GB200
PulseAugur coverage of GB200 — every cluster mentioning GB200 across labs, papers, and developer communities, ranked by signal.
8 day(s) with sentiment data
-
DeepSeek open-sources inference engine, DeepInfer, without KV cache
DeepSeek has open-sourced its high-performance inference engine, DeepInfer, which powers its large-scale production services. The release, available on GitHub under the Apache 2.0 license, includes DeepSeek's custom ker…
-
FastVideo releases FastH3-Omni-Comfy for synchronized video and audio generation
FastVideo has released FastH3-Omni-Comfy, a new diffusion model designed for generating synchronized video and audio from text and visual prompts. The model is available in various formats, including BF16 and INT8, with…
-
NVIDIA's GLM-5.3-Flash local build demands Blackwell GPUs
NVIDIA has released a new version of its GLM-5.3-Flash model, optimized for local use with its Blackwell architecture. This "Local" FP4 build requires significant VRAM, specifically two Blackwell GPUs or DGX Spark syste…
-
Microsoft unveils 2T-parameter MAI-Base-2 MoE model
Microsoft has unveiled MAI-Base-2, a new Mixture-of-Experts (MoE) model boasting 2 trillion parameters, with 80 billion active. This model was trained for 38 days using 16,000 GB200 units on 50 trillion tokens. Internal…
-
Nemotron 3.5 Super VL runs locally with 128GB RAM, bypassing Blackwell GPU needs
NVIDIA's Nemotron 3.5 Super VL, a 120B parameter vision-language model, can be run locally with approximately 128GB of RAM, though GPU power is less critical than memory capacity. While NVIDIA's official deployment targ…
-
JetSpec++ paper introduces faster LLM inference with speculative decoding
Researchers have developed JetSpec++, a novel approach to speculative decoding that significantly enhances the speed of large language models. By employing intra-block self-refinement and inter-block foresight drafting,…
-
NVIDIA Vera Rubin accelerates vLLM throughput by 7.8x over GB200
The vLLM project has announced support for NVIDIA Vera Rubin, a new hardware accelerator. Initial benchmarks indicate that Vera Rubin can achieve over 7.8 times the throughput of the GB200, particularly when used with t…
-
MiniMax M3 runs on NVIDIA Vera Rubin NVL72, boosting throughput 7.8x
MiniMax AI's M3 model is now operational on NVIDIA's Vera Rubin NVL72 infrastructure, leveraging the vLLM framework. This integration has reportedly resulted in a more than 7.8x increase in throughput compared to the GB…
-
AutoTrust releases compact GLM5.3-Flash for NVIDIA DGX Spark
AutoTrust has released a new, compact version of the GLM-5.3-Flash model, optimized for NVIDIA DGX Spark systems. This derivative, autotrust/GLM5.3-Flash-E192-DGX-Spark, retains 192 of the original model's experts and u…
-
New polynomial approximations boost LLM training speed on NVIDIA Blackwell GPUs
Researchers have developed new polynomial approximations for transcendental functions used in large language models (LLMs) to improve computational efficiency. These approximations, tested on NVIDIA Blackwell GPUs, show…
-
OpenAI's custom Jalapeño AI chip prioritizes internal use, hints at future rollout
OpenAI has developed a custom AI inference ASIC named Jalapeño, primarily for its internal use to meet growing compute demands. While the chip is designed to be programmable and capable of running various models beyond …
-
NVIDIA open-sources OSMO for unified AI robotics development
NVIDIA has open-sourced OSMO, a Kubernetes-native workflow orchestrator designed to streamline the development of physical AI. This tool allows robotics teams to define training, simulation, and robot testing pipelines …
-
NVIDIA vLLM supports DeepSeekv4.1 Flash on release; AMD vLLM lags
NVIDIA's vLLM software is functioning seamlessly with the new DeepSeekv4.1 Flash model across all six of its hardware SKUs, including H100, H200, B200, B300, GB200, and GB300. In contrast, AMD's vLLM implementation is e…
-
NVIDIA releases quantized DeepSeek and Qwen LLMs for Blackwell hardware
NVIDIA has released quantized versions of two large language models, DeepSeek-V4-Pro-0813 and Qwen3.8-2.4T-A95B, utilizing their NVFP4 quantization method. The DeepSeek model, with 1.65 trillion parameters, employs Hybr…
-
OpenAI unveils Jalapeño inference chip, challenging NVIDIA's dominance
OpenAI has revealed details about its custom inference chip, codenamed Jalapeño, which reportedly offers significant improvements in efficiency and latency compared to NVIDIA's Blackwell and Rubin-class systems. The chi…
-
OpenAI's custom 'jalapeño' chip benchmarks show it beating NVIDIA hardware
OpenAI has released benchmark results for its custom inference chip, codenamed "jalapeño," which it developed in collaboration with Broadcom. The chip reportedly outperforms NVIDIA's GB300 and GB200 systems in throughpu…
-
OpenAI's Jalapeño ASIC benchmarks show performance gains over Nvidia GPUs
OpenAI has developed its own 700W inference ASIC, codenamed Jalapeño, in collaboration with Broadcom. Benchmarks presented by OpenAI suggest that Jalapeño outperforms Nvidia's GB200 and GB300 GPUs in throughput per kilo…
-
OpenAI's Jalapeño chip shows superior inference performance over Nvidia
OpenAI has revealed initial performance data for its custom-designed "Jalapeño" inference chip, showcasing significant improvements in speed and power efficiency. Benchmarks indicate that Jalapeño outperforms competitor…
-
MiniMax H3 slashes video generation latency by 27x with NVIDIA Sol Engine
MiniMax AI has achieved a significant breakthrough in video generation latency using NVIDIA's Sol Engine and their MiniMax H3 model. By splitting the generation process into a low-resolution draft and a high-resolution …
-
NVIDIA releases Nemotron-Labs-Teacher models with 1M context · 4 sources tracked
NVIDIA has released a suite of Nemotron-Labs-Teacher models, each with 550 billion parameters, though only 55 billion are actively used. These models leverage a LatentMoE architecture incorporating Mamba-2, MoE, and Mul…