Arm Unveils Cortex-A77, Emphasizes Single-Thread Performance

Back-end improvements

Historically, Arm has been somewhat conservative in how much they widen the pipeline. In some designs, they even narrowed the pipeline to achieve better power-performance-area efficiency. With the A77, Arm pulled out all the stops in order to increase the IPC. Arm increased the width of the pipeline by 50% with up to 6 instructions getting renamed and go to dispatch each cycle. This means that when instructions come from the new MOP cache, up to 6 of them can be renamed and scheduled every cycle while the traditional path from the decodes remains four. Arm only moved to a 4-way rename on their A76 last year so this is a very large change architecturally year-over-year.

The out of order window size for all recent Arm microarchitectures going back all the way to the Cortex-A57 remained the same at 128 micro-ops in-flight. Arm has always argued that the return for enlarging this was very small. On the A77, in order to expose more parallelism, the reorder buffer has also been increased by 25% to 160 instructions in-flight. This translates to roughly a 2.5-3% performance improvement at the cost of a 25% larger OoO window.

EUs

The execution units saw a number of architectural changes as well. Firstly, whereas previously there were dedicated issue queues per pipeline, on the A77, the issue queues are now unified for better efficiency. We now have one unified IQ for the integer pipelines, one for the ASIMD, and one for the memory subsystem.

The dispatch is now much wider too. Each cycle, up to 10 µOPs may be issued to the execution units. On the integer side, two new ports were added. Arm added a 4th ALU to improve integer throughput as well as a second branch unit, doubling the branch throughput. On the floating point side, Arm added a second AES unit.

Memory subsystem

In order to cope with a wider out of order pipeline, the memory subsystem has been slightly adjusted as well. The number of in-flight loads and stores have also been increased by 25%. One of the other major changes is with the store data ports. Previously those two pipelines were shared with ALUs. On the A77 there are now dedicated store data ports.

---

Update: With some feedback from Arm, the article was corrected with regard to how the decoders were changed in the A77. We apologize for that.