IBM's Power9 CPU Could Be Game Changer In Servers And Supercomputers With Help From Google, Nvidia

Thanks to adoption from Google and Rackspace, IBM’s Power9 CPU could begin making a dent into the server chip market currently dominated by Intel. The partnership with Nvidia should also help IBM’s Power9 become popular in servers, as well as supercomputers.

Google, Rackspace To Use Power9-Based Servers

Google and Rackspace announced recently that they will be working on a new open architecture for data center servers that will be based on IBM’s upcoming Power9 CPU. Google joined the OpenPower Foundation in 2014, along with IBM, Nvidia and other companies looking to develop servers based on IBM’s OpenPower chips.

The new data center server specification is designed to fit into Facebook’s 48V open rack that Google has been helping design as part of the Open Compute Project, which is meant to standardize IT infrastructure.

Google has been a rather late entrant in a cloud computing market dominated by Amazon. However, it has lately scored some big contracts from companies such as Spotify and Apple. This should show other potential customers that Google now has a serious cloud computing business that’s likely only going to grow and mature.

Reducing Reliance On Intel

Right now, Intel dominates the server chip market almost completely, with 99 percent market share. When a company dominates any market to such a degree, that’s usually not a good sign for the customers, no matter how benevolent and reasonable said supplier has been in the past.

Google has recently shown some signs that it wants to diversify its chip suppliers. The company is now looking not just at the OpenPower architecture as a way to increase competition in the server chip market, but also at RISC-V, the new instruction set architecture that’s fully open source.

The two architectures likely won’t overlap too much for the time being, as RISC-V should become more of an ARM competitor in the medium term. Meanwhile, OpenPower is ready to become a serious competitor to Intel, starting with the Power9 CPU that’s due for release next year.

Nvidia Goes All-In With OpenPower

IBM, Nvidia and Wistron recently announced a new high-performance computing platform for servers based on IBM’s Power8 CPU and Nvidia’s latest Tesla P100, which is the company’s most powerful GPU yet.

Nvidia seems at least as committed to switching to a new CPU architecture as Google is, because Intel has been locking it out of various markets for years now. Intel did it first by bundling its GPUs with its CPUs in notebooks, making dedicated graphics cards less and less relevant in mainstream systems. Then Intel used the same strategy against Nvidia for supercomputers, where it started replacing Nvidia’s GPUs with its own Phi accelerators.

This is likely why we’ve seen Nvidia aggressively adopt ARM CPUs in its mobile chips, even making its own custom “Denver” core. The company is now partnering with IBM to combine OpenPower CPUs with its own GPUs so it doesn't have to worry about Intel anymore when selling its chips to customers.

Power9, The Game Changer

With Power9 right around the corner and with both Google and Rackspace promising to adopt it soon, Nvidia and IBM’s Power8-based chip likely won’t get too much traction this year. However, starting next year, we may see Power9-based chips make a dent into Intel’s market share in both the server and the supercomputer markets.

The Power9 CPU will be built on a 14nm FinFET process co-developed by IBM with Global Foundries. It will have on-chip accelerators and “extreme” optimization for analytics and “big data,” making it well suited for data centers and supercomputers. The Power9 CPU combined with Nvidia’s deep learning-optimized GPUs should also be well received by companies that are looking for fast AI training (such as Google, Amazon, Facebook, etc.).

Back in November 2014, Nvidia and IBM had already landed a contract with the U.S. Department of Energy to build two new supercomputers, one 100+ PFLOPS (5x more powerful than the one it replaced), and another called Summit that will have 150-300 PFLOPS and should be the most powerful supercomputer in the world when it launches.

Summit will displace China’s Tianhe-2 (34 PFLOPS), which uses Intel’s Xeon Phi processors, as the most powerful supercomputer in the world. Summit, which is supposed to come out in 2017, may even be more powerful than Intel's and Cray's 2018 Aurora supercomputer. Aurora was also contracted by the DOE and will have a performance of only 180 PFLOPS.

If IBM and its partners (mainly Google and Nvidia) can keep pushing the performance and adoption of OpenPower chips, then the OpenPower architecture may make a strong comeback in both the server and the supercomputer markets in the next few years, giving Intel a run for its money.

Lucian Armasu is a Contributing Writer for Tom's Hardware. You can follow him at @lucian_armasu. 

Follow us on FacebookGoogle+RSSTwitter and YouTube.

This thread is closed for comments
13 comments
    Your comment
  • hdmark
    Interesting read , but now it made me question something about Nvidia's strategy. Assuming they are selling large quantities of P100's for these super computers and large servers, and these P100's are not cheap at all, how large a segment is this for their company?

    Originally I thought that Nvidia and AMD made most of their money with consumer and workstation GPU's. But I'm thinking 1 of these P100's would probably net them more money than a ton of consumer cards. I also kind of figure that the big companies buy these high powered cards in bulk.
    So do NVidia/amd make consumer/workstation cards almost as a side business? or is the market share ALOT bigger than im assuming
  • InvalidError
    1913428 said:
    So do NVidia/amd make consumer/workstation cards almost as a side business? or is the market share ALOT bigger than im assuming

    Those ultra-high-end chips may net AMD and Nvidia a $1000+ margin per sale but they sell only a few thousand of these per month. Mainstream GPUs on the other hand may only net them $10-50 in gross profit per sale but they sell millions of these a month.

    Mainstream sales may not be as appealing individually but they still account for a bigger chunk of revenue and gross profit. If they want to maximize their revenue from all the R&D they invested into architecture and die optimization with their fab partners, they need to produce chips for every slice of the market where it should be reasonably profitable to do so.
  • JamesSneed
    The DGX-1 with 8 of these P100's is selling for $129,000.00 per. So they probably are making a good deal more than 1K however the mainstream / lower cost sales will more than make up for that in volume. Its not even remotely close either.