Hardware

Google has unveiled two specialized 8th-generation TPUs, the v8t for training and v8i for inference, designed to power the next wave of AI agents. These new chips promise to double raw performance and dramatically improve efficiency. What does this mean for the future of AI infrastructure?

NVIDIA demonstrated Google's Gemma 4 VLA running locally on an 8GB Jetson Orin Nano, enabling real-time edge robotics without cloud processing.

AMD has released Lemonade, a high-performance, open-source server for running LLMs locally. It uniquely utilizes both GPUs and NPUs for efficient AI processing.

An experimental Wine rewrite moves Windows syscall translation into the Linux kernel, with early benchmarks showing up to 2x speedups in syscall-heavy scenarios, per XDA-Developers.

Tiny corp's Tinybox Green v2 hits 3,086 TFLOPS FP16 with 384GB VRAM; the company also lists a 2027 exaFLOP-scale 'Exabox' as a preorder-only product.

NVIDIA and Hugging Face launched Open-H-Embodiment, offering 778 hours of surgical robotics data alongside two open-source physical AI models.

A shutdown at Qatar's Ras Laffan facility wiped out 30% of global helium supply, putting AI chipmakers and South Korean fabricators on high alert.

IBM has released Granite 4.0 1B Speech, an open-source model that cuts parameter count in half while taking top honors on the OpenASR benchmark leaderboard.

Hugging Face and NXP revealed an on-device optimization blueprint that cuts ACT robotics model latency to 0.32s on the i.MX 95 SoC for real-time control.

Researchers reverse-engineer Apple's M4 Neural Engine, bypassing CoreML to achieve direct API execution, zero-copy buffers, and in-memory model compilation.

Apple's M4 iPad Air features a 16-core Neural Engine, 12GB memory, and Wi-Fi 7 connectivity starting at $599. Read our full breakdown of the hardware.

NVIDIA and Hugging Face have introduced a technical guide for deploying open-source Vision-Language Models like Cosmos Reason 2B on Jetson edge hardware.