How the F*⚡$ Does Nvidia Name GPUs?!

Recently, I was trying to figure out the best GPU to use to train some computer vision models that we’re starting to build on my team at ShopRunner. As a relative newcomer to GPUs, I saw a forest of incomprehensible names and acronyms everywhere I looked and kept asking myself the profane question that is now the title of this article. After plowing through dozens of sources and Wikipedia pages, I figured I’d write the single guide that I wish existed when we started this project. So here it is.

What’s in a name?

Probably the most important thing that I eventually realized is that Nvidia uses the name “Tesla” in two completely different contexts. The first is as a so-called “microarchitecture” — one in the series of scientist names that they use for microarchitectures. In chronological order, the microarchitecture generations are Tesla, Fermi, Kepler, Maxwell, Pascal, and Volta (see this incredibly useful Wikipedia table). Apparently the next generation microarchitecture will be the Turing. The scientist microarchitecture generation name tends to make its way into the individual GPU name as a single letter abbreviation (so there are K’s, M’s, P’s, and V’s).

The second way in which Nvidia uses “Tesla” is as a brand name for a whole line of workstation or general-purpose computing GPUs (or GPGPUs). So indeed, it’s possible to have cross-scientist pairings: a Tesla K80 (released November 2014) is a Tesla built on the Kepler microarchitecture, a Tesla P100 (June 2016) is a Tesla built on the Pascal microarchitecture, and the Tesla V100 (June 2017) is a Tesla built on the newest Volta microarchitecture. Judging from the Wikipedia table, the Tesla C1060 from 2009 is the last Tesla built on the actual Tesla microarchitecture. I guess they just decided to recycle the name?

These GPGPUs tend to be the GPUs that you find on many of the cloud providers. The AWS p3 family of instances features Tesla V100’s, and the older p2 family features Tesla K80’s. I was initially thrown by the fact that AWS decided to use the letter “p” for the GPU instance family — it’s a red herring and isn’t related to the Pascal microarchitecture, as you might have guessed.

Consumer Confidence

The second concept that was most useful for me to realize is that Nvidia generally breaks GPUs down into consumer GPUs for desktop computing, graphics, and gaming (but which are also extremely useful for deep learning) and the GPGPU/workstation Tesla’s that I described above, which aren’t targeted at graphics. See this Wikipedia table for a full listing.

Nvidia’s consumer GPUs go under the brand name “GeForce” and have a systematic naming scheme. The best explanation I could find came from Samantha Jaras’s excellent post from 2016:

The easiest way to break down this naming stuff…is to relate it to something that might be a bit more common, like cars. The first thing that you notice about the car is the company brand. Just like General Motors owns Chevrolet, Cadillac, and Buick; NVIDIA owns the GeForce name.

The next part about cars is the body style. Is it a Sedan, SUV, or pickup truck? Each style can handle things a little bit differently. This is what the string of capital letters after the GeForce are. Their sedan of graphics cards has no letters after the GeForce. This version is perfect for everyday use and light video watching in HD. Nothing crazy. The SUV version is called the GeForce GT. So this card is supercharged for better HD quality, 3D movies, and entry level gaming capabilities. Now for the biggest, most heavy hitting line: The GeForce GTX. This graphics card is the ultimate GPU, and provides high performance for any gamers [sic] needs, including VR capacity. This version is the newest and most popular line on the market right now.

Since then, the GTX Titan has emerged as the real heavy-hitter. So I guess that’s the semi-truck of GPUs?

In addition, there are numerical generations for the consumer GPUs. So there’s a 600 series, a 700 series, a 900 series, and a 10 series. (See this and this…also, no idea what happened to the 800 series). So, for example, the top of the GeForce 10 series are the GeForce GTX 1080 (May 2016), the GeForce GTX 1080 Ti (March 2017), the Nvidia TITAN X (August 2016), and the Nvidia TITAN Xp (April 2017), which are all based on the Pascal microarchitecture. The Nvidia TITAN V (December 2017) is the newest heavy-hitter based on the newest Volta microarchitecture.

So which should I use?

Well, that’s beyond the scope of this article, but hopefully I’ve equipped you to understand the alphabet soup of names. Check out these posts for thoughts on which to choose — 1, 2, 3, 4.

Note: If you’re a true expert in all of this and I’ve gotten something wrong or missed something, please leave a comment, and I’ll add a correction.