Blackwell

The GLM5.2 model achieves 2626 tokens per second per node on AMD MI355X hardware, offering over twice the cost efficiency compared to Blackwell, according to a blog post by wafer.ai.

New AI technology, including Nvidia's Blackwell systems, is expected to significantly increase efficiency and reduce AI token prices. Companies like OpenAI and Silicon Data have noted declining token costs, with a spending index showing a drop from 2.06 to 1.75 in early June. Unnamed AI infrastructure firms also predict cheaper tokens due to improved models later this year.

Nvidia has announced the RTX Spark, an Arm-based chip for Windows PCs combining a 20-core Grace CPU with MediaTek, up to 6,144 Blackwell GPU cores, and 128GB unified memory. The chip will power slim laptops and compact desktops from partners like Asus, Dell, and Microsoft, available this fall.

Nvidia's N1X mobile processor, combining an Arm-based CPU and Blackwell GPU, will power new RTX Spark notebooks and mini PCs starting this fall. The chip, previously used in the DGX Spark AI workstations, offers up to 500 teraFLOPS of compute and 128 GB of unified memory, with systems running Windows and targeting high-performance gaming and creative workloads.

The US Department of Commerce has taken steps to block the export of advanced AI chips like Nvidia's Rubin and Blackwell, and AMD's MI350x to Chinese firms operating outside China, addressing a loophole that allowed these chips to reach subsidiaries in countries like Malaysia.