Upgrading cognitive computing with POWER9
IBM introduced its “Minsky” IBM Power System S822LC for High Performance Computing (HPC) in 2016. The S822LC is a high performance deep learning platform, but because of the POWER8’s buffered memory requirement and its power consumption, its cost structure is too high for widespread cloud adoption.
To reduce the cost and increase the performance of its cognitive computing platform, IBM upgraded nearly every feature of new its POWER9-based AC922 “Newell” platform.
The air-cooled four-GPU version of AC922 is available now, while the water-cooled six-GPU version will be available in Q2 2018. IBM showed both versions at the SC17 (Supercomputing) conference in November.
I labeled the processors and GPUs in the above photograph:
-
-
A. POWER9 processor with air-cooling heat sink
-
B. POWER9 processor with water cooling plate
-
C. NVIDIA V100 NVLink module with air-cooling heat sink
-
D. Empty NVLink socket
-
E. NVIDIA V100 NVLink module with water cooling plate
IBM did something with the AC922 design that I have not seen in other NVLink system designs. The AC922 design operates NVLink 50% faster with only four GPUs than with six GPUs. That means it is not straightforward to compare the four-GPU configuration with the six-GPU configuration. Performance of the two variants will scale differently, and they may excel at different workloads (fewer GPU nodes with faster interconnect vs. more GPU nodes with slower interconnect).
Also, NVLink is not just for GPUs anymore, courtesy of IBM’s OpenCAPI initiative. The OpenCAPI Consortium booth at SC17 showed an NVLink to OpenCAPI conversion board. They attached an OpenCAPI cable (which uses the SlimSAS physical format) to an OpenCAPI equipped FPGA card. OpenCAPI cabling opens a new world of high speed acceleration options for POWER9-equipped system designs, including the AC922 motherboard design. The AC922 motherboard includes three PCIe Gen4 x16 slots and one x8 slot.
IBM’s POWER-based AI business model
It is important to remember that IBM exited the x86 server business three years ago, and IBM’s POWER processor-based servers only account for 5% of server unit shipments. IBM doesn’t sell POWER processors to competing OEMs. Nonetheless, IBM is still #3 in worldwide server revenue market share.
At the same time IBM sold its x86 server business to
IBM’s investment and commitment to the POWER program appear solid. The first time I saw IBM’s current POWER processor roadmap was in Q2 2016. IBM is still using that roadmap and still tracking to it. That says a lot.
POWER9 and the OpenPOWER ecosystem
IBM founded the OpenPOWER Foundation to create a third-party hardware and software ecosystem for the POWER processor series, starting with POWER8. The POWER9-based AC922 system will start IBM’s volume sales curve for an ecosystem capable of hosting a very wide array of compute, network and storage acceleration options.
I’ve previously noted
The OpenCAPI Consortium booth at SC17 showed several FPGA cards equipped with OpenCAPI ports.
Also at SC17, Molex showed a stack of its OpenCAPI connected FPGA Storage Accelerator cards. Each card holds eight NVMe cards. With a maximum capacity of 2TB per NVMe card, each storage accelerator card can hold up to 16TB of data. Four fully-populated cards can hold up to 64TB of data alongside a Zaius motherboard in a 2U chassis, with high-bandwidth, low-latency access across OpenCAPI.
That is a lot of very fast storage in a 2U chassis.
In the photos above:
-
-
A. Molex FPGA Storage Accelerator card
-
B. OpenCAPI ports connected to Storage Accelerator backplane card
-
C. Storage Accelerator backplane card
-
D. Rear of full width chassis showing stack of four Storage Accelerator cards
-
E. OpenCAPI (SlimSAS) connector close-up
Take-away
I must believe that if IBM is shipping POWER9 processors to enterprise and CSP customers in AC922 systems, then IBM is already selling and shipping POWER9 processors directly to a few large CSPs, like
That means those CSP customers have started their evaluation cycle in earnest with production POWER9 processors.
And that means that the rest of the OpenPOWER ecosystem is being evaluated by IBM’s CSP customers.
So, while the AC922 platform is the only IBM-branded platform that launched, I believe it is only the beginning for POWER9 adoption in the market.
I can’t wait to see what 2018 holds for the OpenPOWER ecosystem, but I guess I must. Especially the “partner chips” part of the roadmap. When third-party OpenPOWER processors do enter the market, they should find a fairly mature ecosystem waiting for them.
-- The author and members of the TIRIAS Research staff do not hold equity positions in any of the companies mentioned. TIRIAS Research tracks and consults for companies throughout the electronics ecosystem from semiconductors to systems and sensors to the cloud.