IBM is going all in on artificial intelligence (AI) infused hybrid-cloud deployments with its latest generation of Power-series chips based on Samsung's 7-nanometer manufacturing process. The company took the lid off its new chips today at Hot Chips 2020.

IBM's Power10 chips, which will hit the market in mid-2021, not only offers higher performance and data bandwidths than its predecessor, they do so while sipping a third of the power, claims Bill Starke, distinguished engineer for IBM Power. "We've rebuilt our processor architecture from the ground up for energy efficiency in this generation, and we're going to get a pretty massive three-X energy efficiency improvement over our prior Power9 infrastructure," he said in an interview with SDxCentral.

While IBM's chips seldom get the fanfare enjoyed by rivals Intel and AMD, the Power series hits a different subset of the enterprise and hybrid cloud market that demand high bandwidth for processing financial transactions or AI inference.

"We've really historically always been highly differentiated in our ability to bring data bandwidth into our systems," Starke said. "We are in big-iron enterprise servers, the kind of stuff that's behind the scenes — worldwide banking, logistics."

According to Brian Thompto, distinguished engineer for IBM Power, each core offers roughly 30% higher performance than the previous generation of Power processors.

Power10 is a big chip with 15 cores spread across a 600 square-millimeter die, and that's for the single-chip version. IBM also offers a dual-chip module pushing the combined core count to 30.

"The single-chip module is focused on maximizing the core strength. You're feeding the most amount of energy per core... and also maximizing the bandwidth," Starke said.

Meanwhile, the dual-chip module aims to deliver the highest compute density possible at the sacrifice of bandwidth.

By putting two chips on a single socket, IBM is also able to lower the overall thermal design power. "We are actually able to profitably feed 1200 square millimeters of silicon," Starke said.

While 15 or even 30 cores may not sound like much compared to the 28-core and 64-core offerings from Intel and AMD, Starke claims the Power10's cores are more than comparable.

"If you're comparing this to typical x86, think of one of our cores as being aimed at doing roughly two times the work of one of their cores," he said.

Thanks to improvements to IBM's PowerAXON and OMI Memory signaling interfaces, Power10 now supports shared memory across multiple servers.

While the idea of composable infrastructure is nothing new, "we're building that capability into Power10 in a manner where you can do it with reasonable latency," claimed Starke, adding that the problem with existing shared memory schemes is that while it's possible to share memory, it can't be done at acceptable latencies.

"From a cloud economics standpoint, this is all about not having to configure so much memory on each server to meet spike demand because I can always borrow from my neighbor," he said. "On the other hand, I could have the option of offering capabilities that are far beyond what any single server can do."

For example, a cloud provider using IBM's Power10 could in theory offer up to 200 terabytes of memory for memory-intensive workloads because the memory can be shared across multiple servers.

And in theory, a cloud provider could offer up to 2 petabytes of memory, Starke said.

On the topic of memory, each socket can be outfitted with up to 4 terabytes of DDR4 memory for 400 GB/s of bandwidth. Customers that need higher bandwidth and are willing to sacrifice capacity can double that by swapping in GDDR memory dims.

Finally, the chip can support large pools of persistent memory for less bandwidth-intensive workloads.

With Power10, IBM has also stepped up its AI processing capabilities by baking AI inferencing capabilities — traditionally handled by banks of GPUs — right into the chip. AI is nothing new to IBM or its Power line up, however, the company's Power9 chips were largely focused on AI training rather than inference.

IBM claims these new capabilities enable the new chip to achieve AI inference performance 20 times greater than its predecessor.

IBM envisions a hybrid-cloud environment, where rather than having to offload data to the cloud for AI training or inference, it can be processed locally.

"Now we're putting the amount of number crunching into the general-purpose processor core so that we don't have to be offloading it to other places," Starke said. "It runs right at the heart of our systems."