March 11 2020

NVidia Volta

AMD and NVidia prove that competition can be a catalyst to innovation. While AMD provides some of the best-priced consumer GPUs, NVidia seems to excel in designing micro-architecture that is suited for workstations and supercomputers.

One such example of NVidia’s prowess is its Volta architecture, which features in both the Lawrence Livermore National Laboratory’s Sierra supercomputer and IBM’s Summit computer. For many enthusiasts, NVidia’s Volta range flew under the radar.

It seems that many users do not even know that it was released over two years ago. In this guide, we’ll go over Volta’s history, explore its uses, discuss future developments and much more.

What is NVidia Volta?

We find that the average gamer is not very au fait on the intricacies of GPU architecture. While advanced PC gamers and hobbyists tend to be overly pedantic about each part that goes into their PCs. In most cases, it’s enough to win a pissing contest by announcing that you have the latest and greatest graphics card for gaming and how much you’ve overclocked it.

Everything else is superfluous. However, there are gamers who are genuinely interested in GPU micro-architecture and its impact on performance and capability. We suspect that you are obviously one of those gamers.

So what is NVidia’s Volta and what makes it superior to its predecessors? If you’re running a Pascal or Maxwell-based card, is it worth trading it in for a Volta card? The answer to this question may surprise you. First, let us go over Volta’s history.

NVidia Volta History

On the 19th of March 2013, during Nvidia’s annual GPU Technology Conference (GTC), the company’s CEO, Jen-Hsun Huang gave what was one of the most important keynote addresses of the decade (in terms of GPU technology).

For those of you who are unfamiliar with the GTC, it is a global conference covering some of the most important topics, not just related to NVidia, but in all spheres of computing and technology. 2020’s conference will feature topics on AI, IoT, deep learning, simulation, etc.

The conference will be held in the San Jose McEnery Convention Center on the 26th of March. You can learn more about it from NVidia’s GTC page.

Nevertheless, Jen-Hsun Huang’s keynote address was particularly important because he went into great detail about NVidia’s plans for the future. He laid out a roadmap illustrating NVidia’s plans for the next five years.

The first revelation was Kepler’s successor, dubbed Maxwell. Maxwell can be remembered for making graphics cards like the GeForce GTX 750 Ti and the GTX Titan X possible. Surprisingly man gamers still use cards with Maxwell micro-architecture.

From the keynote address, it was understood that Volta would be the one to replace Maxwell. Retrospectively, we can now see that this was false, as GPUs based on the Pascal architecture came out roughly two years later.

We’re not complaining though because Pascal made great graphics cards such as the GTX 1050 Ti, GTX 1070 and the GTX 1080 Ti possible.

On average, NVidia tends to release new micro-architecture every two years. Surprisingly enough, the Volta architecture was released just after a year of Pascal’s release. There’s a good reason for this and we’ll explain further down this guide.

Volta Release Date

During a keynote address on NVidia’s 2017 GTC, Volta was officially announced with the Tesla V100. If you’ve never heard of NVidia’s Tesla range, then you may not be the intended audience (unless you’re a data scientist/computational physicist or any other scientist that works with large sets of data).

It makes sense that the architecture was named after Alessandro Volta, an 18th – 19th-century Italian chemist and physicist, known for his invention of the electric cell and discovery of methane. This was a continuation of NVidia’s tradition of code-naming their micro-architectures after famous scientists.

NVidia’s Tesla range was of course named after everyone’s favorite reclusive inventor, Nikola Tesla. The Tesla V100 wasn’t made for gaming, more for stream processing, computation, geospatial intelligence, simulations for complex calculations and generating visual objects and images for various professions.

The Tesla V100 would be the first card to bare Volta’s architecture on the 1st of June 2017. Later that year, on the 7th of December, the NVidia Titan V was released. Unlike the Tesla V100, the NVidia Titan V came with full graphical processing capabilities.

In other words, you could use it for gaming (if you were very rich that is). None of NVidia’s Volta graphics cards could be considered mid-tier or affordable. At the time of its release, the NVidia Titan V was considered to be the most powerful graphics card. It was priced at $2,999 at launch.

Volta’s Features

Volta was NVidia’s introduction to Tensor cores, which are known to provide superior deep learning capabilities than regular CUDA Cores.

Yet still, the Volta architecture still features CUDA Compute Capability 7.0. CUDA Cores are essential for General-Purpose computing on Graphics Processing Unit, as they enable parallel computing.

The architecture is manufactured through TSMC’s 12 nm FinFET process which allows 21.1 billion transistors. NVidia has always been a testament to Moore’s Law.

It features the second-generation High Bandwidth Memory, HBM2. Which allows up to 8GB (space) and 256 GB/s (speed) memory bandwidth per package. This was very impressive at the time until HBM2E came out a few years later.

Volta also comes with NVidia’s PureVideo, which supports hardware decoding of various video code standards. In addition to this, Volta uses NVLink 2.0, which delivers superior performance and speed when compared to PCIe. Unfortunately, this feature is disabled on the NVidia Titan V.

NVidia’s aspiration for Volta was to create microchips that would kindle the fire of Artificial Intelligence and advance it. To make this possible, Volta chips come with 640 Tensor cores which help deliver over 125 TFLOPS per second.

Of course, you also get your choice of Volta-optimized software and kits. Like CUDA APIs and NVidia Deep Learning SDK libraries.

Volta Graphics Cards and Products

In this section, we’ll cover some of the best technology and products using the volta micro-architecture. If you’re interested in purchasing a graphics card, this section will give you an impression of what’s available to you. Additionally, it should also help you understand the possibilities and applications of the Volta microarchitecture.

NVidia V100 PCie

NVidia V100 PCie upclose

  • Release Date: 21st September 2017
  • Price: $5,923 – $11,458
  • NVidia Tensor Cores: 640
  • NVidia CUDA Cores: 5120
  • GPU Memory: 16GB/32GB HBM2
  • Base Clock: 1246 MHz
  • Boost Clock: 1380 MHz
  • Slot Size: Dual-Slot
  • Single Precision Performance: 14 TFLOPS
  • Tensor Performance: 112 TFLOPS
  • Max Power Consumption: 250 W
  • Best Implementation: PNY Nvidia Tesla v100 16GB

As we’ve previously mentioned, the NVidia V100 was the first graphics card to display NVidia’s Volta architecture. It uses the GV100 graphics processor. Until this day, it still delivers one of the best computational performances. According to NVidia’s tests, it’s 32 times faster than the average CPU.

It supports nearly every deep learning framework. From Caffe2 to Pytorch and MXNet. Upon release, the average price for this particular was $10,644 for the 16GB version and $11,458 for the 32GB version. Today you can get the 16GB version from PNY for less than $6,000.  HP also sells their own version for $5,999 (at the time of writing this article). Alternatively, you can get the 32GB version of the Tesla V100 reference card for $8,509. NVidia also sells the 16GB version of the reference card for $5,995.00.

NVidia V100 SXM2

NVidia V100 SXM2 upclose

  • Release Date: 27th March 2018
  • Price: $10,664 – $27,500.00
  • NVidia Tensor Cores: 640
  • NVidia CUDA Cores: 5120
  • GPU Memory: 32GB HBM2
  • Base Clock: N/A
  • Boost Clock: 1601 MHz
  • Slot Size: Dual-Slot
  • Single Precision Performance: 16.4 TFLOPS
  • Tensor Performance: 130 TFLOPS
  • Max Power Consumption: 250 W
  • Best Implementation: N/A

Over six months after releasing the NVidia V100 PCIe, NVidia released the NVidia SXM2. Because it uses NVIDIA NVLink, it has substantially greater interconnect bandwidth. Almost ten times as much. The V100 PCle delivers 32GB/sec of interconnect bandwidth while the V100 SXM2 gives you a whopping 300GB/sec.

It also delivers slightly better precision performance as well as clock speeds. As expected, it supports computing APIs such as CUD, DirectCompute, OpenCL, and OpenACC.

The con here is that it’s far more expensive than the V100 PCIe and in most cases, you’ll only be able to get it on special order for system builds only. Good luck on trying to find it on Amazon or any other general electronics retailer.

The NVidia SXM2 also uses passive cooling and requires an out of the ordinary power supply unit to run. Your PSU needs to have at least 650 units of Wattage to power the V100 SXM2.

NVidia V100S PCIe

NVidia V100S PCIe with box packaging


  • Release Date: 26th November 2019
  • Price: $12,000 – $14,000
  • NVidia Tensor Cores: 640
  • NVidia CUDA Cores: 5120
  • GPU Memory: 32GB HBM2
  • Base Clock: 1290 MHz
  • Boost Clock: 1530 MHz
  • Slot Size: Dual-Slot
  • Single Precision Performance: 15.7 TFLOPS
  • Tensor Performance: 125 TFLOPS
  • Max Power Consumption: 300 W
  • Best Implementation: N/A

In March of 2019, NVidia announced that it would be purchasing Mellanox, an Israeli-American data center networking company. This deal would prove that NVidia was serious about moving into other technologies that would push their aspirations in IoT, AI, and general-purpose use GPUs even further.

NVidia’s September 2018 release of the Tesla T4 (based on the Turing microarchitecture) wasn’t an indication that they were done with Volta.  Over a year later, they would release the NVidia V100S.

In terms of double-precision TFLOPS, it would improve performance over the original Tesla V100 PCIe by 17% and the SXM2 by 5%. The single-precision performance also saw increases of 17% over the PCIe and 4% for the SXM2.

So far NVidia has only released a 32GB version of the graphics card. However, they’ve increased the memory bandwidth to 1134 GB/s. That’s a 26% overall increase of both its predecessors. On the outside, the NVidia V100s still shares the same design and color pallet of the original Test V100.

There’s also a similarity in the amount of power consumed. Both cards have 250W as their max power consumption. This means NVidia found a way to increase performance while optimizing power usage.

Nvidia Titan V

Nvidia Titan V with box packaging

  • Release Date: 7 December 2017
  • Price: $2,999 – $3,900
  • NVidia Tensor Cores: 640
  • NVidia CUDA Cores: 5120
  • GPU Memory: 12GB HBM2
  • Base Clock: 1290 MHz
  • Boost Clock: 1455 MHz
  • Slot Size: Dual-Slot
  • Single Precision Performance: 13.8 TFLOPS
  • Tensor Performance: 110 TFLOPS
  • Max Power Consumption: 250 W
  • Best Implementation: NVIDIA TITAN V VOLTA 12GB HBM2 VIDEO CARD

Since the Tesla Volta V100 series was made for general-purpose computing, graphics cards in the series do not have monitor outputs. What if we could take the Tesla V100’s massive chip and put it into a graphics card that is actually meant for gaming?

The NVidia Titan V answers this question. For almost a year, it stood as the world’s most powerful consumer graphics card (and the most expensive), until the RTX 2080 Ti dethroned it for half the price.

However, the NVidia Titan V is still more optimized for deep learning and AI. It has more CUDA cores and nearly doubles the tensor cores. What this ultimately means is that the Titan V is suited to both workstations and high-end gaming rigs.

Its outer shell is similar in design to the V100 PCIe and SXM2, but with more gold. This is why the NVidia Titan V transcends its status as an ordinary graphics card. Like a gaudy gold chain or a pair of overly expensive designer jeans, it’s a status symbol.

Nvidia Titan V CEO Edition

Nvidia Titan V CEO Edition package inclusions


  • Release Date: 21 June 2018
  • Price: N/A
  • NVidia Tensor Cores: 640
  • NVidia CUDA Cores: 5120
  • GPU Memory: 32GB HBM2
  • Base Clock: 1200 MHz
  • Boost Clock: 1455 MHz
  • Slot Size: Dual-Slot
  • Single Precision Performance: 14.8 TFLOPS
  • Tensor Performance: 125.33 TFLOPS
  • Max Power Consumption: 250 W
  • Best Implementation: N/A

Even though the NVidia Titan V CEO Edition was a limited edition card, we still felt that we needed to include it here. The card was gifted to around 20 CEOs. Everyday users were also given a chance to win the card from various publications.

The Titan V CEO Edition’s most noteworthy attribute over the original was its increase of memory size. The first Titan V came with 12GB of memory. The CEO Edition more than doubled this. It came with a whopping 32GB of on-board memory.

The CEO Edition also outperforms the RTX 2080 Ti in most benchmarks. However, the RTX 2080 Ti is still more suited for gaming because of its ray-tracing capabilities as well as driver support for a wider range of games.

If you thought the original Titan V was a status symbol, imagine the envy you will attract from fellow gamers when they find out you have this monster in your rig.

Nvidia Quadro GV100

Nvidia Quadro GV100 Volta GPU

  • Release Date: 27 March 2018
  • Price: $8,549.00 – $8,999
  • NVidia Tensor Cores: 640
  • NVidia CUDA Cores: 5120
  • GPU Memory: 32GB HBM2
  • Base Clock: 1132 MHz
  • Boost Clock: 1628 MHz
  • Slot Size: Dual-Slot
  • Single Precision Performance: 16.66 TFLOPS
  • Tensor Performance: 119.5 TFLOPS
  • Max Power Consumption: 250 W
  • Best Implementation: NVIDIA Quadro GV100 Volta GPU 32GB Graphics Video Card

NVidia loves outdoing themselves. If you were impressed with the Titan V or the V100 series and this is the first time you’re reading about NVidia’s Quadro GV100, then your jaw is about to drop.

Similarly, to the Tesla series, the Quadro series was made for workstations, professional settings, and general-purpose computing. However, unlike the Tesla V100 series, the Quadro GV100 comes with four display ports.

The Quadro V100 is currently the best implementation of the GV100 chip and Volta microarchitecture. While it was not intended for gaming, you can get max resolutions of 4096×2160 on 4 120Hz screens, 5120×2880 on 4 60 Hz screens and 7680×4320 on 2 60 Hz screens.

As you’d expect, the Quadro GV100 is also VR-Ready. It supports the latest Vulkan, OpenGL, Shader Model and Direct X graphics APIs. Unlike the V100 SXM2, the Quadro GV100 uses active cooling.

In terms of AI, computing and deep learning, the Quadro supports CUDA, DirectCompute, and OpenCL. You’ll get boost clocks of about 133.325 TFLOPs for deep learning. The Quadro is slightly more expensive than the elite version of the V100 PCIe.

Upon release, Quadro GV100 was priced at $8,999. That’s three times as expensive as the Titan V. Today, you can get it at a slightly cheaper price. If you’re an average consumer with large amounts of money to spare, you can purchase two NVIDIA Quadro GV100 Graphics cards and use them in tandem with the NVIDIA NVLink.

Just remember, you’ll need a set of really good monitors and a powerful CPU to get the best out of the GV100.

Final Words

The Volta microarchitecture has been out for a while now. Even with current microarchitectures such as Pascal and Turing (as well as ones on the horizon Ampere and Hopper), NVidia does not look like it will stop trying to get the best out of its Volta graphics cards.

We predict that there will be at least one more ludicrously priced GV100 based graphics card in the future. NVidia has only just married itself into deep learning and data center technology (relatively speaking). They will not stop producing and tapping into that market any time soon.

In this article, we spoke extensively about the Volta microarchitecture and its products. By the end of this article, you should be a near expert on everything Volta related. If not, then we’ve failed you. Either way, we hope you’ve enjoyed reading this article. Thank you for reading.


{"email":"Email address invalid","url":"Website address invalid","required":"Required field missing"}


Mduduzi Sibisi