The need for speed is something that computer professionals and enthusiasts will probably never lose. We’re always eager to hear about the latest high-end components. Which is good, because they don’t come much more high-end than the new Nvidia Quadro GP100.
If you’ve never heard of Quadro products before, these are the professional line of GPUs meant for people into serious graphics applications.
It’s not so much that Quadro cards are more powerful than standard Geforce cards, but that they are made with professional users in mind. One place where these cards definitely blow “normal” cards away is in price.
It’s not unusual for a high-end Quadro (or FirePro for AMD) to cost as much as a used car. You could easily blow $6000 on a single card before you’ve bought any other component of your computer.
What is it Good For?
So what are you paying for? In a word: reliability.
Quadro users get a lot of perks from Nvidia when it comes to customer support. Nvidia works closely with companies like Adobe and AutoDesk to make sure that their drivers work perfectly with them.
The customers who want Quadro cards understand that their time is money. The high price of the hardware pays for itself very quickly when large rendering or simulation projects don’t get stalled because of wonky hardware or software.
Paying for a Quadro is therefore not just about performance, but about support and professional projects that can’t afford downtime.
The Tesla Effect
To understand where the GP100 suddenly came from, you need to know about one of Nvidia’s other product lines: Tesla.
Tesla cards contain GPUs just like GeForce and Quadro cards, but you won’t find anywhere to plug in a monitor. These cards are optimized to do nothing but crunch numbers for high-performance computing work.
We first saw the GPU that would drive the GP100 in a Tesla card dubbed the “P100”. While Nvidia isn’t all to creative with the names of these business-oriented units, you can believe that they have eye-bleeding performance.
The Party Trick
So what makes this new Pascal-based card so special? One of the key features of the P100 is its ability to handle double precision floating-point operations at high-performance levels.
This is important for the type of engineering work something like the P100 would be used for, where even tiny rounding errors can have significant and even catastrophic results.
The P100 was designed to deal with double precision calculations without breaking a sweat and the Nvidia Quadro GP100 has inherited that same silicon heart.
This means that serious professionals can now buy a desktop workstation card that can actually reach useful performance levels at double precision levels. Although you could process these types of jobs on some previous Quadro products, they were pretty terrible at it and performance levels fell drastically.
Of course, not that many people absolutely need this. So the single precision performance levels also matter. Luckily the GP100 is not a slouch there either.
Reading over the official Nvidia specs for the Nvidia Quadro GP100 is an exercise in trying to keep your jaw from dropping. It has 16GB of HBM2 memory, which is a big jump from the 4GB limit we saw in the first generation of this super-wide memory technology. This RAM has a 4096-bit bus providing up to an insane 717 gigabytes per second of bandwidth.
Even more impressive is the fact that you can connect two GP100 cards together and have access to 32GB of HBM2. That may sound obvious, but multi-GPU setups have not allowed for the combination of RAM before.
For example, when you play video games using SLI technology, each card’s RAM has a duplicate of the current data in it. The NV Link technology and HBM2 bandwidth are so fast that the cards can share their onboard memory and add the totals together.
That’s a pretty big deal because other workstation cards are exceeding the 16GB mark, which is a technological limit. There are definitely workloads out there that need more than 16GB of GPU memory, so this removes any excuses to steer away from the Nvidia Quadro GP100 on that count.
Central Processing Zone
The Pascal-based GPU itself has a staggering 3584 CUDA cores onboard and eats a very reasonable 235W at peak. It’s also only a dual-slot card, which makes it about as hefty as a Titan XP, which it handily outperforms.
It’s not just rendering muscle either since you can drive four 4096×2160 monitors at 120Hz or 4x 5120×2880 screens at 60Hz. All thanks to the four Displayport connections on the back. Surprisingly there’s also a DVI port, which they removed in the new 1080 Ti to improve airflow.
The main card that this has to be compared to is, of course, the Nvidia P6000, which is the current flagship Quadro card. On paper, the P6000 can actually look better than the Nvidia Quadro GP100.
For example, the GP100 has fewer CUDA cores and has a faster boost clock. It has more texture units and much faster memory speed. Most importantly it has eight more GB of RAM.
Those comparisons are not so straightforward, however. The GP100 can do more per CUDA core and trounces the P6000 in terms of bandwidth. Also, the ability to combine the memory of two Nvidia Quadro GP100 cards solves that final issue. Albeit in an expensive way.
The Final Frontier
The GP100 represents the pinnacle of what Nvidia can do with a consumer product today. The numbers this beast can put out are just staggering. The Nvidia GP100 can stand up to four old titans in quad-SLI.
Let that sink in for a moment as you think about how much performance they’ve squeezed into that tiny 16nm FINFET chip. If you have the bones to pay for it, there can be little doubt that the GP100 is set to rule the roost for a while.
If however, you’re a little strapped for cash our cheap VR guide may help stretch those dollars a bit. On the other hand, if you are a high-roller and need the perfect accessory, check out my article on ultrawide monitors.
All Images Official NVIDIA Product Shots (Fair Use Rationale)