[Ng said:] “The ability of individuals in the company to make decisions like that and move infrastructure quickly is something I really appreciate about this company.”
That might sound like a kind deference to Ng’s new employer, but he was alluding to a clear advantage Baidu has over Google.
“He ordered 1,000 GPUs [graphics processing units] and got them within 24 hours,” Adam Gibson, co-founder of deep-learning startup Skymind, told VentureBeat. “At Google, it would have taken him weeks or months to get that.”
Weeks or months? I wouldn't expect Google to get them in 24 hours, but has it really gotten that bureaucratic there?
That doesn't sound like a fundamental difference between Baidu and Google so much as it sounds like he has a lot more authority within Baidu. A lack of red tape is not inherently a good thing- if we read some other story about some other Baidu employee who asked for and received 1000 GPUs within 24 hours for a project that ended up being completely pointless, that story would be about wastefulness and lack of oversight at Baidu, not about how refreshingly free of bureaucracy it is.
I'm not necessarily disagreeing with any of the points here- there are just many different ways of looking at these data points.
Good point. Though, my impression of Andrew Ng's tenure at Google was that he was running a high visibility deep learning project[1] that was attracting some big names, and one would think he'd have had some authority to get hardware provisioned sooner rather than later. Certainly his level of authority is much greater within Baidu now than it was at Google.
Google has a LOT of highly famous AI/ML/DL researchers like Hinton, Kurzweil and Norvig to name some off the top of my head. Ng is certainly very highly regarded but he was one of many there. It seems like at Baidu he's top dog.
Non-standard hardware kind of implies dedicated servers. And multiple dedicated machine pools instead of a single shared one add a lot of ops complexity and are very bad for getting good utilization. If the GPUs aren't standard hardware, I'd be surprised if you could get them deployed in a Google data center in just months.
Imagine 16 different groups all want 1000 GPU's, each decided they want a different type.
Ignoring even the GPU's for a second, somebody has to support programming them (now you get a whole bunch of incompatible GPU toolchains), etc.
Especially for something expected to be long term, it would be stupid to let this happen en masse without at least some different folks sitting down and talking about it, which is red tape.
It depends, 1000 GPUs is a lot of money. Even if each GPU was $500 each, that's still $500,000
Granted, in Google terms that's pocket change, but still, I'd imagine there is a rigorous requirement for a solid business/project plan before management sign the dotted line.
Depends on the size of the company for them $500,000 is not a huge amount these days.
Back in the day I looked at using a neural network for an experiment to measure the efficiency of toilets that would have been £250,000 in 1982 just for the hardware - that was a bit to much.
Though back then for a single piece of HP test gear we spent around a 100 grand.
Comments
FTFA:
[Ng said:] “The ability of individuals in the company to make decisions like that and move infrastructure quickly is something I really appreciate about this company.”
That might sound like a kind deference to Ng’s new employer, but he was alluding to a clear advantage Baidu has over Google.
“He ordered 1,000 GPUs [graphics processing units] and got them within 24 hours,” Adam Gibson, co-founder of deep-learning startup Skymind, told VentureBeat. “At Google, it would have taken him weeks or months to get that.”
Weeks or months? I wouldn't expect Google to get them in 24 hours, but has it really gotten that bureaucratic there?
That doesn't sound like a fundamental difference between Baidu and Google so much as it sounds like he has a lot more authority within Baidu. A lack of red tape is not inherently a good thing- if we read some other story about some other Baidu employee who asked for and received 1000 GPUs within 24 hours for a project that ended up being completely pointless, that story would be about wastefulness and lack of oversight at Baidu, not about how refreshingly free of bureaucracy it is.
I'm not necessarily disagreeing with any of the points here- there are just many different ways of looking at these data points.
Good point. Though, my impression of Andrew Ng's tenure at Google was that he was running a high visibility deep learning project[1] that was attracting some big names, and one would think he'd have had some authority to get hardware provisioned sooner rather than later. Certainly his level of authority is much greater within Baidu now than it was at Google.
[1] http://www.wired.com/2013/05/neuro-artificial-intelligence/a...
Google has a LOT of highly famous AI/ML/DL researchers like Hinton, Kurzweil and Norvig to name some off the top of my head. Ng is certainly very highly regarded but he was one of many there. It seems like at Baidu he's top dog.
Maybe things improved after he left. Remember it's been a while. I'm the one quoted in this article btw, I'd be happy to clarify anything.
Non-standard hardware kind of implies dedicated servers. And multiple dedicated machine pools instead of a single shared one add a lot of ops complexity and are very bad for getting good utilization. If the GPUs aren't standard hardware, I'd be surprised if you could get them deployed in a Google data center in just months.
Yeah, that sounds like a bunch of red tape, justified or no.
It is a bunch of red tape.
Imagine 16 different groups all want 1000 GPU's, each decided they want a different type. Ignoring even the GPU's for a second, somebody has to support programming them (now you get a whole bunch of incompatible GPU toolchains), etc.
Especially for something expected to be long term, it would be stupid to let this happen en masse without at least some different folks sitting down and talking about it, which is red tape.
It depends, 1000 GPUs is a lot of money. Even if each GPU was $500 each, that's still $500,000
Granted, in Google terms that's pocket change, but still, I'd imagine there is a rigorous requirement for a solid business/project plan before management sign the dotted line.
Depends on the size of the company for them $500,000 is not a huge amount these days.
Back in the day I looked at using a neural network for an experiment to measure the efficiency of toilets that would have been £250,000 in 1982 just for the hardware - that was a bit to much.
Though back then for a single piece of HP test gear we spent around a 100 grand.