Mobile, desktop and cloud — the right tool for the right job
When the pandemic hit and the whole world virtually shut down I, like many others, ended up working from home. I was the lucky one — still had a job and the ability to do it from home and for that I am grateful! Two things, however, made this transition much less painful work-wise and both came courtesy of Lenovo.
Full disclosure: I am not sponsored nor paid by Lenovo in any way to write this article. I simply decided that I wanted to share my experience working on two systems that Lenovo makes and was kind enough to lend me so I could finish some of the work I have been doing on COVID-19 scientific literature analysis; you can read about it here: https://towardsdatascience.com/building-an-understanding-of-viruses-by-mining-covid-19-scientific-corpus-f1d27c33b48.
Mobile or desktop? Why not both!
I was lucky to test two workstations: the Lenovo ThinkPad P53, and the Lenovo ThinkStation P920. Both are beast of machines with so much much to offer and having both provides so much flexibility.
The ThinkPad P53 offers 15" touch 4K screen, Xeon CPU, 64GB RAM (configurable up to 128GB) and 2TB NVMe SSD drive in a package that might be a bit heavier than your regular, off-the-shelf 15" system, but still packaged into a very versatile and convenitent form factor. The versatility stems also from the fact that the model I got for testing was equipped with NVIDIA Quadro RTX 5000 with 16GB of VRAM! That’s 3072 CUDA cores, 384 Tensor cores and a whole lot of fun! On the go nonetheless!
On the other hand we have a true workstation: in a form factor that might be perceived as slightly larger tower design, a behemoth of desktop computers sits inside: dual-Xeon Gold 6226 CPUs (with 24x cores / 48x threads, configurable to max. 56 cores / 112 threads), 192GB of PC4–2933 ECC RAM (with this machine you can get up to 2TB of PC4–2933 ECC RAM) and 1TB of M.2 NVMe SSD drive (up to 64TB of storage, combination of NVMe, SSD and HDD). But as you might think the beauty of this machine does not only come from the CPU side: the model I tested came equipped with dual NVIDIA Quadro RTX 8000 GPUs with a whopping 48GB of VRAM each (you can configure your system with up to 3 NVIDIA Quadro RTX 8000s!!!) And this 96GB of VRAM monster of pure GPU power, with over 9216 CUDA cores (4608 in each card), 1152 Tensor cores (576 in each card) is sitting quietly under my desk and I can barely hear it. Seriously, even my older box with i5 CPU and NVIDIA TITAN RTX is louder than this…
One thing that I also need to mention that makes these systems ideal for data scientists or machine learning practitioners is the operating system. The computers came pre-packaged and pre-configured with everything you need to get you going from the get go. Both of them came with Ubuntu 18.04 and Docker images of RAPIDS and the most popular deep learning preinstalled: if your tool-of-choice is Tensorflow — it’s there, prefer PyTorch — off you go, need more Caffe — served within! No need to spend time downloading CUDA, installing drivers and downloading Docker images. You can literally start building models within 10 minutes of opening the box.
Why do I need both?
You may be asking the same question. Well, that was my initial reaction, too… However, I quickly realized how one complements the other.
The mobile workstation is precisely that — a beast on wheels! I can take it with me wherever I go and never miss a beat. Travel these days is limited but for anyone who depends on GPUs to deliver results it offers the flexibility of working wherever. Hours in an airplane seat? No problem! Commuting to work on the bus or train and tired of constantly losing the connectivity? It can handle this! Camping with family in a remote location? Wait — should not be working… Spend the time with your family. Still, it’s another scenario where this machine can help deliver so much more productivity with limited access to either desktop or cloud computing: plug it in and off you go. Just be aware that it comes with a 230W PSU and drains quite a bit of energy. So do plug it in if you do some serious work.
While the NVIDIA Quadro RTX 5000 with 16GB of VRAM would make even a really decent desktop blush, you probably won’t be working on enormous datasets going into hundreds of GBs using it. P53 is not a replacement for a desktop station — it completes it. Having access and be able to sync between both stations, I found that I can deliver results much quicker and have more time to spend with my family. How? Simple. I can write code and test it while on the road on a smaller, sampled version of the dataset while still have access to a powerful GPU. I can then sync that code with GitHub when connected and deploy it on a larger data on the P920 workstation. I truly enjoyed this experience and will definitely be missing the P53 and the P920.
Why do you need desktop then?
While the mobile workstation gives the flexibility and access to a powerful GPU in a mobile form factor, the brute power is delivered by the desktop. With up to 96GB of GPU RAM, and almost 10k of CUDA cores, training Deep Learning models, or running RAPIDS distributed workloads, I can tackle larger datasets to further refine my approach before training my model in the cloud.
In the same manner as the mobile workstation complements the desktop, the desktop can symbiotically co-exist with on-prem or cloud infrastructure: data scientists equipped with these machines can build and test the models on their desktops, changing and tweaking the data features or the model architecture independently, in the process saving the cloud or on-prem resources for training the large model. This way you can train and tweak more models, and since you have more resources available in the cloud or on-prem, these will finish training quicker as well. Win-win!
For me, personally, the P920 delivered the power required by the work I was doing at the time. Shortly after the pandemic hit, the cloud resources I had access to got more scarce as customer behavior changed (just think of all the online meetings we have held over the time) as well because of the increased demand for compute resources aimed at helping finding the cure or the vaccine for COVID-19. The P920 and P53 perfectly fit within my workflow: start small wherever, refine on a larger data at home, and then train in the cloud only when truly ready.
How much?
These ain’t cheap. But, these are not systems for an average Joe… These are refined pieces of machinery aimed at a specific target of customers and their requirements: data scientists or machine learning practitioners to help them be more productive no matter what or where.
Is it worth it? It’s a question each company needs to answer. However, here’s my two cents: simply yes. If you can increase the productivity of a data scientist by 20%, you are saving anywhere between $30-$100k a year. With an estimated retail price of the desktop workstation I tested is about $30k, over the lifetime of this system (assuming 3 years), a company may be getting back anywhere between $60k–$270k in increased productivity. Not bad ROI. At the same time, the faster time to market for products or services your company offers, can result in more revenue opportunities and better market standing/competitiveness.
As for me — yes — I will miss these two systems!
