Today morning I got a newsletter form a tech site which I always read.
THENEXTPLATFORM
Last year when I first read about the Commodity Power8 servers, I know immediately I will need it for my researches.
I was amazed the NVLINK between the CPUs and the GPUs
We have 4 P100 Pascal GPU in the server, and two 2x100Gb Infiniband connection, which together provides 400Gb connection to the outside world.
The P100
But now this is the past only, with the new Power 9 servers. Let's see what's new:
* image source
It is a two socket server
You can choose between a 16-core version of the Nimbus chip that has a base clock speed of 2.6 GHz and that turbos up to 3.09 GHz, or a 20 cores running at 2 GHz that turbos up to 2.87 GHz.
The chips are real power hungry ones with 190 watts TDP which is a little lower than past generations of Power processors.
Maximum memory capacity will be 2TB with 2666MHz DDR4 modules.
The memory speed is really impressive: 306 GB/sec of aggregate memory bandwidth across the two sockets.
The system can be configured with either four or six of the Volta G100 accelerators.
With 4 Volta GPU the bandwidth to the GPU is rated at 150GB/s, with 6 it is limited to 100GB/s
The server has a shared two-port 100 Gb/sec network interface card mounted right onto the motherboard, and this NIC interconnects with two PCI-Express 4.0 x8 slots coming off the controllers on the Power9 die.
Each socket also has a native PCI-Express 4.0 x16 slot that is enabled with the CAPI 2.0 protocol.
This is an excellent connection for non-GPU accelerators and persistent memory devices like flash and 3D XPoint or ReRAM.
From storage perspective it has two 2.5-inch drive bays and a SATA storage controller for flash drives.
The flash drives can be 960 GB, 1.92 TB, and 3.84 TB capacities.
If we have enough fast network speed for this server we don't need to much local storage.
I'm really curious to see this in live. Last times when we compiled GPU based Database MAPD to the Power GPU platform it was amazing. On one billion record size database the average query response time was 10ms on average.
I think this will be the future of the realtime analytics, but from these I will talk later in another daily tech talk. Stay tuned!