2024
CPU Architecture DEEP DIVE
New AMD and Intel processors take center stage
by Jarred Walton
AMD AND INTEL have traditionally released new processors every year, often with the biggest updates arriving for the fall season. 2024 continues that trend, but there are new processors, and there are new processors. In what has become mostly a two-year cadence, both AMD and Intel have major architectural overhauls going on, promising bigger changes than the off-year refreshes.
It’s not just about desktop CPUs, either. As another sign of the current state of affairs, where mobile solutions continue to outsell desktop parts, both companies have laptop-centric designs coming out. Intel even leads off with its mobile-focused Lunar Lake architecture, which has some important differences from the upcoming desktop-centric Arrow Lake designs that we’ll cover here. AMD likewise has its Granite Ridge line of CPUs that target desktops, with Strix Point processors going after the mobile market. Both use AMD’s latest Zen 5 architecture, but there are plenty of underthe-hood differences to discuss. There are server designs from both companies happening as well, but we won’t get into those.
All told, we have a collection of CPU architectures to discuss, with numerous changes from the traditional approach we’ve seen in the past. Intel is moving away from being the sole or even primary manufacturer of the silicon used for many elements of its latest CPUs (see last month’s Tech Talk, MPC233), turning to rival TSMC for many aspects of its multi-tile solutions. AMD has used TSMC for years, and will continue to do so for this round of updates, but there’s news to discuss there as well.
There’s a lot to cover, so turn the page to join us as we seek AMD’s Zen for the fifth time while swimming in some Intel lakes.
The block diagram overview of Zen 5 shows significant changes compared to prior Zen designs, with larger buffers and a wider pipeline.
© AMD
AMD Zen 5 Architecture
AMD IS FIRST out of the gate with its new Zen 5 microarchitecture, powering the latest Ryzen 9000-series desktop CPUs, as well as the Ryzen AI 300-series laptop parts. Officially, AMD has announced four desktop chips and three laptop parts, but we can expect that number to increase over time. Zen 5 gets further split into ‘regular’ Zen 5 cores, as well as density-optimized Zen 5c cores (see Zen 5c sidebar), but most of the functional units are the same.
Every new CPU architecture looks for ways to improve overall performance, using some combination of improved IPC (Instructions Per Cycle) and higher clock speeds. Adding performance alone isn’t enough, however, as we learned in the clock speed wars of the mid-2000s. Ideally, new architectures offer improved performance along with better efficiency, which is why clock speeds haven’t seen massive changes over the past couple of decades.
"Every new CPU architecture looks for ways to improve overall performance, using some combination of improved IPC (Instructions Per Cycle) and higher clock speeds
The first CPUs to hit 1GHz came out in March 2000 with AMD’s Athlon 1000 and Intel’s 1GHz Pentium III models. Clock speeds doubled to 2GHz and more just 18 months later with Intel’s Pentium 4 processors, and by late 2002 there were 3 GHz Pentium 4 chips available. That’s when CPU designers ran into a wall and clock speeds began increasing much more slowly. We didn’t get a retail 4GHz processor until ten years later with the AMD FX-4170! AMD’s infamous FX-9590 broke the 5GHz mark in 2013, but with rather extreme power draw for the time, and we only just hit 6GHz CPUs with the Core i9-13900KS in early 2023.
What we’re saying is that increasing CPU clocks generally hasn’t been the best way to boost performance for over two decades, so instead, CPU designers have focused on boosting IPC. There are multiple tactics that can improve IPC, including wider data paths, deeper buffers, more execution units, better branch prediction, simultaneous multi-threading (SMT), and larger caches. Zen 5 basically ticks off every one of those boxes in its quest for higher instruction throughput.
AMD reworked the front end for Zen 5 to include dual 4-wide instruction decoders, along with an 8-wide dispatch unit. Previous Zen architectures were 6-wide designs, so this represents a potential 33 percent increase in instruction throughput. The branch prediction has also been improved to have lower latency and improved accuracy—key factors for increasing performance.