleaning and structure

Model Development

Education and resource allocation

Model Training

Optimisation and scaling

Model Integration

Inferencing and deployment

Governance

Maintenance and compliance

It may not be immediately obvious, but ensuring your project scope is realistic and achievable has a large impact on what AI development hardware you’ll ultimately require. Similarly, undertaking adequate data preparation, including classification, structuring and cleansing of data, ensures you know the precise size of datasets you’ll be dealing with - thus providing an indication for GPU sizing, memory capacity and storage choices. Perhaps most crucially, data size also impacts the ability to scale your model as you iterate numerous options and evaluate outcomes. This data preparation stage can be up to as much as 60% of the entire time of your entire AI project, indicating how important it is to get it right. You can learn more about these crucial pre-development stages by reading our A-Z of AI White Paper.

AI Model Development

When it comes to using your prepared data in model development, chances are your project isn’t unique, so there may well be pre-trained frameworks or optimised foundation models (FMs). These save time and effort when building an AI pipeline, and removing the need for much of the subsequent development work. A prime example is NVIDIA AI Enterprise (NVAIE), which features in excess of 30 distinct and interlinked pre-trained frameworks, optimised for NVIDIA GPUs, designed for end-to-end implementation of AI projects such as medical imaging, autonomous vehicles, avatars, drug discovery, robotics, generative AI and many more.

Understanding the frameworks or FMs available to you may also have a significant impact on your AI development platform choice, as it may dictate a minimum GPU memory size or level of performance required to complete this phase within a given timeframe. When it comes to model size, some FMs may already number into the several billion-parameter space, before you start fine-tuning them into your own tailored (often even larger) model. It is therefore key to understand the relative size of models, how they might scale and the GPU hardware that will be capable of handling this. Just for context, between the 1950s and to 2018, AI model size grew by seven orders of magnitude (from 000’s to 30M) – yet from 2018 to 2022 alone, it has grown another four orders of magnitude (from 30M to 20B). Today 400B parameter models are not uncommon and ChatGPT-4 is rumoured to have 1.8T parameters.

AI Model Development

Mike and the CreativeBloke team feel that cloud-based workflows are invaluable for creative professionals seeking high-performance solutions without being tied to a fixed workspace. Freelancers, studios, and remote teams alike can benefit from its scalability and sustainability. As a next step, Mike intends to integrate AI-powered tools and explore collaborative opportunities by connecting multiple remote systems into a unified pipeline. The presumption is that this would further streamline workflows while enabling even greater creative freedom.

Type of AI Model	Typical Use Case	Parameters	Dataset Size	Typical GPU(s) Required
SLMs	Limited scope chatbots / NLP / On-device applications	100M - 7B	1 - 8GB	Consumer GPU(s) in laptop / desktop
LLMs	Content creation / advanced chatbots / Real-time translation	7B - 20B	10 - 15GB	Professional GPU(s) in workstation
Generative AI	Advanced content creation / Personalised experiences / Drug Discovery	20B - 70B	20 - 40GB	Multiple professional GPUs in workstation
Agentic AI	Personalised interactions / Data-driven insights / Autonomous vehicles	70B - 200B	60-150GB	Multiple professional GPUs in servers / cluster
Physical AI	Robotics / Digital Twins	200B - xT	200-750GB	Multiple professional GPU servers in cluster

It is worth clarifying that this table is for guidance only - absolute size (in GB) of any model is determined by number of parameters and the size of each parameter. Similarly, there will be small agentic AI models if their function is very focused and there will be very large LLMs if translation of many languages is the goal. It is also worth pointing out that the dataset size mentioned is the likely final size, so you need to consider capacity for numerous versions and many many iterations before reaching the final model. You can find more in-depth information about model development by reading our A-Z of AI White Paper.

AI Hardware

The following development options are compared and discussed in light of the model sizes above, offering recommendations and advice about the most suitable option for various scenarios. However, as previously stated thorough planning and scoping phases will lead to much more accurate provisional model sizes and better insight into hardware choice.

Traditional AI workstations are built around the industry-norm of an x86 CPU (either AMD or Intel) plus one or more NVIDIA RTX GPUs connected via the PCIe bus. The major benefit of this approach is that as a well-established system architecture over decades, such systems are extremely cost-effective and very easy to expand with more powerful processors, memory, storage and networking as your needs grow.

The downsides are that the CPU and GPU have separate pools of memory and that NVIDIA RTX GPUs are currently limited to 96GB of memory, so require multiple cards to run generative AI and reasoning AI models effectively. A new approach, using system-on-chip (SoC) provides a single CPU/GPU resource connected to a large combined memory pool.

Click the tabs below to explore these options, or of course you can contact our AI team for more information or advice.

NVIDIA DGX Spark

The DGX Spark is NVIDIA’s latest appliance for developers looking for a desktop solution to develop and fine-tune generative AI and reasoning AI models. These are particularly challenging workloads as they have a large memory footprint, so are too large to run on many existing GPUs. The DGX Spark has been built from the ground-up to provide a single large memory pool. It achieves this by replacing the old discrete CPU and GPU paradigm with a SoC, combining CPU and GPU cores together in a single unit, making the unit very compact despite its considerable performance.

Architecture

Known as the GB10 Grace Blackwell Superchip, this SoC comprises a Blackwell GPU and 20 Arm CPU cores, sharing a unified 128GB of memory. The GPU element of the SoC is equipped with the latest 5th gen Tensor cores, which are optimised for performing FP4 calculations, the most commonly used precision level for generative AI and reasoning AI models, while the CPU element has ten Cortex X925 and ten Cortex A725 cores.

Connectivity
Wi-Fi, Bluetooth, USB

ConnectX
NCCL, RDMA, GPUDirect

4 TB SSD

GB10 Superchip Blackwell GPU
1 PetaFLOP FP4 AI Compute

Grace GPU
20 Arm Cores

High Bandwidth Unified Memory
128 GB Low Power DDR5X

The SoC is supported by 4TB of NVMe SSD storage, alongside a 10GbE Ethernet port, ConnectX-7 SmartNIC and WiFi 7 for ingesting data. Just like a typical desktop PC, the DGX Spark has a HDMI port to connect a monitor, plus four USB Type C ports to connect peripherals. Furthermore, two DGX Spark units can also be clustered using a special LinkX cable, effectively creating a combined 256GB memory pool for running extremely larger models.

The DGX Spark runs DGX OS, which is a NVIDIA-customised version of Ubuntu Linux. This enables developers to unleash the full potential of their DGX Spark with a proven software platform built around a complete AI software stack including access to NVIDIA NIM microservices, Blueprints and AI Enterprise. This is of course the same software stack as other NVIDIA AI appliances, making it easy to transfer your newly-developed AI model onto more powerful MGX, HGX or DGX systems, whether they’re in your own datacentre or in the cloud.

Relative Performance & Capability

The DGX Spark has an FP4 performance of 1,000 TOPS, which as you can see in the table below, isn’t the highest - however with a GPU memory capacity of 128GB (or 256GB if you combine two Sparks), it clearly has the best potential for natively and cost-effectively developing extremely large AI models.

System	NVIDIA DGX Spark	3XS AI Laptop with GeForce RTX 5090	3XS AI Workstation with GeForce RTX 5090	3XS AI Workstation with RTX PRO 6000 Blackwell	3XS AI Workstation with RTX PRO 6000 Blackwell Max-Q	NVIDIA DGX Station GB300
AI performance per GPU (FP4)	1,000 TOPS	1,824 TOPS	3,352 TOPS	4,000 TOPS	3,351 TOPS	20,000 TOPS
Memory per GPU	128GB	24GB	32GB	96GB	96GB	784GB
AI model size per GPU (FP4)	200 billion	38 billion	51 billion	153 billion	153 billion	1.2 trillion
GPU(s)	1	1	Up to 2	Up to 2	Up to 4	1
Maximum AI model size (FP4)	400 billion (2x DGX Spark)	38 billion	102 billion	307 billion	614 billion	1.2 trillion
Cost	£	£	£	£££	£££	£££££

Conclusion

Thanks to its innovative SoC architecture, the DGX Spark is the only reasonably-priced desktop device that can develop AI models with a large number of parameters, such as generative AI and reasoning AI models. Being a desktop device, the DGX Spark is also a practical choice for organisations that are not able to upload their data into a cloud service. Having said that, if performance is more important to you than outright GPU memory, you should consider either an a (link to tab 2) or 3XS AI Development Workstation (link to tab 3).

Ready to buy?

Click the link below to view the range of AI development solutions. If you still have questions on how to select the perfect system, don't hesitate to contact one of our friendly advisors on 01204 474210 or at [email protected].

DGX Spark

3XS AI Development Laptop

3XS AI Development Laptops are custom-built for flexibility whilst maintaining the ability to effectively develop and debug your AI projects. Featuring an 18” QHD+ screen, an x86 CPU combined with the latest NVIDIA Blackwell GPU enabling you to identify the MVP (Minimum Viable Product). This mobile platform is ideal to learn valuable insights from failed models before rolling out your code to a compatible NVIDIA training system, such as an MGX, HGX or DGX server.

3XS Systems has been hand crafting PCs, workstations and servers for more than 30 years. We pioneered the AI dev box back in 2016, so we have a huge amount of experience in building highly reliable systems that deliver the most performance for your budget. Here are the key reasons to trust a 3XS AI Development Laptop.

NVIDIA Elite Partner

Scan has been an accredited NVIDIA Elite Partner since 2017, awarded for our expertise in the areas of deep learning and AI.

AI Optimised

Our in-house team includes data scientists who optimise the configuration and software stack of each system for AI workloads.

Trusted by you

Scan 3XS Systems AI dev boxes are trusted by organisations including the NHS, University of Liverpool and University of Strathclyde.

7 Days Support

Our technical support engineers are available seven days a week to help with any queries.