Nvidia Boosts Open-Source AI with Acquisition and New Nemotron Models
▼ Summary
– Oracle Cloud Infrastructure has launched new A4 Standard instances powered by Ampere’s AmpereOne M silicon, available in both virtual machine and bare metal configurations.
– The virtual machines offer up to 45 OCPUs (90 cores) and 700GB memory, while bare metal instances provide 48 OCPUs (96 cores), 768GB memory, and 3.84TB storage.
– Oracle claims the A4 instances deliver up to 35% better core-for-core performance than the prior A2 generation, citing higher clock speeds and an improved memory controller.
– Oracle has divested its stake in Ampere, with leadership stating a new commitment to “chip neutrality” and working with multiple CPU/GPU suppliers rather than relying on its own chips.
– Unlike proprietary Arm chips from Amazon (Graviton) and Microsoft (Cobalt), Oracle is selling these Ampere-based instances outside its own cloud platform.
Oracle Cloud Infrastructure has expanded its compute offerings with the launch of its new A4 Standard instances, powered by Ampere Computing’s latest AmpereOne M silicon. These instances are available in both virtual machine and bare metal configurations, providing customers with flexible options for running demanding workloads. Notably, Oracle continues to offer these Arm-based chips to its customers despite having recently divested its ownership stake in Ampere Computing itself.
In this architecture, each pair of AmpereOne M cores constitutes what Oracle terms an OCPU, or Oracle CPU unit. Virtual machine configurations can scale up to 45 OCPUs, which translates to 90 physical cores, and are equipped with 700GB of memory. For users requiring direct hardware access, the bare metal option provides 48 OCPUs (96 cores), a substantial 768GB of DDR5 memory, and 3.84TB of onboard storage. Both instance types support additional block storage and can deliver network bandwidth of up to 100Gbps.
Oracle states that the new A4 instances deliver a significant performance uplift over the previous generation. The company claims up to 35% higher core-for-core performance compared to A2 instances, attributing this gain to a 20% increase in clock speed and the implementation of a 12-channel memory controller. It is important to note that the older A2 instances still offer larger maximum configurations, supporting up to 78 OCPUs and 946GB of memory. Pricing for the A4 instances is set at $0.0138 per OCPU per hour, with memory priced at $0.0027 per GB per hour.
This move by Oracle occurs within a broader industry trend where major cloud providers are aggressively developing their own custom silicon. Amazon Web Services recently unveiled its 192-core Graviton5 CPU and new Trainium3 AI accelerators, while Microsoft introduced its second-generation Cobalt CPU featuring 132 Arm Neoverse V3 cores. Google, for its part, is promoting the capabilities of its Ironwood TPU v7 accelerators, suggesting performance on par with Nvidia’s latest Blackwell GPUs. A key differentiator for Oracle is that, unlike these competitors, it is selling its Ampere-based instances to customers outside of its own cloud ecosystem.
The strategic shift behind Oracle’s product direction was clarified by founder and CTO Larry Ellison, who confirmed the company’s exit from its investment in Ampere. “Oracle sold Ampere because we no longer think it is strategic for us to continue designing, manufacturing, and using our own chips in our cloud data centers,” Ellison stated. He emphasized a new corporate philosophy of “chip neutrality,” indicating that Oracle intends to work closely with all its CPU and GPU suppliers rather than relying on proprietary silicon. This approach is designed to provide maximum flexibility and leverage a diverse supply chain. Oracle has not, however, specified whether future generations of Ampere processors will be integrated into its cloud platform.
(Source: TechRadar)
