Synchronized with AMD’s AI pep rally instead of system availability, Intel launched Arrow Lake-S (ARL-S) desktop processors, branded Core Ultra 200S to signify their contemporaneousness with the Ultra 200V (Lunar Lake) laptop processors. Systems employing the chiplet-based 200S multichip module (MCM) will be available by the end of the month.
The Three Ps
Power—When its process-technology advancements stalled, Intel eked out annual performance gains by raising its desktop processors’ power thresholds. While Arrow maintains the previous-gen chip’s (Raptor Lake’s) high TDP figures, it requires less power in normal operation. Intel says power declines 44% on the Procyon Office Productivity benchmark and 58% when running Zoom. Power—and performance—on demanding workloads isn’t clear from Intel’s statements.
Process—A year ago, Intel showed off an Arrow wafer fabricated in the Intel 20A technology (the fourth node in its 5N4Y plan). Economics, not yields, may have led the company to outsource computing-die production to TSMC. Ramping a node for only PC processors would’ve incurred excessive fixed costs relative to production volume.
Packaging—One of MCMs’ advantages, however, is that they limit the blast radius of late design changes, such as updating the GPU to the newest architecture or adding NPU hardware to raise performance. This is true, albeit to a lesser degree, of the 2.5D silicon stacking Arrow employs where the dice fit tightly together. More importantly, this chiplet approach is expensive to manufacture compared with AMD’s implementation. If Arrow’s development didn’t occur during a period when Intel needed to backstop its manufacturing plans, it may have preferred a monolithic design.
The XPUs
CPUs—Arrow integrates the same CPUs as Lunar. The Lion Cove performance CPU (P-core) is a beast. Decoder and renamer widths are eight instructions, compared with six in the previous-gen P-core, for example.
Also an eight-way machine, the Skymont efficiency CPU (E-core) is beefy compared with most licensable CPUs. However, other constraints limit its per-cycle throughput compared with Lion, and it doesn’t clock as fast. In return, it requires less than half the area and much less power. Skymont also doesn’t implement AVX-512, which prohibits the P-cores from executing AVX-512 instructions. These messy x86 extensions do provide useful features. As AMD demonstrates, an implementation can save area by breaking 512-bit operations into smaller chunks while still providing compatibility.
On the basis of Intel’s discussion of Lunar’s power efficiency running Microsoft Teams, we expect much of Arrow’s Zoom savings come from shunting execution from the P-cores to the E-cores. However, the practical benefits of mixing CPU types in a Windows machine remain murky, especially for desktop computing.
GPU—The 200S’s GPU is less advanced than Lunar’s. Desktop systems require a basic integrated GPU for mainstream users but will add a high-performance discrete GPU for gamers. The forthcoming Arrow Lake-H laptop processors will get Intel’s newer GPU architecture, like Lunar has, which will also improve the GPU’s AI performance.
NPU—Intel endowed Arrow with a 13 TOPS NPU, too little for Microsoft to bless. Sure, a developer can obtain additional AI performance by simultaneously using the CPU and GPU, entailing extra work. An add-in GPU card would provide more performance, too, but also render the integrated NPU moot. We suspect that Intel finalized Arrow’s features more than a year ago and before Microsoft established its 40 TOPS threshold for Copilot+ branding and elected to keep the little guy in there to claim bragging rights for the first AI-accelerated x86 desktop processor. It’s possible that Arrow’s reported Zoom power improvement partially comes from the NPU handling effects such as background blur.
Bottom Line
Competition—although Apple and Qualcomm processors provide a comparison basis, Intel really competes with only AMD for desktop sockets. Despite offering competent, if not superior, desktop processors, AMD has lagged far behind Intel in market share. It’s gradually catching up, but the recent Ryzen 9000 (Zen 5) launch wasn’t smooth. Nonetheless, the Core Ultra 200S does little to affect the competitive dynamic beyond addressing power—Raptor’s key shortcoming.
Customers—we doubt mainstream users care much about their PC processors absent a notable new capability. Vendors’ hopes of an AI-fueled upgrade cycle aren’t backed by consumer enthusiasm for any particular AI-enhanced applications. Should that change, Arrow lacks the requisite NPU, if not for the applications, then at least for the branding.
Despite all this, Intel deserves credit for successfully executing a simultaneous microarchitecture and process-technology upgrade (tick and tock) and for backstopping its Intel 20A plan with a TSMC N3B option to give it flexibility. As boring as PC-processor upgrades are, this one at least delivered one real product improvement—lower power for mainstream workloads.