IBM, Samsung Ink Agreement to Build Future Power Chips

When GlobalFoundries announced it was pulling the plug on its 7nm process node and would instead invest more heavily in custom foundry work around its 22FDX and 12FDX processes, one immediate question was what would happen to IBM and its own Power line of products. Just over four years ago, GlobalFoundries and IBM signed a deal for the former to provide manufacturing capabilities and upgrades to Big Blue. With GF gone, IBM and Samsung have signed a deal instead for Samsung to provide manufacturing capacity for IBM on 7nm and beyond.

Over at NextPlatform, Timothy Morgan has written a deep dive on the history of cooperation between IBM and Samsung. The two companies have worked together for years; Samsung and IBM both used to belong to a trade industry group known as the Common Platform Alliance. This was an agreement between GF, Samsung, and IBM (back when IBM built chips) and collaborated on 32nm development. Samsung and GF continued that collaboration down to 14nm, at which point GF licensed Samsung's process node technology while IBM left the industry altogether. Moving its business to Samsung's fabs shouldn't be a huge problem for IBM, particularly given that it has a few years to work out the kinks.

While there are earlier roadmaps that show Power10 debuting on 10nm, Next Platform shared this more recent image, which doesn't actually specify a node for IBM's next CPU designs. NextPlatform has also written about IBM's overall segment results extensively, but it's actually quite difficult to tease out how well IBM's Power systems are selling, specifically, since the company doesn't share its revenue split between mainframe (System z) and Power hardware. It's an interesting examination of why IBM continues to invest in its own custom architectures and platforms even after the vast majority of the world has moved to x86 hardware, and why such continued investment is a functional necessity for the company's overall success.

At the same time, there's no doubt that IBM will face new pressures in 2019 and 2020. Both Intel and AMD are moving to higher CPU counts, with up to 64 cores on-die for AMD and 48 for Intel with Cascade Lake AP. AMD will introduce PCIe Gen 4 itself this year, and Intel is pushing up to 12 memory channels with its Cascade Lake AP. This could put pressure on Big Blue, which may have superior memory bandwidth, but won't be fielding nearly as many CPU cores.

Samsung's 7nm process has already been cleared for mass production, but we aren't surprised at all to see IBM launching chips in 2020 as opposed to 2019. High core-count server CPUs often take longer to build than more traditional desktop parts, and while it's not unheard of for companies to lead with server launches (as AMD is expected to do with Epyc this year), it's also not unusual for server parts to lag other chips due to their intrinsic complexity.

This announcement also gives Samsung something it has lacked for its 7nm EUV push — a major top-tier customer. We're not trying to imply that Samsung literally has no 7nm customers except itself, but it hasn't made many customer announcements for the node, at least not yet. That's a concern given how the total customer base for new nodes is shrinking; according to GlobalFoundries, the reduced number of customers for the node was a major factor in its decision to cancel development.