It’s been almost a decade since ARM unveiled ARMv8, its 64-bit architecture. The company has now developed ARMv9 to focus on extending the length and breadth of its addressable ecosystem.
Unlike ARMv8, ARMv9 isn’t a ground-up overhaul of ARM’s entire instruction set. The new architecture includes various improvements ARM has released to ARMv8 as optional standards; they’ll now be incorporated into the baseline. ARM is also including a follow-up SIMD instruction set to Neon, dubbed SVE2. SVE2 is an extension of ARM’s earlier Scalable Vector Extensions SIMD implementation, which focused primarily on high-end computing. The fastest supercomputer in the world, Fugaku, is based on a Fujitsu A64FX processor that implements SVE, but most ARM chips today still rely on Neon.
One of the big differences between SVE2 and competing Intel standards like AVX-512 is that SVE2 offers variable vector sizes from 128 to 2,048 bits. Developers are only supposed to need to compile their code once to take advantage of this improvement. If ARM decides to implement wider SVE2 registers, pre-existing code will be able to take advantage of them. SVE2 also introduces new instructions to improve its performance and overall capabilities relative to SVE. Accelerators are expected to provide most AI processing needs, but SVE2 can improve the performance of AI calculations running on the CPU.
Confidential Compute Architecture
We’ve seen a number of high-profile hardware attacks in the past few years and ARM has clearly been paying attention. The company is announcing its Confidential Compute Architecture (CCA). Available detail is currently high-level, ARM will give more information on the feature later this summer.
CCA introduces Realms, which are sandboxed application containers intended to execute the program from the hardware it’s running on. Neither the OS nor hypervisor can see into these containers. Realms don’t use a conventional hypervisor, they rely on a “Realm Manager,” which is roughly 1/10 the size.
CCAs are intended to shrink the degree of trust required when running applications in the cloud on an unknown platform. As the cloud becomes more prevalent, there’s a greater need to offer secure hosting to companies working with sensitive data. With that said, this looks like another closed approach to security. One point of disagreement between Intel, AMD, ARM, and the general security community is how much public information companies should provide about these solutions to enable thorough testing.
ARM also shared some performance estimates for future products. It expects the upcoming Matterhorn and Makalu CPUs to deliver a cumulative 30 percent performance improvement. This suggests ARM is having more trouble finding ways to boost performance than it once did, though a 15 percent annual rate of increase is still quite good. ARM also intends to introduce ray tracing features on future Mali GPUs. Ray tracing tends to hit desktop and mobile PCs quite hard, so it’ll be interesting to see if game developers can bring this capability to much smaller devices.
ARMv9 CPUs won’t be available until 2022, so we won’t see any devices using the new ARMv9 architecture this year. The performance and security benefits from the new ISA are tuned to address the evolving computer market and while the pace of improvements may be slowing, ARM is confident there’s plenty of gas in the tank.
Now Read:
0 Commentaires