“Polaris” AmpereOne M Arm CPUs Sighted In Oracle A4 Instances

Japanese tech conglomerate SoftBank announced a $6.5 billion acquisition of Arm server CPU upstart Ampere Computing back in March, and that deal has not yet closed. And that means business has to continue as usual – well, as close to usual as possible – until the dominant shareholders of Ampere Computing have traded their stock for cash.

And so, it was not surprise when Ampere Computing said last week at Oracle AI World that Big Red was launching instances on Oracle Cloud Infrastructure based on the “Polaris” AmpereOne M processor, which has been shipping in volume since Q4 2024. It is a bit of a wonder why it took so long to get the Polaris machines inside of OCI, given that Oracle is the only one of the major clouds and hyperscalers that does not have its own homegrown Arm-based CPU and that Oracle is also a 32.3 percent shareholder in of Ampere Computing. (Private equity firm The Carlyle Group, which put together Ampere Computing from some bits and pieces of Arm server upstart Applied Micro and Intel, has just under a 60 percent stake of the company.)

The Polaris AmpereOne M CPU, which we gave its codename because Ampere Computing stopped giving us synonyms after the “Siryn” AmpereOne from 2023, has 192 cores and twelve channels of DDR5 memory and was quietly announced in December 2024. The Siryn AmpereOne (no letter after the name) also had 192 cores, but only had eight DDR5 memory channels. So Polaris had a more than 50 percent boost in memory bandwidth, which is important for database workloads as well as AI inference workloads running on CPUs.

The ”Magnetrix” AmpereOne MX chip with 256 cores and twelve DDR5 memory channels was in fabrication last July, and Ampere Computing was working on the 512-core “Aurora” Ampere CPU for further down the road with integrated AI accelerators on the chip package at some point in the future. The key thing is that Aurora has a homegrown scalable mesh interconnect to link Arm cores with the accelerators, as well as a third generation of homegrown Ampere cores that we presumed would have lots of vector units as well as an external matrix math unit on the package or inside of each core.

Here’s a roadmap refresher:

According to a statement put out by Ampere Computing, which has been quiet as a mouse in church with a hungry cat watching from the lectern, OCI will have new A4 instances based on the AmpereOne M processor in November, in both bare metal and virtualized variants. For whatever reason, there is not an A3 instance type, but the A1 instance on OCI is based on the 3.0 GHz 80-core “Mystique” Ampere Altra processor and the A2 instance is based on a 160-core variant of the Siryn AmpereOne chip also running at 3.0 GHz.

The Ampere Computing CPUs do not offer hyperthreading, which the company believes is more secure than having hyperthreads behaving like cores. But with the A2 instances as well as the new A4 instances, Oracle is pairing up to physical, single-threaded cores to look like a single, two-threaded core so it looks like an X86 processor with hyperthreads. So the A4 instance looks like it has 96 two-threaded cores running at 3.6 GHz. Those Polaris cores are running 20 percent faster, which is important for both integer and vector workloads. The extra 50 percent plus in memory bandwidth is also important. The Polaris chip also has 100 Gb/sec Ethernet ports in the package.

Add it all up and you can run AI inference on these CPUs, says Ampere Computing. The data is a bit vague, but Ampere Computing and Oracle say that running the Llama 3.1 8B transformer model, the A4 instance on OCI will offer 83 percent better price/performance than an instance running the Nvidia “Ampere” A10 GPU accelerator that is four years old. The benefit is that you can run it all on the CPU, and you don’t have to move to Nvidia’s AI Enterprise stack and CUDA-X framework.

Oracle says that the A4 instances will deliver up to 45 percent more oomph on “cloud native workloads” compared to its A2 instances and are expected to deliver 30 percent better price/performance than its E6 instances that run on the semi-custom 128-core Epyc 9J45 processor from AMD.

Oracle adds that it has over 1,000 customers renting time on the A1 and A2 instances.

The database, middleware, and application software supplier also says that its Fusion Applications ERP software suite, which is written in Java, are currently running on the AI instances and are being ported to the A4 instances, which have a lot more oomph and better bang for the buck. Oracle’s eponymous database management system has been ported to the AmpereOne chips and is implementing the memory tagging feature that Ampere Computing has added into its architecture, which helps secure memory accesses as well as recover from memory faults better.

While this is all well and good, and will help Carlyle and Oracle get their money from SoftBank, what we really want to know is how SoftBank, through its Arm, Ampere Computing, and Graphcore subsidiaries is going to help OpenAI create its own CPUs and XPUs to build its own iron so it doesn’t have to pay so much to Nvidia and AMD. Oracle will be happy to build such systems, being agnostic about compute engines.