Intel Xeon 6 Performance Feature Benchmarks: Latency Optimized Mode

[erek]

"On a geo mean basis the performance went up by 17% with the Latency Optimized Mode to make for a net win in performance and power efficiency. But ultimately it comes down to what workloads the Intel Xeon 6 Granite Rapids server is most engaged in whether there would be any measurable performance benefit and if the increased power use would be warranted. In any event, for those unfamiliar with the new Latency Optimized Mode with Intel Xeon 6, here are some reference numbers for those curious or are learning about this feature for the first time. It's certainly worth considering for those wanting to maximize the performance potential of Granite Rapids."

"Intel's documentation on the Latency Optimized Mode new to Birch Stream is for helping ensure better latency by maintaining higher uncore frequencies consistently unless limited by the RAPL/TDP power or other platform constraints. By default Latency Optimized Mode is disabled due to the impact on power consumption.

This round of testing with two Intel Xeon 6980P Granite Rapids processors on the Giga Computing R284-A92-AAL1 server and running Ubuntu 25.10 while running a Linux 6.18 kernel development version. Tests were done with the BIOS defaults and then repeating the tests when enabling the Latency Optimized Mode from the BIOS. In addition to the raw performance, the overall system power consumption was also monitored via the BMC for seeing the impact on power use.
14 Comments"

Source: phoronix.com/forums/node/1589440?__cf_chl_tk=G07Mf7lavOZ7AAOd6H9FwJOcE.xEYex4xthHPeLLDf8-1762459285-1.0.1.1-fBKzlXG7us5uU2eTSdzJKs9F8E72aSqPlLhsO_6iiW0

 

[DukenukemX]

The performance increase is rather insane. It even makes the CPU more energy efficient. It does use more power but it gains far more in terms of performance.

 

[Lakados]

The Latency Optimization stuff is a big deal for systems hosting lots of Virtual Machines....

When a VM spikes, because of how the virtual CPUs are assigned out it doesn't always (very rarely actually) cause the correct frequency increase because the vCPUs may all be on a small subset of the core, so it may only result in a 2-3% increase in overall CPU utilization on the physical CPU.

So the latency optimization should provide the hypervisor with the information it needs to correctly allocate its own resources, making the system more responsive and more efficient.

Traditionally, for both Intel and AMD systems, you need to get into some specific BIOS and Host OS tuning to optimize the system for your use cases. But if you don't keep up with it as system or BIOS updates occur, it can make things worse, not better.

This feature, assuming it works as advertised, is a huge deal. ARM has had this feature for a few years now. I want to say they call it Dynamic Voltage and Frequency Scaling, and it's part of the platform's Hardware Virtualization Extensions.
ARM having it while the others don't was one of the big benefits ARM held in the efficiency space; it's nice to see Intel working to level the playing field there.