Mapping the PC SoC landscape

From tick-tock to SoC-o’clock

I spent an unusually large chunk of my professional life—about 17 years—doing one thing: testing and writing about CPUs and GPUs. That period of my career ended abruptly just over a decade ago when I went to work in AMD’s graphics division.

At the time, the world of CPUs was honestly not very interesting, at least in terms of the competitive landscape. The answer to “which CPU should I get?” was, for a long period, an Intel processor of some sort—probably a Core i5 or i7. Intel was ticking and tocking along, delivering worthwhile improvements to its CPUs via process tech and architectural updates on an annual cadence. AMD and others in the market couldn’t keep up.

This fact made the business of publishing CPU reviews rather challenging, as you might imagine. Nobody really needed a comparative review to make a buying decision. On the GPU side of things, a similar dynamic was playing out with Nvidia’s dominance in graphics.

Then a funny thing happened. My last CPU review for The Tech Report was the then-new Intel Skylake processor fabricated on a 14-nm process. Shortly after that, I went to work at AMD and stayed there for five years. I then took a job at Intel, in its graphics division. When I walked in the door there, Intel was still shipping 14-nm Skylake processors as its primary consumer products—somewhat revised, but fundamentally the same.

For five years, the ticks and tocks had stopped cold.

Eventually, in place of Intel’s singular dominance, a broad constellation of alternatives flourished. Fast forward to today, and the consumer CPU landscape is more interesting than ever. There are more competitors and a broader variety of technologies on offer. Yes, recent supply constraints have put a damper on things to some extent, but this, too, shall pass.

Of course, we don’t call them CPUs anymore. These products are now called SoCs, or systems on a chip.

Oddly enough, CPUs weren’t really just CPUs for a very long time—they integrated FPUs, memory controllers, last-level caches, and I/O interfaces before folks embraced the SoC label. Now, SoCs are no longer just systems on a chip, either; many of these products involve packaging together multiple chips and perhaps memory, as well.

Whatever you call them, though, they are at an intriguing place in their evolution. Some of these products are so capable and refined that they’ve turned traditionally challenging personal-computing workloads into solved problems. Meanwhile, the challenge of running generative-AI inference locally is driving a new generation of big SoCs with unified memory. I’ve already written about how I think those big SoCs are going to keep growing and may supplant traditional graphics cards for a lot of gamers and enthusiasts. In my version of events, this is a Good Thing for all involved.

So many open questions

You probably know the key players in the PC SoC race: Apple, AMD, Intel, Nvidia (with buddy MediaTek), and Qualcomm. (Eventually, Samsung might get in on the action, and possibly another dark horse like Huawei.) Each of them has some version of each of the key IPs needed to build an SoC, and they all have products shipping in systems now, with further plans on their roadmaps.

They face a host of fascinating decisions as they figure out what to do next, and I expect some of their choices to differ substantially from one another. The more compelling questions include:

What scale and component mix? — This is the most obvious one, and I’ve already alluded to the rise of big SoCs with large and fast GPUs and memory subsystems. Apple has thrown down the gauntlet yet again with the M5 Ultra. Who else will step up and compete there? And how will folks vary the mix of capabilities at different product scales?

What to do about CPU cores? — After some relative stagnation, we’ve seen real progress on per-thread performance in recent years. In the period before that, PC processors arguably had too many cores and too few advances in per-thread performance—in part because the primary development target for many CPU cores was server chips, which live in a different, more thread-parallel sort of world.

The most obvious indicator of this problem, in my view, was the unbelievable prominence of Cinebench scores in marketing materials as an indicator of consumer CPU performance. Please, show me again how good your CPU is at doing a GPU’s job. I always enjoy graphics rendering work measured in seconds per frame.

Meanwhile, Amdahl’s Law is undefeated.

Fortunately, new cores have moved the needle on IPC and per-thread perf, and I am here for it.

Another still-open question is what, exactly, to do with heterogeneous CPU cores or asymmetric multiprocessing. Most silicon providers seem to believe that offering different core types makes sense for efficiency, at least in some cases, but the solutions on offer differ substantially.

NPUs, GPUs, or both? — The NPUs in today’s PC chips have evolved to handle at least some generative AI workloads, but many of them have their roots in running small networks efficiently for real-time workloads like face detection for mobile device cameras. NPUs themselves are typically smaller and more efficient than GPUs, but they have limitations in terms of flexibility and supported operation types. In order to handle future AI workloads, NPUs will have to add more general computing chops. In doing so, will they lose their efficiency advantages over GPUs?

Bigger picture, does it make sense for NPUs and GPUs to coexist? If so, what is the mix of work that gets assigned to each one?

What happens with packaging? — Multiple chips from different fabs sharing a package, the stacking of chips on top of one another, various vertical interconnect technologies, and all sorts of related voodoo make this space one to watch. Each player seems to have its own set of technologies on this front, and they are wildly differentiated from one another.

How will traditional barriers erode? — Some structural dynamics in the PC market have changed thanks to software translation and compatibility layers, and more change is coming. Apple moved from x86 to Arm with few hiccups, and Qualcomm now ships competitive Windows laptops with Arm cores. PC gamers run Windows games in Linux on the regular. The rise of agentic AI coding tools may erode such barriers further, making the work of, say, porting PC games to macOS or Linux trivial and common. Something tells me things will get truly freaky on this front soon, with unforeseeable consequences.

Those are just a few of the questions in play today. The choices the silicon providers make in navigating them will be consequential. Past me would be kind of jealous.

Let’s take a tour

As long as I keep doing this Damage Labs thing, I get to cover some of the drama in this new era of consumer SoCs. But I have been distracted and preoccupied by corporate work and crazy hobbies for a while now. I kind of need to write myself into shape and better familiarize myself with some details of the current landscape.

To make that happen, I have cooked up a hare-brained plan to review each of the major PC SoC makers, one by one. My intention is to review their current offerings and to look a little closer at the key types of IP they possess.

At first blush, I’d say the categories to be covered for each firm will be something like this:

• CPU cores

• GPU

• NPU

• SoC integration—glue, cache, interconnects, etc.

• Memory

• Packaging

• Fabrication tech

• Software and ecosystems

That may or may not cover it, but hopefully it’s a reasonable template to get us started. My vision for this is one article per SoC maker, but that’s a ton of ground to cover. I guess we’ll see what happens. If you’re at all interested, please take a moment to subscribe—and perhaps even share with a friend. Thanks!