5 ms·

Google launches the third version of its A.I. chips, an alternative to Nvidia's

62 points by mauliknshah 8y ago

throwaway2016a 8y ago
> Also, if the new version is anything like its predecessor, it will also become accessible to third-party developers through Google's public cloud service. I wish this could be bought for home use. Cloud based machine learning is great (I mean that, I love it) but sometimes a one time hardware purchase is easier to swallow than a hourly rate.
- theDoug 8y ago
  Dell is selling NVIDIA V100s for only about $20,000 for "home use". http://www.dell.com/en-us/work/shop/accessories/apd/490-bedx http://www.dell.com/en-us/work/shop/accessories/apd/490-bedx At its priciest in GCP ($2.48/hr) that's still about 7800 hours before buying was an advantage (15700 hours on preemptible gear). (Disclosure: I work at Google, but this is all public math)
  - leeoniya 8y ago
    > Dell is selling NVIDIA V100s for only about $20,000 for "home use". or... https://www.thinkmate.com/product/nvidia/900-2g500-0000-000 https://www.thinkmate.com/product/nvidia/900-2g500-0000-000
    - theDoug 8y ago
      You just saved the parent commentor $10k. I hope he gives you a commission from the savings!
    - KaoruAoiShiho 8y ago
      If it's actually home use you can buy the Quadro gv100 for 8999 on the nvidia store: https://www.nvidia.com/en-us/design-visualization/quadro-store/ https://www.nvidia.com/en-us/design-visualization/quadro-sto...
      - scottlegrand2 8y ago
        Or a TitanV for $3000 because Tesla is more branding than anything else.
  - jacquesm 8y ago
    Even at that price that's less than a year for continuous use and I'm sure that NVIDIA would be happy to make you a deal if you ordered enough of them. Now if you needed a few 100 of them for two weeks or so that would be a different matter.
    - IntelMiner 8y ago
      7800 hours = 325 days It's -barely- less than a year, with 24/7 usage
      - jacquesm 8y ago
        And the price is vastly inflated. More likely you will end up sub $10K, which makes your payback with continuous use about 180 days. And typically people that are busy with ML will buy these cards because they intend to use them full time. Model training takes insane amounts of computing power.
  - KaoruAoiShiho 8y ago
    Can you explain to us exactly how fast TPUv3 is compared to v2? There's been some confusion over how much of the increase is per chip vs sticking more chips in a pod.
    - theDoug 8y ago
      Definitely not my job to spill beans on performance of a just-announced thing but details will definitely come. I also hope to see a lot of public clarity over this stuff. "More powerful" is incredibly vague and could literally mean just consumes more power. $/time to train models that other chipsets can will be a thing I advocate for transparency with.
  - throwaway2016a 8y ago
    I saw a neat app idea that lets you crowdsource response time. I bet there are a lot of people that would gladly pay, say $10k, but personally $20k is too steep. I also don't know where this sits in relation to power and efficiency. Just because it costs Nvidia a small fortune doesn't mean someone else can't do it cheaper. Especially if the chip is laser focused to a specific instruction set (vs being a general purpose video card GPU -- which I'm not sure if the Nvidia one is). I assume from your response (and your disclosure) that is not the case though.
theDoug 8y ago
*announces, not launches Plenty of details to come. The TPU v2 in beta is available for customers currently. This is v3, not currently available.
jacquesm 8y ago
It's only an alternative by the time you can buy them. Until then owning GPU hardware is a lot cheaper than renting it if you use it for a substantial fraction of the time.
voice_dosa 8y ago
This would be an interesting way to launch a denial of service attack on a restaurant in the real world.
- what_ever 8y ago
  That doesn't even make sense. The restaurant booking, as shown during the presentation, can only be triggered through Assistant. It's not an API that you can program. You will still have to input it to the Assistant that book this restaurant on this date at this hour. And even if you do program that, Google can control it at the account level that things are not abused. You are assuming as if Google hasn't dealt with DDOS situations before. Disc: Googler but nowhere close to the project.

62 points by mauliknshah 8y ago

throwaway2016a 8y ago

> Also, if the new version is anything like its predecessor, it will also become accessible to third-party developers through Google's public cloud service. I wish this could be bought for home use. Cloud based machine learning is great (I mean that, I love it) but sometimes a one time hardware purchase is easier to swallow than a hourly rate.

theDoug 8y ago
Dell is selling NVIDIA V100s for only about $20,000 for "home use". http://www.dell.com/en-us/work/shop/accessories/apd/490-bedx http://www.dell.com/en-us/work/shop/accessories/apd/490-bedx At its priciest in GCP ($2.48/hr) that's still about 7800 hours before buying was an advantage (15700 hours on preemptible gear). (Disclosure: I work at Google, but this is all public math)
- leeoniya 8y ago
  > Dell is selling NVIDIA V100s for only about $20,000 for "home use". or... https://www.thinkmate.com/product/nvidia/900-2g500-0000-000 https://www.thinkmate.com/product/nvidia/900-2g500-0000-000
  - theDoug 8y ago
    You just saved the parent commentor $10k. I hope he gives you a commission from the savings!
  - KaoruAoiShiho 8y ago
    If it's actually home use you can buy the Quadro gv100 for 8999 on the nvidia store: https://www.nvidia.com/en-us/design-visualization/quadro-store/ https://www.nvidia.com/en-us/design-visualization/quadro-sto...
    - scottlegrand2 8y ago
      Or a TitanV for $3000 because Tesla is more branding than anything else.
- jacquesm 8y ago
  Even at that price that's less than a year for continuous use and I'm sure that NVIDIA would be happy to make you a deal if you ordered enough of them. Now if you needed a few 100 of them for two weeks or so that would be a different matter.
  - IntelMiner 8y ago
    7800 hours = 325 days It's -barely- less than a year, with 24/7 usage
    - jacquesm 8y ago
      And the price is vastly inflated. More likely you will end up sub $10K, which makes your payback with continuous use about 180 days. And typically people that are busy with ML will buy these cards because they intend to use them full time. Model training takes insane amounts of computing power.
- KaoruAoiShiho 8y ago
  Can you explain to us exactly how fast TPUv3 is compared to v2? There's been some confusion over how much of the increase is per chip vs sticking more chips in a pod.
  - theDoug 8y ago
    Definitely not my job to spill beans on performance of a just-announced thing but details will definitely come. I also hope to see a lot of public clarity over this stuff. "More powerful" is incredibly vague and could literally mean just consumes more power. $/time to train models that other chipsets can will be a thing I advocate for transparency with.
- throwaway2016a 8y ago
  I saw a neat app idea that lets you crowdsource response time. I bet there are a lot of people that would gladly pay, say $10k, but personally $20k is too steep. I also don't know where this sits in relation to power and efficiency. Just because it costs Nvidia a small fortune doesn't mean someone else can't do it cheaper. Especially if the chip is laser focused to a specific instruction set (vs being a general purpose video card GPU -- which I'm not sure if the Nvidia one is). I assume from your response (and your disclosure) that is not the case though.

theDoug 8y ago

*announces, not launches Plenty of details to come. The TPU v2 in beta is available for customers currently. This is v3, not currently available.

jacquesm 8y ago

It's only an alternative by the time you can buy them. Until then owning GPU hardware is a lot cheaper than renting it if you use it for a substantial fraction of the time.

voice_dosa 8y ago

This would be an interesting way to launch a denial of service attack on a restaurant in the real world.

what_ever 8y ago
That doesn't even make sense. The restaurant booking, as shown during the presentation, can only be triggered through Assistant. It's not an API that you can program. You will still have to input it to the Assistant that book this restaurant on this date at this hour. And even if you do program that, Google can control it at the account level that things are not abused. You are assuming as if Google hasn't dealt with DDOS situations before. Disc: Googler but nowhere close to the project.