Hexagon NPU 6 in the Snapdragon X2 Elite Extreme: 80 TOPS performance that is up to 95% faster than Apple M4 and 122% faster than Intel Lunar Lake
Qualcomm recently unveiled its next gen Snapdragon X2 Elite Extreme and X2 Elite SoCs for laptops. The new X2 Elite chips not only come with overall CPU, GPU, and memory improvements but also with significant enhancements to the Hexagon NPU.
The company claims that the latest Hexagon NPU is the fastest NPU on laptops with up to 80 TOPS of AI processing power. That's nearly a 78% increase from the 45 TOPS we've seen with the Snapdragon X lineup.
X2 Elite Extreme is built for agentic workflows
Qualcomm highlighted how a single LLM agent with a 1024 context length and 14 billion parameters (14B) would require up to 143 TOPS of AI processing power while agentic use cases with two (3B) or four agents (7B), each running an LLM and LVM concurrently, can require up to 72 TOPS.
This also brings with the challenge of optimizing for performance per Watt and low latency.
Qualcomm says that Snapdragon AI inferencing works with the Windows AI software stack to enable next gen agentic workloads. Supported open ML frameworks include PyTorch and TensorFlow.
Together with the ONNX Runtime offered by Windows ML, Snapdragon AI inferencing can involve the CPU, GPU, and the NPU to accelerate on-device AI processing.
Snapdragon X series SoCs laptop CPUs currently support more than 1,000 models and more than 300+ experiences spanning audio, video, photo, generative AI, agentic AI, and reasoning.
NPU-powered experiences can be leveraged in popular apps including Adobe Premiere, Blender, VS Code, Affinity Photo, and more.
Two decades of Hexagon evolution: 143% increase in scalar and vector throughput
Qualcomm detailed how Hexagon originally started off as a digital signal processor that primarily involved scalar processing to the current NPU 6 form that also does vector, and integer and FP16 matrix acceleration.
Scalar processing now spans across 12 threads with a 4-wide very long instruction word (VLIW) per thread that results in a 143% increase in throughput with reduced latency.
NPU 6 also features a 127% faster bus and eight parallel threads for 143% faster vector processing with support for FP8 and BF16.
The matrix unit features an independent power rail and is now up to 78% faster with 2-bit weight and FP8+BF16 support.
Hexagon NPU 6: Up to 4x the performance at iso-power
According to Qualcomm, the Hexagon NPU in the X2 Elite Extreme can offer up to 1.6x performance compared to the one in the X Elite at 5 W iso power during an Enhanced Deep Super-Resolution (EDSR) workflow.
Hexagon NPU 6: World's fastest laptop NPU
Qualcomm is pitching the Hexagon NPU 6 in the Snapdragon X2 Elite Extreme as the world's fastest, with up to 5.7x the performance and 3.8x higher efficiency compared to the Core Ultra 9 285H's NPU in Procyon AI Computer Vision.
Geekbench AI sees the Hexagon NPU 6 as being nearly 70% faster than the NPU in the Apple M4 while leading the Core Ultra 9 288V by 82.4%.
Of course, these are manufacturer numbers, so we would hold off any final conclusions until we get to test laptops based on the X2 Elite Extreme SoC early next year.
Buy the Microsoft Surface Laptop with Snapdragon X Elite (12 cores) on Amazon)
Source(s)
Qualcomm Press Brief
