GB10 Die-Shot: 3nm Blackwell GPU Larger Yet Faster

🔑 Enhanced Key Takeaways

• •GB10 is a Grace Blackwell Superchip SoC designed for NVIDIA Project DIGITS and DGX Spark, delivering up to 1 petaFLOP (1000 TFLOPS) of AI performance at FP4 precision.[1][2][3]
• •CPU consists of 20 Armv9.2 cores: 10 Cortex-X925 and 10 Cortex-A725, developed in collaboration with MediaTek for power efficiency.[2][3][4]
• •Features 128GB unified coherent memory via NVLink-C2C interconnect, 256-bit L5X-9400 interface with ~301 GB/s bandwidth, and 140W TDP.[1][3]
• •Includes 24MB shared L2 cache for CPU/GPU coherency, NVDEC/NVENC engines, SR-IOV support, and up to 4 displays (3x DP, HDMI 2.1a).[1]

🛠️ Technical Deep Dive

• •GPU performance: 31 TFLOPS FP32, supports Blackwell features like DLSS, ray tracing, 5th-gen Tensor Cores for FP4/FP8, next-gen CUDA cores.[1][2]
• •NVLink-C2C low-power chip-to-chip links enable unified physical/logical memory on a 2.5D interposer; GPU appears as PCIe device to OS.[1]
• •Large 24MB L2 cache in GPU enables hardware-managed CPU/GPU coherency and Address Translation Services (ATS) for physical tagging.[1]
• •Supports up to 4TB NVMe storage in Project DIGITS; two units link via ConnectX networking for 405B-parameter models.[3]

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline