At Computex 2016, NVIDIA’s CEO Jen-Hsun Huang (“Jensen”) has shown the company’s latest products to a small media audience. The products had already been launched in the weeks leading to Computex and we now know that the GeForce GTX 1080 has seen a huge public interest, and that NVIDIA’s foray in deep-learning A.I datacenters will be one of NVIDIA’s fastest-growing business. The most interesting part of this meetup was to get a peek at “what NVIDIA thinks” of the industry, and how it is positioning itself.
He was asked if NVIDIA would compete again in the handset/tablet processor business with its Tegra chips. NVIDIA’s CEO answered unequivocally: “we are no longer interested in that market”. He adds, “Anybody can build smartphones, and we’re happy to enjoy these devices, but we’ll let someone else build them”.
The question keeps coming back but the answer isn’t really new: NVIDIA has announced last year that it would get rid of its ICERA LTE modem business, a key component required to be a serious contender in the mobile SoC business. This signals a rather definitive intent to depart from this sector, and the ingteresting question is “why”?
NVIDIA says it will focus on where it can truly separate itself from the competition, and do what it does best: build massively parallel computer architectures that can crunch away “computationally unbounded” problems.
“Computationally unbounded” is a key concept to grasp if you want to understand NVIDIA’s success in building some of the most complex and biggest chips. The term refers to computing problems that, by nature, benefit from a near-infinite amount of computing.
NVIDIA strayed away from these when it went after the smartphone SoC market with its various Tegra mobile chips. It seemed to make sense (at the time) for NVIDIA to try entering the smartphone market, but two things ultimately led to the exit:
1/ Smartphones aren’t pushing the boundaries of computing at a level where NVIDIA would have a natural advantage. Phones may be getting faster and faster, but the truth is that compute performance or compute density isn’t a “primary need” for this kind of product. Gaming performance was never a factor to buy a phone, and there is no such thing as a “gaming phone”.
Additionally, Tegra wasn’t always the fastest option by the time Tegra handsets would show up, if they did at all. Qualcomm, Apple and Imagination were never in a situation where they would be “put to shame” by any performance difference in a sustained way. They even went on the offensive many times, beating NVIDIA GPUs at key benchmarks.
The truth is: compute performance has a supportive role in smartphones. Smartphones are first and foremost communications devices and Qualcomm remains the leader in this market because it is first and foremost a communications company that has made itself relatively indispensable and convenient to its OEM customers."THERE IS NO SUCH THING AS A GAMING PHONE"
2/ Better business opportunities for NVIDIA have appeared since. Deep Learning AI datacenters, in-car computing and VR/PC Gaming are growth markets where NVIDIA can leverage a single GPU architecture to create a tremendous differentiation and put a severe pressure on competitors. This has all been made possible because of A.I breakthroughs realized in the past few years, and because those innovations can be efficiently accelerated by GPUs.
At the end of the day, NVIDIA was and remains a GPU company. Massively parallel computing at NVIDIA started with computer graphics (CG) because it is “computationally unbounded” application in itself. With programmable GPUs, accelerating physics or A.I became possible, but “graphics” remains the foundation of everything NVIDIA does. It all comes from the GeForce line of products. It all exists because of the graphics business.
- Graphics funds the R&D required to build those massively parallel chips.
- Graphics is the primary market for NVIDIA.
- Graphics remains the first “reason of being” of every NVIDIA architecture, including the latest one, Pascal.
Expanding outside of graphics has required more than a decade of investment in Architecture, Software, Research and Developer Relations. But this investment seems to be paying more and more dividends as the very lucrative datacenter business develops thanks to deep learning A.I and other non-graphics scientific applications. This is something that other GPU vendors such as AMD, Intel or ARM have not been able to reproduce at scale.
Jen-Hsun Huang mentioned that making cars safer through better electronic “awareness” could strongly reduce the 1M death/year in the U.S. The social contribution would be undeniable, but it also happens to be a business with great potential going forward. That’s why NVIDIA works on platforms such as NVIDIA DRIVE.
The most recent quarterly results impressed investors enough to lift the stock, which is at an all-time high at ~$46 (stock: NVDA). At the moment, NVIDIA seems unstoppable in its field, and even Google’s AI processor announcement doesn’t seem to impact investor confidence.
Google’s didn’t disclose enough information for the community to confirm (or dismiss) a real threat to NVIDIA, and because it looks like Google’s chip would be better at Inference (which is “using” the AI’s already acquired knowledge) than at Training (“teaching” the AI and build knowledge) which is much more compute-intensive. In fact, there are other inference-chips manufacturers such as Movidius who also claim to have a much better power-efficiency than GPUs.
Google’s own assessment of its chip’s competitiveness (in performance/Watt) was based on NVIDIA’s last-gen architecture called Maxwell. Since then, the NVIDIA Pascal architecture was launched with improvements in AI inference and overall higher efficiency. NVIDIA claims 2X higher performance, and 3X better power efficiency. Servers like the NVIDIA DGX-1 have an incredible compute-density which is equivalent to “hundreds” of classic servers, according to NVIDIA’s estimation.
"GPUS ARE ECONOMICALLY EFFICIENT"Despite the ups and downs, GPUs have outperformed other types of processors when it comes to solving those “computationally unbounded” problems. GPUs are economically efficient as well and thisreality has come true because GPUs are supported by the PC market, which is effectively the gate to the (GPU-driven) HPC (high performance computing) market.
Despite the ups and downs, GPUs have outperformed other types of processors when it comes to solving those “computationally unbounded” problems. GPUs are economically efficient as well and thisreality has come true because GPUs are supported by the PC market, which is effectively the gate to the (GPU-driven) HPC (high performance computing) market.
Despite the ups and downs, GPUs have outperformed other types of processors when it comes to solving those “computationally unbounded” problems. There is simply no end to this in the foreseeable future. This reality has come true because GPUs are supported by the PC market, which is effectively the gate to the (GPU-driven) HPC (high performance computing) market."THERE IS SIMPLY BETTER BUSINESS FOR NVIDIA ELSEWHERE"
The smartphone SoC business remains huge but it is no longer the kind of platform where NVIDIA wants, or can, contribute to. The smartphone SoC battle had to be bruising for NVIDIA, but the company was smart enough to retreat when it realized that there was no path to victory (possibly as early as 2013). There is simply better business for NVIDIA elsewhere.
For NVIDIA fans, there is one more chance to see a GeForce GPU in a smartphone, one day: NVIDIA could license its GPU designs just like ARM or Imagination are… but this is another story, and there’s no sign of that happening anytime soon.