Monday, November 16, 2026 Edition No. 240
Search
Daily News Fully automated · 5 min read

NVIDIA AI Updates: November 14, 2026

Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled

1. NVIDIA Serves the 753B-Parameter GLM-5.3 Across Two DGX Stations With vLLM, Peaking at 608 Tokens per Second

NVIDIA published a walkthrough for serving GLM-5.3 NVFP4, a quantized mixture-of-experts model with 753 billion parameters (40 billion active per token), on two DGX Stations linked over ConnectX-8 RDMA. The checkpoint is about 464 GB, more than the 252 GB of HBM3e on a single GB300 GPU, so vLLM v0.30.0 splits the weights with tensor parallelism (TP=2), leaving about 231 GiB in use per GPU after loading. The configuration uses multi-token prediction with 4 speculative tokens, FP8 KV-cache storage, prefix caching and a 98,304-token context length. On an AIPerf run of the SPEED-Bench Coding set, overall throughput was 107.4 tokens per second at concurrency 1 and peaked at 607.8 at concurrency 16, then fell to 385.0 at concurrency 32 as mean time to first token rose from about 483 ms to 899 ms. Source

2. cuBLAS in CUDA Toolkit 13.4.2 Adds Ozaki-II and 2M Algorithms for Faster Double-Precision Matrix Multiplication

NVIDIA added two algorithms to cuBLAS in CUDA Toolkit 13.4.2 that emulate double-precision (FP64) matrix multiplication on integer Tensor Cores. Ozaki-II uses the Chinese Remainder Theorem and needs roughly half the Tensor Core GEMMs of Ozaki-I at FP64-matching accuracy, while the 2M method computes a complex GEMM with two real GEMMs instead of four. NVIDIA reports DGEMM with Ozaki-II running 5.5x faster than native FP64 on Vera Rubin NVL72 and 3.9x faster on GB200, with throughput up to 175 TFLOPS, and ZGEMM up to 5x faster than the conventional 4M method on GB200. End to end, a CCSD quantum chemistry workload ran 2x faster on Vera Rubin NVL72 at full FP64 precision, and BerkeleyGW ran 6.6x faster with matching results to 10 decimal places. cuBLAS chooses between the algorithms with heuristics and falls back to native FP64 when the required mantissa bits exceed the configured maximum; samples are in the NVIDIA/CUDALibrarySamples repository. Source

3. SOMA Retargeter v0.2 Maps Human Motion Capture Onto Five Humanoid Robots

NVIDIA detailed the v0.2 release of SOMA Retargeter, an open source tool that adapts human motion to a robot’s proportions, joint structure and range of motion. It takes BVH animation on the SOMA base skeleton and writes CSV files with per-frame root position, orientation and joint angles, which can be converted to motion-library formats for SONIC and ProtoMotions. Reference configurations and retargeted examples ship for the Unitree G1 and H2, Booster T1, and AgiBot X2 Ultra and A3T3. The pipeline scales the motion to the robot, solves inverse kinematics with Newton on NVIDIA Warp, and applies post-processing such as foot-plant correction, with a Robot Configurator, an experimental IK weight optimizer and headless batch retargeting. The post gives no performance numbers and does not state the license. Source

4. cuda-bindings 13.3.0 Brings cuDLA to Python for Deep Learning Accelerator Inference on Jetson AGX Orin

NVIDIA documented cuda.bindings.cudla, a Python layer over the cuDLA C++ API introduced in cuda-bindings 13.3.0, which exposes thirteen functions, nine data types and eight enums. It targets hybrid-mode workflows in which CUDA streams coordinate inference on the Deep Learning Accelerator (DLA) with GPU preprocessing and postprocessing on a Jetson AGX Orin. Running an existing loadable needs only the libcudla.so that ships with JetPack, while compiling one uses TensorRT 11.4.1 with CUDA 13.4 Update 1 on JetPack 7.2.2 or later; Python 3.10 is the minimum. Standalone mode is not supported in this release and raises CudlaError(ErrorUnsupportedOperation). The post includes no performance measurements. Source

5. cuPhoton’s xBeam Correlator Processes 10 TB of Event Horizon Telescope Data in 70 Seconds on 64 GPUs

NVIDIA described xBeam, the FX correlator in cuPhoton v0.1.3, which turns radio telescope voltage recordings into visibilities for Event Horizon Telescope imaging. On a Vera Rubin NVL72 system with 64 GPUs and GPUDirect Storage in compatibility mode, a 10 TB dataset was correlated in 70 seconds, against 31,992 seconds (8.89 hours) on an x86 CPU baseline, about 460x faster. Data-loading bandwidth scaled from about 20 GB/s on 4 GPUs to about 170 GB/s on 64, and xBeam can run on POSIX I/O where GPUDirect Storage is unavailable. The correlator was developed with the Harvard-Smithsonian Center for Astrophysics and MIT Haystack Observatory, and the code, sample recordings and job files are in the NVIDIA/cuPhoton repository. Source