AI Magazine December 2025 | Page 202

LATEST FIGURES

Gaudi 3:
1,835 BF16 / FP8 TFLOPS
performance at approx US $ 15,000 per accelerator
Rack Scale:

64

accelerators per rack, 8.2TB high-bandwidth memory capacity
Cost Efficiency:

92 %

better performance-per-dollar vs Nvidia H200( Signal65 testing)
Arc Pro B60:

197

TOPS AI inference, 24GB memory for workstation applications

FOCUS AHEAD

Crescent Island GPU development targeting 16 GB memory with air-cooled efficiency for enterprise
Gaudi 3 PCIe cards shipping H2 2025 with OEM partnerships including Dell PowerEdge integration
Arc Pro B-Series GPUs expanding workstation AI inference capabilities with Xe2 architecture
Open Ethernet networking strategy avoiding vendor lock-in with standard infrastructure