In early 2016, NVIDIA and Google announced that the NVIDIA Tesla P100 was coming to the Google Cloud Platform in early 2017. We are not sure how early September is in 2017, but it has finally happened. You can read the formal announcement here.
One can now attach up to 4x NVIDIA Tesla P100 GPUs or 8x NVIDIA Tesla K80 GPUs (equivalent to four K80 cards you can purchase) to a single instance. Beyond just GPUs, you can attach up to 3TB of local SSD space for high-IOPS and capacity local storage. By using the instances on demand you do not need to get a NVIDIA DGX-1 in order to take advantage of the high-end Tesla P100 GPUs.
Launch pricing is as follows:
|SKU – LOCATION||ON DEMAND PRICE DIE / HOUR (USD)||SUSTAINED USE DISCOUNTS DIE / HOUR (USD) (UP TO 30% OFF)|
|P100 – US||$2.30||As low as $1.61|
|P100 – EUROPE/ASIA||$2.53||As low as $1.771|
|K80 – US||$0.70||As low as $0.49|
|K80 – EUROPE/ASIA||$0.77||As low as $0.54|
That is certainly some premium over the K80 GPUs however justified. Google lowers the pricing of GPUs over time.
The NVIDIA Tesla P100 is an important card even though the new Volta generation is coming online. You can read more about the new NVIDIA Tesla V100 from our Hot Chips coverage. We have heard those are starting to ship but it generally takes some time for cloud providers to stock the cards. NVIDIA instead sells its flagship cards to those willing to pay a hefty premium first, then to the cloud providers that generally wait for higher discounts.
If you need fast half or double precision performance for machine learning, AI, or other applications like simulation, then GCP can help you scale quickly and efficiently to get Tesla P100’s immediately and without having to specifically provision GPU racks. With the GCP sustained use pricing, the cost is not too bad especially if you know you will have sustained use only for a few months.