ARM Announces New Cortex-A77 CPU Architecture
Computex buried us in a flurry of announcements from Intel and AMD earlier this week, but the two x86 manufacturers weren’t the only companies showing off new hardware improvements. ARM also announced new designs for its own products, including the Cortex-A77 and Mali-G77.
The Cortex-A77, codenamed Deimos, is intended to continue delivering on ARM’s 20 percent annual performance target uplift. This is the second CPU developed as part of ARM’s Austin CPU family, and it’s an evolutionary improvement over the existing Cortex-A76.
Unlike its predecessor, however, the Cortex-A77 isn’t arriving alongside a new node shrink or significant same-node performance improvement, and that limits its potential upside a bit. The improvements ARM is delivering this cycle will come courtesy of IPC.
New improvements introduced with the Cortex-A77 include more fetch bandwidth, a larger out-of-order window, an L0 instruction cache (macro-op cache), new integer ALU pipeline, and double the branch prediction bandwidth, along with improvements to branch prediction accuracy. The overall function of the macro-op cache seems similar to AMD and Intel’s micro-op cache.
The overall goal of these improvements and changes is to keep the wider core fed and prevent pipeline stalls. The introduction of a macro-op cache should also reduce the effective latency of a branch prediction miss from 11 cycles down to 10, even though the CPU technically has a 13-stage pipeline. The decoder width is also increased, to 6-wide, up from 4-wide.
Anandtech has some additional details on the specific core-level improvements, but the end result is that ARM expects some significant overall performance gains.
These gains could actually put ARM in competition with Apple at the top of the mobile FPU performance charts, but raw power consumption could also bump higher, even if, as ARM claims, absolute power efficiency remains the same. Part of the reason that ARM is hitting these sorts of gains generation-on-generation, while x86 CPU vendors struggle to deliver anything like the same improvement level at vastly higher power consumption is because the constraints of high performance silicon are actually harder to budge these days than low-power chips. Intel’s largest performance improvements over the past seven years, for example, have been in its CPU families that consume the least amount of power.
ARM seems likely to continue delivering significant performance gains year-on-year. Smartphone upgrades have slowed substantially as devices last longer, but those updating after several cycles should still see measurable performance jumps.