AI and Machine Learning

ML and AI Workstations

In machine learning and artificial intelligence science, there is a great diversity of software programs and applications; a standard ML AI Workstation does not exist. Ikbenstil Computers creates a suitable custom solution based on the specific software maker’s recommendation, the available high-end hardware and the many years of experience of our engineers.

Recommended system requirements for ML AI Workstations

ML and AI hardware:

Processor (CPU): The CPU is generally subordinate to the GPU; GPUs are very important in ML AI science and have already led to many scientific breakthroughs. The processors do have an important role in the workstation, but that is mainly to provide the right platform to allow GPU calculations and simulations to run as quickly as possible. Intel with the Xeon and AMD with the Threadripper Pro and EPYC lines are the applied processors in the ML AI Workstation line. The clock speeds and the large number of cores, Intel up to 64 cores and AMD up to 128 cores, up to 12-channel memory, up to 3TB DDR5 ECC Registered memory, the large number of PCIe lanes, AVX512 instruction set, the bus speeds and many more features make the professional systems the most powerful processors, and are therefore often preferable to the regular Core and Ryzen types. Which CPU is the best choice for you depends on the application and software maker; our specialists can advise you on this.

The graphics card (GPU) The graphics cards are an essential component in ML AI Workstations. The most commonly used GPUs in machine learning and artificial intelligence science are the professional graphics cards from Nvidia, because of the Nvidia open-source Rapids software widely used in science. These Pro cards come in sizes from small to large, NVidia RTX 2000 Ada to the NVIDIA RTX PRO 6000 96GB Blackwell Max-Q Workstation Edition. Due to the compact width form factor of the large cards, it is possible to position them directly next to each other, allowing up to four GPUs to be used per workstation. This is much more difficult with the regular RTX 5080 or 5090 cards because the format does not allow it and it therefore remains limited to a maximum of two GPUs. A card worth special mention is the NVIDIA RTX PRO 5000 Blackwell with 72 GB GDDR7 VRAM. AMD also makes professional GPUs with the Radeon PRO and FirePro, which is supported by the AMD ROCm software stack. This software is suitable for AI, Large Language Models, image recognition science, etc. This software is promising, but not yet widespread in science. Also, the fastest AMD Pro GPUs are not as fast as those from Nvidia. We are closely following AMD and expect that much will happen here in the near future in the scientific field.

Memory (RAM) The application of these Xeon and AMD Threadripper PRO and EPYC processors has a great advantage in the memory area; a very large amount of memory can be placed on some mainboards up to 3TB; configurations with 256GB and 512GB are common. For Intel, the memory bus is a maximum of 8 channels and AMD EPYC a 12-channel memory bus. Stable working memory can also be used (Registered, RDIMM) and with the ECC function, which acts as memory error correction; the modules are often indicated as ECC Registered. These modules are produced up to a maximum of 256GB DDR5 each. In machine learning and artificial intelligence GPU Workstations, the rule is to place twice the total amount of VRAM as RAM; a configuration with 4x RTX PRO 6000 96GB needs 768GB RAM.

Storage For the operating system and programs, we choose a fast primary SSD. These come in different types, sizes and interfaces. For ML AI Workstations, it is decisive which application is required, and the associated hardware platform may have specific overviews of which SSDs, brand and type, are supported. Of course, storage for content is also needed, this can be on an SSD or optionally large volume HDDs, file server or a NAS.

Custom build Ikbenstil Computers has made workstations that align with standard workflows and the associated software. If your choice is not included here, you can contact one of our specialists, who can offer you a custom solution.