Deploying-Heterogeneous-Hardware-in-The-Era-of-Agentic-AI-Presented-by-Intel

At OCP Global Summit 2025, Intel VP Saurabh Kulkarni presented Intel's strategy for deploying heterogeneous hardware in the agentic AI era, arguing that ~80% of data center AI compute will shift to inference by 2028 and that hardware diversity is not optional but foundational. The session showcased Intel's silicon portfolio (Xeon, Gaudi 3, Crescent Island GPU, Arc), rack-scale systems (64-Gaudi 3 OAM rack), and open standards commitments (UALink for scale-up, Ultra Ethernet for scale-out), with on-stage partners Dell Technologies (Bhavesh Patel) and Astera Labs (Mike Hendricks). Intel outlined a predictable annual GPU cadence: current foundations, then the inference-optimized Crescent Island GPU sampling in H2 2026, then next-gen Shores product line.

Key Takeaways

Inference will account for approximately 80% of AI compute cycles in data centers by 2028 (Gartner, Aug 2024), fundamentally shifting the workload profile and the optimal hardware mix.
Intel's Crescent Island GPU offers up to 160 GB LPDDR5x memory at 1.5 TB/s bandwidth in a PCIe card form factor optimized for low-cost inference, with samples targeted for H2 2026.
Intel Gaudi 3 is available in three form factors (OAM card, Universal Baseboard, PCIe CEM); the rack-scale system packs 64 Gaudi 3 OAMs with 76.8 TB/s HBD bandwidth and 8.2 TB HBM2e at 120 kW.
UALink (open scale-up interconnect) delivers 800 Gbps full-duplex in x4 config and scales to 1,024 GPUs; Intel and Astera Labs are both board-level promoter members of the UALink Consortium.
Dell Technologies launched the PowerEdge XC7740 — a 4U air-cooled enterprise server with up to 8 PCIe Gaudi 3 accelerators — and is scaling it into multi-node scale-out topologies.
Intel's open AI software stack provides day-0 PyTorch/JAX/vLLM support, OpenAI Triton enablement, OneCCL for multi-GPU communication, and a VTune-based performance analysis toolchain.
Hardware diversity is framed as a structural necessity rather than a feature: complex agentic pipelines spanning context generation, reasoning, LLM prefill/decode, tool calls, and guardrails each require different silicon optimized for cost, latency, and power.
Intel's OEM ecosystem for Gaudi includes ASRock Rack, ASUS, Dell Technologies, Fsas Technologies, Gigabyte, Inventec, QCT, and Supermicro.