Core Tuning and the Misdirection of Latency Benchmarks
It is easy to get caught up in chasing the lowest latency number or the highest synthetic score, especially when screenshots are passed around, appearing to show lower-than-average latency.
However, synthetic benchmarks, by their nature, are isolated and often detached from how real workloads behave. A single data point in AIDA64 or a quick spike in a CPU test might look impressive, but it rarely tells the full story. Real-world performance is far more nuanced, shaped by scheduling, memory behaviour, and workload distribution, not just nanoseconds on a chart. And while some will inevitably say that a five-nanosecond drop in latency “feels snappier,” the truth is that changes on that scale are quite literally beyond human perception. What is important, though, is framerate!
So, what mode is best?
Recently, AIDA64 has improved how it reflects AMD Core Tuning behaviour in its latency benchmarks. While it may not explicitly display whether Level 1 or Level 2 tuning is active, you can observe noticeably better latency figures in Level 2 mode compared to previous versions (765.7400 current latest).
Despite all that, unlike the age-old rather ridiculous claim that the human eye cannot see beyond 60 frames per second, we are now dealing with time differences as small as five to ten nanoseconds. At that scale, the variation between 60 and 70 nanoseconds is completely imperceptible to the human brain. Pursuing such marginal gains can sometimes mean you are tuning for the benchmark rather than improving actual performance.
What Core Tuning Config Does as described by AMD Engineers
With recent BIOS updates incorporating AGESA 1.0.8x and newer, AMD introduced Core/Thread Tuning Modes that allow users to modify how the CPU presents its topology to the operating system. The modes are Legacy, Level 1, and Level 2, were officially explained by an AMD employee in a Reddit post during their rollout:
“Legacy disables all topology enumeration and prefetching guidance. Level 1 is partial enumeration with limited topology exposure, and Level 2 is full enumeration and prefetching guidance enabled, which should help Windows do better scheduling decisions.”
Up until recently, most synthetics could show better latency in "Legacy" mode due to something called Prefecting and topology enumeration.
Prefetchers are mechanisms in the CPU that perform preemptive data loading. They monitor patterns in how your applications access memory and try to predict and load data into cache before it’s needed, resulting in less wait time.
Heavily simplified, Topology enumeration allows the operating system to understand the “map” of the CPU, which is critical for efficient resource use, especially on those of you sporting CPUs with multiple CCDs where intercommunication is essential.
But this doesn't always translate to better performance; it depends entirely on the workload. Synthetic tools measure latency using simplified or narrow loops that often don't reflect typical application behaviour. They're a fixed data pattern. Games are far more dynamic. They rely on unpredictable memory access, varying thread loads, and complex asset streaming, which is where these functions can improve latency.
Reddit Source
System Specs
Configuration:
2x24GB 6400MT/s tCAS 28 (MCLK/UCLK 1:1)
9800X3D @ 5.425GHz
2200 FCLK
ROG CROSSHAIR X870E HERO
ROG Astral GeForce RTX™ 5090 OC
AMD Ryzen™ 7 9800X3D
Synthetic Benchmarks
Gaming Benchmarks
Pick your Poison
As the benchmarks show, the differences between modes vary slightly depending on the workload, but overall performance deltas are minimal. Time Spy Extreme shows a small but measurable gain at Level 2, likely due to improved inter-core communication and scheduling. In games, the results are more mixed, with some titles seeing a modest uplift in average or minimum frame rates, while others show no meaningful change at all. What they do show, however, is that Level 2 isn't necessarily the best option - and certain titles benefit from certain prefecting functionality being either partially or completely disabled.
AIDA64 latency figures reinforce the idea that Level 2 can now reflect slightly better latency than Legacy mode, thanks to improved handling of prefetchers and topology, a reversal of what synthetic tools often reported in earlier versions.
That said, changing these tuning modes requires a memory retrain and cannot be toggled on-the-fly via the OS, so it's a case of pick your poison. Users with workloads that span across CCDs or rely on consistent thread scheduling may benefit more from Level 2 than single CCD parts, but for gaming, the performance delta here is truly minimal.
