NVIDIA Blackwell Ultra: NVFP4, 15 Petaflops and 288 GB HBM3e, but No FP64

NVIDIA has revealed details about the Blackwell Ultra accelerator: the chip has 208 billion transistors, 160 SMs, 640 fifth-generation tensor cores, supports the new NVFP4 format, and is equipped with 288 GB of HBM3e. The throughput of NVFP4 computations has reached 15 petaflops, and the company has abandoned FP64 development.

NVIDIA Taiwan Semiconductor Manufacturing Company

Mistral AI Engineers Uncover True Cause of Memory Leak in vLLM Using BPFtrace

The Mistral AI team discovered a slow memory leak in vLLM when serving the Mistral Medium 3.1 model with graph compilation enabled. Using pmap and BPFtrace, they found that the leak occurs not in the heap but through direct mmap calls that bypass the glibc library.

Mistral