When it comes to chips aimed at accelerating math-heavy workloads, such as the ones typically used in the high performance/scientific computing market, NVIDIA's (NASDAQ: NVDA) GPU-based Tesla cards have led the market in terms of product performance as well as in market segment share.
Indeed, although the data is about a year old, NVIDIA had about 84% of the market for dedicated high performance computing accelerators, with alternative solutions combined comprising just 14% of the market.
Server chip giant Intel (NASDAQ: INTC) has been actively pursuing this market with its Xeon Phi family of co-processors. The company released its first iteration of its Xeon Phi family, known as Knights Corner, back in mid-2013. The product has seen some success (I suspect Intel comprises the majority of that 14% number cited above), though NVIDIA's parts still clearly lead.
As my fellow Fool Timothy Green pointed out back in mid-2014, the follow-on to Knights Corner -- known as Knights Landing -- has the potential to be quite the threat to NVIDIA's Tesla accelerator business.
However, I believe that as a result of the apparent delays that Intel has seen in actually launching Knights Landing, NVIDIA may have caught quite the break. Here's why.
Narrowing the manufacturing gap
NVIDIA is expected to launch a new line of graphics processors at some point in 2016 based on its Pascal architecture. Pascal, as the company formally revealed in its CES pre-show press conference, will be built on TSMC's (NYSE: TSM) 16-nanometer FinFET Plus manufacturing technology.
Intel was originally supposed to launch Knights Landing, based on its 14-nanometer chip manufacturing technology, in the second half of 2015. Although according to Intel's James Reinders in an interview with The Register there are a few systems out there with early Knights Landing silicon, investors should expect to see a "gradual ramp-up as the new year starts."
Reinders also told The Register that the company hasn't said when "general availability" will happen.
Although I believe that Intel's 14-nanometer technology is still more sophisticated than TSMC's 16-nanometer FinFET Plus technology, the manufacturing gap is unlikely to be anywhere near what it would have been if Intel had launched the part in the second half of 2015 as originally planned.
NVIDIA may still have the architectural edge; likely to iterate quicker
Although Intel's Knights Corner chip had a manufacturing advantage over the comparable NVIDIA parts based on the latter's Kepler architecture, NVIDIA was able to keep the lion's share of the market to itself. I believe that Intel's first Knights Corner architecture was generally inferior to the Kepler architecture, which helped NVIDIA maintain such high share.
With Knights Landing, Intel has made substantial architectural enhancements, but it is very likely that NVIDIA has made some very significant architectural improvements to Pascal. Given their respective track records, I wouldn't be surprised to see Pascal pull ahead of Knights Landing here, as well.
Furthermore, I believe that NVIDIA will iterate its products here at a quicker pace than Intel will. Intel's next generation Xeon Phi is known as "Knights Hill." That product is expected to be built on Intel's 10-nanometer process, which means that it probably won't arrive until sometime in the first half of 2018.
NVIDIA, on the other hand, has said that its follow-on to Pascal, known as Volta, will find its way into systems that will be delivered to customers starting in 2017 .
In the near-to-medium term, I expect NVIDIA to maintain overall product leadership in the world of high-performance compute accelerators. I do think Intel will continue to improve upon its product offerings and could very well become much more competitive over a couple more product generations, but for now NVIDIA appears to be in the driver's seat.
