Intel – Data dreaming

Reply to this post

Intel is doing what it must to keep its position in the Data Centre. 

• Intel has launched its attack on the field of Artificial Intelligence (AI) with a series of initiatives aimed at easing the compute load but seems to completely miss the fact that the real challenges of AI are based on software not hardware.
• At a special event in San Francisco, Intel detailed its strategy to address AI and at the same time relaunched an AI compute platform it acquired from Nervana.
• Intel’ strategy includes:
  • First: A new Intel product called Knight’s Crest that tightly integrates Xeon processor with Nervana’s computing platform that is built from the ground up for AI.
  • Second: An update to the Xeon Phi processors (Kinghts Mill) that will be available in 2017 that will deliver a 4x improvement in AI computations.
  • Third: Intel has launched a series of initiatives to ensure that the platform it is offering is both easy to use and as widely available as possible.
  • This includes the release of developer tools and a series of initiatives aimed at driving engagement with the AI community, higher education and schools.
• This strategy is exactly what Intel needs to be doing as it plays directly to its strengths in terms of designing the best performing processors but its commentary shows that it has not understood what the big challenges of AI are.
• Intel confidently expects that the Intel Nervana platform will provide a dramatic reduction in the time required to train neural networks and promises to deliver a 100x improvement in performance by 2020.
• However, I think that Intel has missed the fact that the big challenges faced by AI today have very little to do with the ability to crunch data.
• The biggest problems with AI are:
  • A vast amount of data is needed to train an AI because the algorithm needs to see a lot of examples before it can draw any conclusions.
  • A trained AI cannot transfer what it has learned to any other task.
  • Building and adjusting the models during training is a manual and very time consuming task.
• I believe that it is these issues that are holding up progress in AI and no amount of raw horsepower is going to meaningfully speed up the solution of any of these problems.
• Consequently, while Intel might deliver a 100x increase in performance in number crunching, it won’t be until the programmers have figured out how to train AIs with less data or to have the machines build their own models that AI takes a big leap forward.
• The net result here is that Intel has produced a product line-up that should help preserve its dominant position in data centre processors but it will not suddenly make Intel a nerve centre for AI.
• If Intel can encourage all of the AI industry to run its models on Intel processors, then the threat from ARM in the data centre will be meaningfully reduced.
• Intel is merely doing what it must to ensure that its hugely dominant and highly profitable products are relevant for the next generation of data centre computation.