It may be surprising to hear, but the core concepts behind artificial intelligence have been around since the 1950s. What hasn’t been around until recently, however, is the hardware necessary to finally take full advantage of some of those concepts to provide tangible, real-world benefits for the average business. While exciting, this creates many new challenges for companies both big and small.

As a subset of machine learning, deep learning relies heavily on data. In essence, you’re using the past (i.e., data) to provide educated guesses about the future (i.e., insights). For instance, a business could predict supply chain issues, machine failures, or cyberattacks, and work toward solving those problems before they start. It can also be used to predict things like marketing trends or customer behavior. This is all done by training models via an iterative optimization algorithm.1 It may seem intimidating at first, but conceptually it’s actually very simple

Example of a DeepNeural Network:


See? Simple.

That said, no matter how extensively trained, models are only as good as the data you give them. If the datasets are old, or not conditioned properly, no amount of training will fix it. To get a better model, you need to optimize your data (e.g., regularization, algorithmic tuning, etc.), as well as provide new, clean data that must also be tested and trained in order to continue to obtain accurate and actionable results. It’s an iterative and cyclical process. Data is the foundation of machine learning; if you lay a shoddy foundation, the entire process suffers. For example, labeling and annotating data in the initial storage phase helps to streamline the process later.

As expected, data scientists are the ones doing much of the legwork in this new arena—but they can’t do it alone. They need IT leaders and others to help redesign their infrastructures in a way that is geared toward ingesting, conditioning, training, and validating very large datasets. That involves asking some important questions. For example, do you want inferencing to run locally, or on the cloud? Locally will be faster, as well as alleviate many security and privacy concerns that are inherent to cloud computing.

However, neural networks can run in perpetuity on dedicated systems off-site, and can be updated with better models more easily. It’s up to you to decide which strategy is best for your business.

Whatever your business model, it’s vital to start off on the right foot. To do that, you’ll need hardware designed for dense compute and low-latency workloads. Specifically, you’ll want a processor with a high core count and a large CPU cache. The larger your cache, the fewer trips the processor must make to the memory for the next instruction. CPUs can handle trillions of calculations per second, and a larger cache translates to higher throughput. High memory capacity and bandwidth are also important for very similar reasons, as is the unmatched parallel processing capabilities of professional-grade GPUs.

If you’re looking for a compact, ultra-fast, deskside workstation to test, train, and visualize large datasets, the APEXX W3 from BOXX is a perfect match. It houses a single 18-core Intel® Xeon® W processor and can hold up to 512GB of memory, as well as three full-size GPUs. For optimal performance for data science projects, BOXX recommends two NVIDIA® Quadro RTX™ 8000s which provide a staggering 96GB of GDDR6 memory and a total of 1,152 hardware-accelerated Tensor Cores for unprecedented performance.

If you require even more cores, the APEXX D4, featuring dual 2nd gen Intel® Xeon® Scalable® processors2—with higher memory bandwidth3 and space for four full-size GPUs—will blast through any data science project you throw at it. Both BOXX models are purpose-built, ready-made solutions that come pre-loaded with AI frameworks,4 as well as GPU-acceleration libraries based on NVIDIA’s CUDA-X architecture that unleashes the full potential of their Quadro RTX™ cards. These tools come preconfigured for your workflow and simplify the AI development process significantly so you can hit the ground running and iterate faster.

Still not sure which model is best for you? Consult with a BOXX performance specialist. As a trusted workstation manufacturer for over 23 years, BOXX is doing its part to put companies on the right track.


1 Stochastic gradient descent, for the record. Say that five times fast.

2 Up to 56 cores per CPU.

3  Up to 4TB of DDR4-2933 ECC memory across six memory channels.

4  RAPIDS suite; Docker 18.06.1, NVIDIA Docker Runtime v2.0.3; TensorFlow, PyTorch, and Caffe 2.