Video generation has become one of the most discussed topics in the field of artificial intelligence in 2026. Developments such as ByteDance Seedance 2.0, MiniMax H3, and Seedance 2.5 have propelled Chinese teams into global rankings, and applications are expanding into the film industry, short dramatic storytelling, and e-commerce.
However, behind the grand statements lie technical challenges: inference costs remain high, the complexity of multi-threaded video understanding is growing, and general-purpose graphics processing units (GPUs) expend computational power, energy, and time on tasks that do not align with their original purpose.
SmarCo HT Tech, one of the first Chinese teams to implement a dataflow architecture, has introduced its new Jingang GC3 chip. This chip features 12 RISC-V cores, provides 200 TOPS of INT8 computation, and includes 128 GB of LPDDR5 ECC unified memory. CTO Wu Dongdong stated that the chip is 'born for video': the same silicon array, capable of decoding and encoding multiple streams, can execute video generation models, and its memory subsystem is designed to handle the intermediate tensors required by modern video transformers.
Wu explained why standard GPUs fall short in video analysis tasks. When analyzing video, raw streams arrive at the server, are decoded into frames, and then processed using inference on the GPU. He noted that the bottleneck is rarely pure FLOPS; rather, the issue lies in the latency of moving each frame between memory blocks and compute blocks when handling numerous parallel streams.
In the case of video generation, the GPU runs a diffusion or transformer model that creates frames, which are then compressed by an encoder. Although the model itself requires significant computational resources, every intermediate result must still be read and written, and these tensors can reach hundreds of gigabytes. Excessive frequent reading leads to silicon idling.
This architectural discrepancy underlies both problems. Central Processing Units (CPUs) and GPUs are Von Neumann machines with a program counter that sequentially executes instructions. Such a control structure is highly versatile, making it mediocre for video processing. Parallelism at the pixel and matrix level outpaces the capabilities of the central controller for task scheduling, leading to memory traffic saturation. Many cores spend more time waiting for barriers than performing useful work.
The dataflow architecture fundamentally changes this model. The program is represented as a graph of operations, where data dependencies act as edges. An instruction executes as soon as its input data becomes available, without the involvement of a program counter dictating sequence. Wu compared this to an assembly line: each station begins work as soon as parts arrive and immediately passes the result along, without a central dispatcher and with almost no downtime.
The Jingang GC3 integrates dataflow as the computational engine with RISC-V as the control core and unified memory for CPU, VPU, GPU, and NPU paths. The industry is currently testing whether dataflow can replace the controlled flow as the primary mechanism for AI computation. The Jingang GC3 is SmarCo HT Tech's answer to this question, using video as the proving ground.
The application period for the President AI Award 2026 continues in Uzbekistan—this is the country's first presidential competition focused exclusively on startups operating in the field of artificial intelligence. The deadline for submitting registration documents is set for August 15.
The competition includes a total investment fund of one million dollars. Support will be provided to fifteen AI startups whose solutions address areas such as public administration, industry and entrepreneurship, healthcare, education, green economy, and agritechnology.
The main requirement for participation is the existence of a functioning minimum viable product (MVP); projects that are only at the conceptual stage are not eligible to participate. To register, the team must provide a description and presentation of their project, as well as a video no longer than three minutes, a link to the prototype, and participant information.
Applying teams must consist of 3–8 people over eighteen years of age. Both citizens of Uzbekistan and stateless persons with permanent residence in the country are eligible to participate.
Three winners will be selected in each of the five thematic categories. The prize places include the following investment amounts: first place will receive $100,000, second place $60,000, and third place $40,000.
The investments will be provided by IT Park Ventures in two stages, contingent upon achieving predetermined performance indicators. Additionally, twenty-five of the most promising teams will participate in an acceleration program, the finalists of which will demonstrate their developments to investors at the Demo Day event scheduled for December.
Applications can be submitted free of charge until August 15 via the platform awards.gov.uz. Current information and important announcements about the competitions are published in the Telegram channels @PresidentTechAward and @president_ai_award. The media partner of the competition is the publication The Tech, and the entertainment partner is the mobile application OQ.
Celestial bodies known as rogue planets, which travel through space without orbiting a star, have always been studied by astronomy. However, there was uncertainty about what happens to the moons of these planets when they are ejected from their original systems. A new study sought to clarify this issue by conducting thousands of stellar encounter simulations.
Scientists from the Leiden Observatory, located in the Netherlands, examined approximately 34 thousand scenarios where a star passes near a planetary system, modifying its gravitational dynamics. The findings, published in the arXiv preprint repository, demonstrate that the survival of a moon is fundamentally linked to the distance at which it orbits its planet. Additionally, the characteristics of these orbits can help astronomers understand how the ejection of these systems occurred.
This work was conducted by Yannick Badoux and Simon Portegies Zwart, monitoring the fate of planets and their respective moons during close approaches with other stars. According to the simulations, the determining factor is the so-called Hill radius, which defines the area around a planet where its gravitational attraction is stronger than the influence of the host star. It is within this limit that a moon can maintain its orbit.
The researchers found that moons located up to about 40% of the Hill radius tend to follow the planet even after it is ejected from the system. Conversely, those located beyond this limit quickly lose their gravitational bond. From approximately 50% of the Hill radius, the separation between the planet and the moon becomes virtually inevitable.
The scientists suggest that this behavioral pattern may help distinguish rogue planets ejected after passing another star from those ejected by gravitational interactions with other planets, since each process tends to leave distinct signatures on the orbits of the surviving moons.
The conclusions obtained were applied to a well-known example from our own Solar System. The study points out that if Jupiter were removed from the Solar System by a nearby passing star, the moons Io, Europa, Ganymede, and Callisto would continue to orbit the gas giant. This is because these four moons are very close to the planet, occupying less than 1% of Jupiter's Hill radius.
The authors emphasize that the gravitational encounter leaves different marks on the remaining moons. Those that remain in closer orbits tend to maintain nearly circular trajectories, while those that barely escape exhibit more inclined and elongated orbits.
Furthermore, the simulations indicate that in approximately 87% of ejection events, the distance between the planet and the moon varies by less than 10%. The researchers believe that this characteristic allows astronomers to reconstruct the history of these systems just by observing the moons' orbits.
The team also used the model on an object already identified through the gravitational microlensing technique, cataloged as MOA-2011-BLG-262L. This object could be a rogue planet accompanied by a moon smaller than Earth's, although current data still allow for other interpretations.
If this hypothesis is confirmed, the simulations suggest that the system probably originated between 1.3 and 5.9 astronomical units from its star, with a slight preference for about 5.2 astronomical units, a distance comparable to that separating Jupiter from the Sun.
The study also highlights a notable consequence of these moons remaining with their planets. Certain moons can generate internal heat due to the gravitational forces exerted by the planet they orbit, regardless of starlight. Europa, for example, has a subsurface ocean whose existence is attributed to this heating. Thus, if a planet is expelled taking its moons, this heating mechanism can continue even in interstellar space, keeping these satellites warm without the need for a star.
