Sometimes a boring product launch is exactly what a company needs. Intel has launched Granite Rapids, its new P-core Xeon 6 processors. The company previewed the family a year ago, promising delivery this year, and it has delivered. Available with 128 cores, the new Xeon has more CPUs than the comparable AMD Epyc family on sale now. Forthcoming AMD Turin Epycs will overtake Granite in this metric, but Intel now leads for the first time in years.
For a 16-vCPU configuration, Intel claims to match Epyc on integer tasks and exceed it on other workloads. Other than for floating-point workloads, Granite excels owing to special features such as AMX AI instructions and QAT encryption offloads. However, deployments relying on discrete AI and crypto accelerators won’t benefit from these special Intel features.
From a manufacturing perspective, Granite employs chiplets and the Intel 3 manufacturing process. The chiplet approach facilitates economical scaling, allowing Intel to integrate one, two, or three computing dice in a package. The Intel 7 process used for I/O allows Intel to use a technology better suited for mixed-signal circuits than the new-fangled Intel 3. The only other Intel 3 product is Sierra Forest, the E-core Xeon 6 processor. Commercialization of Intel 3 is an important milestone in Intel’s 5N4Y plan, but its limited use hurts Intel’s ability to amortize its development and deployment costs.
The company joins chiplets with silicon bridges, an approach Intel has used for generations that is like the comparatively immature TSMC Cowos-L said to be the source of Nvidia Blackwell delays. Each compute die also integrates memory controllers, an approach Epyc abandoned after its first generation to mitigate NUMA problems.
Bottom Line
A goal when Intel hired CEO Pat Gelsinger was to get product- and technology-development back on track. The launch of the first Xeon under his watch, Sapphire Rapids, suffered delays and didn’t catch up to Epyc. Since then, Emerald Rapids, Sierra Forest, and now Granite Rapids have gone well, meeting expectations the company has been setting for the past two years. Despite its economics, the Intel 3 process has also met expectations. Closing the say-do gap is essential to restoring confidence in the company. Having accomplished this with Xeon products, Gelsinger must do the same with Intel’s income statement and manufacturing.