When evaluating workstations, CPU and GPU horsepower for CAD, BIM and visualisation are no longer the only benchmarks that matter. AEC firms should consider each machine’s AI capabilities as well.
When architecture, engineering and construction (AEC) firms evaluate new workstations, the questions are somewhat predictable. How well will they run Autodesk Revit®? Will they be able to handle large multi-disciplinary BIM models? Will they be fast enough for real-time visualisation? Will they keep pace with increasingly complex multi-application workflows?
Those questions still matter. But there’s one more worth adding to the list: how well will the workstations run AI?
AI’s impact on AEC is already showing up, and things are moving unbelievably fast. But in an industry built on multi-year projects, nothing’s going to be replaced overnight. Instead, AI is quietly working its way into the tools firms already rely on for CAD, BIM, visualisation and reality modelling, while entirely new AI-powered applications and workflows are emerging alongside them.
For firms investing in workstations today, that raises an important point: the hardware they buy to support their AEC teams over the next three to five years may need to handle some significant AI workloads as well.
AI is becoming part of the AEC workflow
Most of the discussion around AI in AEC so far has been about the cloud. This includes frontier large language models (LLMs) like OpenAI’s ChatGPT and Anthropic’s Claude, image generators like Google’s Gemini (Nano Banana), and a growing list of AI-powered collaboration platforms and specialist AEC tools accessed through a web browser.
Cloud AI will undoubtedly remain an important part of the AEC technology landscape. But more and more AI workloads can now run directly on the workstation, bringing new efficiencies to existing desktop workflows and helping uncover new insights from CAD, BIM and other project data.
These include coding assistants, automation tools, image generators, and open-weight LLMs. In AEC, AI is progressively being used to automate repetitive BIM tasks, check models for code compliance, identify coordination issues, assist with scripting, generate visualisations, and even build agents.
Running AI locally has real upside for AEC firms. Performance is more predictable, particularly when working with large datasets or iterative workflows. Sensitive intellectual property, client information and project data remain under tight control, and there may also be cost benefits overpaying for subscription or usage-based cloud AI services. But local AI also introduces a new set of demands on workstation hardware.
AI changes the hardware equation
Workstation specs have traditionally been defined by three things: CPU performance, GPU performance and system memory. That model still applies, but AI adds another dimension.
A lot of AI models are memory hungry. Running a sophisticated local language or generative model can eat up tens of gigabytes of memory, depending on its size and precision. That can be a real problem for conventional workstation architectures, where CPU memory and GPU memory are two distinct pools. A professional GPU might offer exceptional graphics horsepower but still choke on a big AI model simply because its VRAM is capped. And it can be even worse on mobile workstations, where laptop GPUs typically have far less VRAM than their desktop equivalents.
This comes at a time when AEC workflows are already demanding. Few architects or engineers work in a single application all day. More likely, it’s Autodesk Revit or McNeel Rhino™ running alongside a real-time visualisation tool such as Chaos Enscape™ or Epic Games Twinmotion™, a browser full of cloud services, collaboration platforms, and office tools. Throw reality modelling or computational design into the mix, and the workstation soon needs to juggle several applications at once.
In short, AI just piles onto workflows that are already demanding — and that’s exactly where a new approach to workstation architecture could prove significant.
A different workstation platform
The AMD Ryzen™ AI Max PRO platform, found in machines like the new HP ZBook Ultra G3a mobile workstation and HP Z2 Mini G1a desktop workstation, takes a genuinely different approach.
Instead of treating the CPU, GPU and AI accelerator as separate components, integrate them within a single system-on-chip (SoC), combining high-performance Zen 5 CPU cores, solid AMD Radeon™ integrated graphics and a neural processing unit (NPU), all sharing a single pool of high-bandwidth unified memory. That architecture can be particularly efficient for workloads that need to move data between CPU, GPU and AI accelerators. And, for AI specifically, memory capacity can matter as much as raw processing performance.
The importance of unified memory
Perhaps the most distinctive feature of the Ryzen™ AI Max PRO platform is its flexible memory. The new HP ZBook Ultra G3a, for instance, can be configured with up to 192 GB of unified memory, but instead of having a fixed amount for graphics, the system can dish out memory dynamically based on what the workload actually needs – up to a whopping 160 GB specifically for the GPU.
In AEC, AI is progressively being used to automate repetitive BIM tasks, check models for code compliance, identify coordination issues, assist with scripting, generate visualisations, and even build agents
That could prove particularly relevant as local AI models become larger and more capable. A conventional discrete GPU might have plenty of processing performance, but not enough dedicated VRAM to even load the model, whether that’s simply querying a local LLM or developing a demanding autonomous agent. A unified memory pool gets around that problem by making substantially more memory available to graphics and AI workloads when needed.
That same flexibility helps everyday AEC workflows as well. Large BIM models, complex geometry, high-resolution textures, visualisation scenes and AI models are all fighting over system resources — and dynamically allocating memory is more flexible than permanently splitting it between system and GPU.
That said, firms must be realistic about what processors can do. They should evaluate it against the applications their teams use, since different CAD, BIM and visualisation workflows all have their own processor and graphics requirements. Discrete professional GPUs still matter a lot, especially for demanding visualisation and certain reality modelling work. AMD processors can handle very large models, but it’s not going to out-muscle a high-end discrete GPU when it comes to raw processing power.
Buy for today, but prepare for tomorrow
No workstation can genuinely be considered future-proof — software is changing, AI is moving fast, and no one really knows what tools the industry will be using in five years. Still, it’s hard to argue against giving serious thought to future workloads when buying.
When evaluating professional hardware AI readiness should now sit alongside CPU performance, GPU capability, memory capacity and application certification. That doesn’t mean AEC firms should abandon the old checklist — CAD, BIM and real-time visualisation performance still matters. But it’s worth thinking ahead to what happens when AI needs to run alongside those applications, when teams begin searching large document collections or BIM models with local LLMs, or when AI-powered automation gets baked into desktop BIM workflows.
The workstation is fast becoming the hub of a more complex computing environment. For practices investing in new hardware, workstations such as the HP ZBook Ultra G3a and HP Z2 Mini G1a make a compelling case. Built on AMD they bring together strong CPU performance, capable integrated graphics, dedicated AI acceleration, and an unusually large, flexible unified memory pool in one tightly integrated package — designed to handle today’s AEC workflows with a fair bit of headroom for whatever AI throws at it next.
So, the question isn’t just “How fast will this run my CAD, BIM and viz software?” anymore. It’s “What else will this workstation need to do over its lifetime?” AI is likely to be a big part of that answer.
