When Google recently announced its TPU Custom ASIC, with imited details of chip per-se, it has ruffled the feathers of many players. What Google has stated in a crisp note as part of their product line announcement,
"The result is called a Tensor Processing Unit (TPU), a custom ASIC we built specifically for machine learning — and tailored for TensorFlow. We’ve been running TPUs inside our data centers for more than a year, and have found them to deliver an order of magnitude better-optimized performance per watt for machine learning. This is roughly equivalent to fast-forwarding technology about seven years into the future (three generations of Moore’s Law)."
A few points to ponder upon,
1. Can Google make a custom ASIC of that power and test and deploy so quickly ?
2. It should have done this project in resource optimised way, negotiating well with various vendors -- EDA, Foundry, Packaging and test house. Did it manage well the supply chain ?
3. Who is the loser in this in-house custom ASIC ? --- NVIDIA with its Pascal GPU or Intel or any other top semicon company ?
4. Is it one-off kind of big effort and market leaders in semicon need not read too much ?
5. We may never come to know if it was best hardware piece OR software running around it is giving an edge in the performance ?
6. Finally -- Do other big vendors like Apple, Amazon, MicroSoft, etc. take this path of customised ASIC for their flagship products, instead of relying on beating around standard products ? If this is so, then it could be a paradigm shift in the industry.
"The result is called a Tensor Processing Unit (TPU), a custom ASIC we built specifically for machine learning — and tailored for TensorFlow. We’ve been running TPUs inside our data centers for more than a year, and have found them to deliver an order of magnitude better-optimized performance per watt for machine learning. This is roughly equivalent to fast-forwarding technology about seven years into the future (three generations of Moore’s Law)."
A few points to ponder upon,
1. Can Google make a custom ASIC of that power and test and deploy so quickly ?
2. It should have done this project in resource optimised way, negotiating well with various vendors -- EDA, Foundry, Packaging and test house. Did it manage well the supply chain ?
3. Who is the loser in this in-house custom ASIC ? --- NVIDIA with its Pascal GPU or Intel or any other top semicon company ?
4. Is it one-off kind of big effort and market leaders in semicon need not read too much ?
5. We may never come to know if it was best hardware piece OR software running around it is giving an edge in the performance ?
6. Finally -- Do other big vendors like Apple, Amazon, MicroSoft, etc. take this path of customised ASIC for their flagship products, instead of relying on beating around standard products ? If this is so, then it could be a paradigm shift in the industry.
No comments:
Post a Comment