Nvidia Unveils Spark Supercomputer & Blackwell GPU: AI Hardware Revolution
AI Hardware
Nvidia Spark Desktop AI Supercomputer Launch: Nvidia's Spark desktop AI supercomputer is set to launch this summer, capable of 10^15 operations for around $3000, aligning with Ray Kurzweil's predictions about AI computational capabilities.
Nvidia Introduces Blackwell GPU Architecture: NVIDIA has released a new GPU architecture called Blackwell, which is claimed to be 25 times faster than the previous Hopper architecture. There is ongoing debate about the validity and comparability of these benchmarks due to differences in precision and batch sizes.
Nvidia DGX Spark (Project Digits) Specs Released: NVIDIA has released the specifications for its DGX Spark, previously known as Project Digits. The DGX Spark is a workstation designed for AI inference and fine-tuning, featuring 273 GB/s memory bandwidth and 128GB of RAM, priced at $3,000. It is intended for running large AI models and supports various AI tools and frameworks.
- Nvidia digits specs released and renamed to DGX Spark
- NVIDIA DGX Spark (Project DIGITS) Specs Are Out
- DGX Sparks / Nvidia Digits
AI Models
Meta Releases New 3D Model: Meta has released a new 3D model that is described as lightning-fast. The model appears to be a splatting style algorithm, but it does not fill in the angles and details that are not in the input.
Meta Releases VGGT Model: Meta has released a new AI model called VGGT (Visual Geometry Grounded Transformer). This model can compute LIDAR-like point clouds from images or video frames, offering a fast and efficient way to estimate depth and camera angles without needing metadata.
Nvidia Releases Llama Nemotron Reasoning Models: Nvidia has released two new Llama-3-based models, the Llama-3_3-Nemotron-Super-49B-v1 and the Llama-3.1-Nemotron-Nano-8B-v1, both optimized for reasoning capabilities in AI agents. The models are available on Hugging Face, and Nvidia has also shared their post-training dataset.
Benchmarks for Various AI Models: Benchmark results for various AI models, including Llama Nemotron Super 48B, Gemma 3 27B, and Mistral Small 3.1, highlighting their performance across different categories such as reasoning, coding, mathematics, and language tasks.