CHIP “FEVER” AND NEW NVIDIA CHIPS.

27/08/2025
Updated on 02/11/2025

A sharp increase of interest in training and deployment of neural networks is noted by many analytical online media outlets in the field of advanced technologies.

Technology observer The Register writes that since more than two years ago the company OpenAI introduced ChatGPT, the interest in training and deployment of AI models has sharply grown, and graphics processors (GPU) have become the main accelerator for all these efforts, especially those produced by the American chip manufacturer Nvidia.

The Register also refers to the data of the research company Jon Peddie Research (JPR), according to which at present no fewer than 121 companies speak about producing artificial intelligence processors: from tiny chips designed for embedded applications and devices of the Internet of Things class (IoT — ecosystem of “smart” things — Ed.), to products for hyperscale data centers. Not less than 42 of them are based in California (USA). Many clients are beginning to report insignificant or zero return on the billions invested in AI development, believing (or perhaps hoping) that this will lead to a jump in performance.

The Register also cites the words of JPR president Dr. Jon Peddie, who says: “Artificial intelligence processors are experiencing a Cambrian explosion, reminiscent of the 3D graphics boom of the late 1990s and the XR (extended reality) wave of the 2010s,” says Dr. Peddie. “We expect rapid consolidation in the coming years, and by the end of this decade, out of the 121 players we are tracking today, about 25 will remain.”

According to the news agency CNN, the American chip manufacturer Nvidia has made an impressive march to record figures and jumped by about 21% this year thanks to its leading role in the development of artificial intelligence. As a result, Nvidia’s market value in July 2025 exceeded the mark of 4 trillion dollars.

At the same time, the research company Jon Peddie Research, referring to Reuters, describes the creation by Nvidia of a new chip B30A for training tasks of Chinese customers. It uses a single-chip architecture that provides about half the computing power of the flagship accelerator B300. The design uses 144 GB of HBM3E memory, which is 50% of the B300’s memory resources. Engineers preserved the functionality of NVLink interconnects for multiprocessor scaling, similar to the H20 model. Consolidation of components on a single die simplifies production and ensures compliance with export restrictions.

Nvidia also announced two modules of the NVIDIA® Jetson Thor series — Jetson T5000 and the more affordable Jetson T4000. On the basis of the new system module Jetson T5000 will work the most powerful mini-PC — Jetson AGX Thor Developer Kit. Jetson T5000 is built on the Blackwell architecture, includes 2560 CUDA cores and 96 fifth-generation tensor cores, supports Multi-Instance GPU, works in conjunction with a 14-core Arm Neoverse-V3AE CPU, and is equipped with 128 GB LPDDR5X RAM.

The performance of Jetson T4000 is declared at the level of 1200 teraflops (FP4, Sparse), the number of GPU cores is 1536, the memory volume is 64 GB LPDDR5X, the power consumption range is from 40 to 75 W.

Both modules are designed for flexible use in robotics, industrial automation, sensor data processing, and artificial intelligence tasks.

A new class of graphics processors designed for large-scale contextual inference, Rubin CPX, was announced by Nvidia on September 09, 2025. Rubin CPX is specifically built to handle encoding millions of tokens and generative video applications. Rubin CPX delivers up to 30 petaflops of computation with NVFP4 precision for the highest performance and accuracy. It features 128 GB of energy-efficient GDDR7 memory to accelerate the most context-demanding workloads. In addition, it provides three times faster attention processing compared to NVIDIA GB300 NVL72 systems, enhancing AI models' ability to handle longer contextual sequences without reducing speed.

Adoption of the new processor has been confirmed by Cursor, Runway, and Magic. Sales of the Rubin CPX processor are planned for the end of 2026.

Nvidia also collaborates with the processor industry leader Intel and the Taiwanese fabless manufacturer of mobile processors MediaTek.

The collaboration between Nvidia, the U.S. defense contractor General Atomics, and a group of other partners resulted in the launch of a high-precision digital twin of a fusion reactor, supported by artificial intelligence and featuring interactive performance. The project enables virtual exploration of fusion scenarios.

Visit the SAID-Test to practice identifying fake generations.

said correspondent🌐