What happened

Oracle Corp said it will start deploying Google’s TPUs in its cloud data centres next year. The TPUs will use Google’s ICI copper-based and OCS optical data-exchange technologies, with OCS avoiding the light-to-electric conversion step to cut power use. Oracle will link the chips to the rest of the host data centre via its own Acceleron networking stack, which relies on a custom SmartNIC.

The company also disclosed plans to add Cerebras wafer-scale accelerators and OpenAI’s Jalapeño inference processor to its portfolio. The Jalapeño chip, announced by OpenAI in August, can run some open-source models 1.5 times faster than competing silicon. Together, these hardware options broaden the AI compute choices for Oracle Cloud users.

In parallel, Oracle revealed new software services built around OpenAI models. An AI Workspace will embed ChatGPT into Oracle’s suite of cloud applications, allowing the chatbot to act on behalf of users across order-processing, inventory checks and other business tasks. A new low-code tool called App Composer will let non-technical staff create custom applications using AI assistants, while an AI Security Mesh will scan those applications for vulnerabilities.

Oracle also launched a free-to-start Oracle AI SMB Bundle for small businesses, bundling AI Workspace, App Composer, NetApp QuickStart, ChatGPT Business and Codex, and offering $12,000 in cloud credits per year.

Why it matters

The partnership marks the first time Google supplies its TPU chips to a hyperscaler other than its own cloud. By offering TPUs alongside its existing compute, Oracle gives customers the ability to run large-scale machine-learning models with lower latency and higher throughput, especially where optical data paths reduce power consumption.

Acceleron’s SmartNIC provides hardware-level tenant isolation, a high-performance fabric and zero-trust packet routing, which Oracle says moves data between databases, storage and TPU clusters with high throughput and low latency. This architecture aims to keep AI workloads secure and performant in multi-tenant environments.

Adding Cerebras wafer-scale chips and OpenAI’s Jalapeño processor further diversifies the hardware ecosystem, allowing customers to match the right accelerator to each workload. The integration of ChatGPT-driven services also blurs the line between AI infrastructure and AI-enabled applications, potentially speeding up automation of routine business processes.

What this means for enterprise IT teams

IT teams will need to evaluate which accelerator best fits their models, TPUs for TensorFlow-heavy workloads, Cerebras for massive parallel training, or Jalapeño for inference-focused tasks. The presence of multiple accelerators in the same cloud may simplify procurement but adds complexity to capacity planning and performance benchmarking.

The Acceleron SmartNIC’s tenant isolation and zero-trust routing can ease security concerns when running AI workloads alongside other enterprise workloads. Teams should review the new AI Workspace and App Composer tools for opportunities to automate repetitive tasks, but also assess the AI Security Mesh for vulnerability management.

Finally, the free AI SMB Bundle offers a low-cost entry point for small teams to experiment with AI, though larger organisations will need to consider the pricing model for the accelerator-as-a-service offering.

In the words of Oracle Senior Vice President Karan Batta: "Acceleron’s SmartNIC with hardware-level tenant isolation, High Performance Fabric, and Zero Trust Packet Routing moves data between Oracle databases, storage, and TPU clusters with high throughput and low latency."

The bottom line

Oracle’s addition of Google TPUs, Cerebras wafers and OpenAI’s Jalapeño chips expands the AI hardware palette in its cloud, giving enterprises more flexibility and potentially lower power costs. Coupled with new ChatGPT-integrated services and security-focused tooling, the move could accelerate AI adoption across a range of workloads, while placing new responsibilities on IT teams to manage a more heterogeneous accelerator environment.

Sources

  • siliconangle.com