Tackling Throughput Computing with Sierra Forest

Don Soltis describes his job to his wife as a CPU fanatic. His actual title is a little more official. He’s a senior principal engineer and chief architect for Xeon Efficient-core, but if you ask Soltis, he’ll just say he’s worked on some of “the coolest, best processors” over his 38-year career. That includes Itanium, a joint venture with HP to deliver an uber-efficient 64-bit processor, though Soltis switched tracks to Xeon when Itanium lost out to x86.

Now, he is leading development for a next-generation Xeon processor, code-named Sierra Forest.

When it arrives in the first half of 2024, Sierra Forest will be the first time Intel has offered two breeds of Xeon alongside each other. For Soltis, it’s familiar territory and years in the making. Every one of the 20 or so processors he’s worked on has been for the data center, and before Sierra Forest, he worked on the Atom C3000 series. Formerly known as Denverton, the Atom-based server solution offered up to 16 cores with industry-leading performance per watt, low thermal design power (TDP) and loads of configurable, high-speed I/O. Denverton was for networking, storage, Internet of Things (IoT), scalable solutions and life on the edge.

“It’s something we’ve been trying to do for a while,” he says, explaining that market requirements, stiff competition and a startup mindset are what ultimately put Sierra Forest on the Xeon roadmap next to Granite Rapids, Intel’s upcoming but established Performance-core offering. “The data center market is growing in terms of the breadth of performance requirements,” he says, pulling up a work-in-progress slide for his Hot Chips presentation. Not every customer is chasing raw performance across all workloads. Instead, it’s an X-Y axis where core performance sits on the vertical Y axis and core density is on the horizontal X axis.

​Intel’s next-gen server platform in 2024 will deliver strong performance and efficiency for critical workloads. (Credit: Intel Corporation)

View full image

“Taylor Swift is the best way to explain it to people,” he says.

Wait, what?

Finding the Balance: Power Play Meets Power Efficiency

Soltis says it’s the simplest example for why a customer might want lots of smaller, energy-efficient cores as opposed to the most powerful solution out there.

Late last year, millions of people flocked to Ticketmaster for tickets to Taylor Swift’s Eras Tour, and the company’s system struggled to keep up. A total of 1.5 million pre-registered fans were joined by hordes of bots and unregistered fans all vying for the same thing. Ticketmaster said it received 3.5 billion system requests, four times its previous peak.

Throughput computing (like serving up the same page to millions of desperate fans) requires lots of cores, and they need to be as power-efficient as possible to keep costs from getting out of control. More complex workloads, like artificial intelligence, are better served by fewer, more powerful cores that cost more to run. With Sierra Forest and the brawnier Granite Rapids, Intel can offer customers both, and the chips’ shared roots make for even more flexibility.

A Shared Foundation for Greatness

Balancing performance and power efficiency is not a new concept. It’s why every generation of Intel Xeon contains a multitude of options for different workloads, from telecommunications to AI, and why you’ll find Xeon inside everything from IT servers in a closet at your office to the world’s most powerful supercomputers. But Sierra Forest and Granite Rapids are different for a couple of reasons.

For one thing, there’s overlap between the E-core-based Sierra Forest and the P-core-based Granite Rapids. That gives customers additional choice in selecting the right chip for the chore – depending on workload, they might want a Granite Rapids chip or a Sierra Forest chip. The two support the same software stack and are platform-compatible to provide the right combination for each customer. Throughput computing better is better on Sierra Forest than previous Xeon processors: Last month, Intel revealed Sierra Forest offers 2.5x rack density and 2.4x performance per watt than Sapphire Rapids and a TDP as low as 205 watts.¹

“Customers will have to make the same decisions they did in the past, choosing among the different SKUs, but our E-core solution outclasses our previous throughput SKUs in both performance and performance per watt,” Soltis explains. “More performance and better performance per watt are important for customers like cloud service providers – they pay for electricity, it’s a large portion of their total cost of ownership.”

Show Me the Transistors

Soltis says this is part of the reason why Intel doesn’t plan to combine E-cores and P-cores on a hybrid chip, like it does in its client processors. “There is a breadth of E-core to P-core ratios, the whole range that customers would want,” says Soltis. “It doesn’t make sense to overprovision (and waste silicon area, which is expensive). Provisioning at the server level, rather than at the SoC level, makes more sense.” If Customer A wants an 80/20 split and Customer B wants a 60/40 split, it’s easy to make it happen.

At Hot Chips in August, Soltis took the stage to detail the Xeon E-core and the common platform that makes Sierra Forest and Granite Rapids the most dynamic duo. The shared intellectual property (IP), firmware and OS software stack has been a boon in more ways than one. “We saved a lot in design and in validation. Having the same I/O die gives us that exact same hardware compatibility and software compatibility with the platform,” says Soltis. It’s also working to our advantage with customers who need to see it to believe it.

“I always think of Missouri’s unofficial state motto: the Show-Me State. To a large degree, I believe all of our customers need to see it themselves, like ‘Show me the difference,’” he says. “They’re saying, ‘It’s not that I don’t believe you, just … show me.’ And we can. We have both available and we can say, ‘Okay we’ll run it here and we’ll run it here, and then you’ll see.’”

It sounds so simple, but so much had to go right to get to this point. “It’s unbelievably complicated. We work really hard to get 100 billion transistors connected just perfectly,” Soltis says.

The team driving Sierra Forest isn’t made up of Xeon veterans. Despite the daunting challenge of delivering against what Soltis dubs “Intel’s most important roadmap,” the excitement is palpable. With limited knowledge of historical decision-making, they don’t bear the burden of every yes or no of projects past.

“Of the 20 processors I’ve worked on, it’s one of the most enjoyable because everyone is enthusiastic about working on the next cool thing. Everything we do is a challenge, but the challenge is part of the fun.”

More: Intel Unveils Future-Generation Xeon with Robust Performance and Efficiency Architectures

_{¹ Based on architectural projections as of Aug. 21, 2023, relative to 4th Gen Intel Xeon processors. Your results may vary.}