NVIDIA has released beta version 0.7 software suite RTX Neural Texture Compression (RTXNTC) — a technology that uses neural networks to compress texture materials. The solution is aimed at significantly reducing the volume of game data and more efficient work with textures during loading and rendering.
The updated version of the SDK adds preliminary support for Shader Model 6.9 via new driver and porting part of DirectX 12 Cooperative Vector logic to the standard Microsoft Agility SDK — previously, custom NVIDIA extensions were used. In addition, now the transformation of weights for inference is performed directly on GPU, which speeds up the execution of operations.
Also implemented custom extensible glTF specification “NV_texture_swizzle”, designed to describe NTC storage of materials. In addition, the feedback mode has the ability to batch transcoding of tiles, and in the sample mode, the stability and performance of inference have been improved. The logic of color channel matching and the overall quality of the LibNTC library code have also been improved.
Although the SDK is optimized for NVIDIA graphics, it can also be used on GPU Intel and AMD, including Linux platforms. According to the developers, RTXNTC will continue to develop as a universal tool for future games and 3D scenes, with an emphasis on compression without loss of visual quality.
Huawei is changing its strategy in the AI segment: instead of specialized ASIC components, the company is moving towards developing general-purpose graphics processors, supporting CUDA code. The sources say the new architecture will use an intermediate layer that automatically converts instructions from the NVIDIA platform into compatible commands for Huawei chips.
Chinese tech giant bets on reducing dependence on US restrictions, which banned the supply of NVIDIA's top AI chips to China. Despite the embargo, demand for American solutions remains, and Huawei intends to occupy this niche by adapting its chips to the familiar ecosystem of developers. In addition, the company wants to repeat the architectural approach of NVIDIA and AMD, abandoning narrow-profile solutions.
The production will be handled by SMIC, but this contract manufacturer is limited by sanctions - its process technology does not go below 7 nm. This means that Huawei will have to find a balance between compatibility, performance and the real capabilities of the Chinese semiconductor industry. However, for Chinese companies like Alibaba и Tencent the transition may become a forced necessity.
Huawei is preparing not just to replace NVIDIA, but to create an alternative AI chip platform, designed for the local market, with potential for export. It is not yet clear when the first GPU with the new architecture will go on sale, but the transition from ASIC to a universal architecture is already the beginning of a new phase.
Company Team Group announced new NVMe SSD P250Q-M80, focused on industrial and mission-critical tasks. Its main difference is Dual mode emergency data erase, including flash memory hardware self-destruction.
S/W Quick Erase function (pressing the button for 5-10 seconds) allows you to quickly delete all data programmatically, while maintaining the functionality of the drive. Even if the power is turned off, the process will be completed the next time you turn it on.
H/W Quick Erase function (holding the button for more than 10 seconds) starts high voltage breakdownWhich irreversibly destroys NAND chips. In this case, smoke and light signals are visible on the board - the SSD really burns out. Recovery after such erasure is impossible.
In addition, the drive has:
Reading speed up to 7 GB / s, records - up to 5,5 GB / s
Interface PCIe Gen4 x4, without DRAM buffer
Operating temperature range from 0 to 70°C, in non-working condition - from -55°C to 95°C
Shock resistance, vibration resistance and MTBF over 3 million hours
The P250Q-M80 is available in versions 256 GB, 512 GB, 1 TB and 2 TB, but prices have not yet been announced.
As part of the GeForce Summer campaign, the company NVIDIA Shows Limited Edition GeForce RTX 5090 in style Cyberpunk 2077. The card is made in a characteristic yellow color scheme with full-size decorative design and logos of the game. This is a special version of the Founders Edition, which will not go on sale, will be raffled off among the competition participants.
Participation is simple: to get a chance to win a video card, you need to leave a comment with the tag #GeForceSummer under NVIDIA's post on X (Twitter). The map is visually similar to the recent Superman version, but this time it's completely Cyberpunk-themed, with characters and elements from the Ultimate Edition.
The company is not disclosing whether this announcement will be followed by another wave of custom GPU, but for now the RTX 5090 x Cyberpunk remains exclusive lot, intended exclusively for fans and collectors.
Intel Completes One of the Most Important Milestones in Processor Development Nova Lake S - debugging chips on 2nm TSMC N2 process technology. This means that the company officially adopts a flexible approach, combining Intel's own 18A node with external production at TSMC, which will help avoid bottlenecks in mass production.
Processors Nova Lake S represent an extremely complex multi-component architecture with 52 cores, including:
16 performance cores (P-cores),
32 energy efficient cores (E-cores),
4 dedicated ultra-low power cores.
This set makes the chips universal - from highly loaded tasks to background services in power saving mode.
The graphics subsystem will be divided into roles: Xe3 Celestial will provide game and computational rendering, and Xe4 Druid will handle video processing and display interfaces. This reflects Intel's new approach to separating GPU- loads between different blocks within the SoC.
The memory controller has also received an upgrade, with support for modules up to DDR5-8800, which will improve throughput when running intensive scenarios including AI and graphics.
After debugging is complete, the stage begins complex validation, during which the chips undergo detailed testing for stability, temperature and electrical conditions. Only after this is it possible to launch mass production, which traditionally requires several more months.
If everything goes according to plan, Nova Lake-S to be released in Q2026 XNUMX, becoming one of Intel's first volume products to use both 18A and TSMC N2.
An AMD patent has been discovered online that describes multi-chip architecture GPU with intelligent data transmission system. Although the document formally speaks of CPU, the structure and mechanics indicate application in gaming graphics processors next generation.
The main problem with MCM (Multi-Chiplet Module) game designs is high latency when accessing memory between crystals. A new patent offers a solution in the form of "smart switch" within the Infinity Fabric network, optimized for GPU. This node analyzes each memory access, determining which is better. move task or duplicate data, reducing latency to nanoseconds.
The architecture also includes local L1 and L2 caches in each graphics chipletand shared L3 memory (or stack SRAM) that is accessed by all blocks through the switch. This structure reduces the dependence on global memory and provides fast data exchange between crystals. The approach is reminiscent of the implementation of 3D V-Cache in AMD processors, but is adapted for GPU.
Unlike data center solutions (where HBM is used), here it is proposed to use regular GDDR or stacked DRAM, connected by means of InFO‑RDL technologies from TSMC. This will allow us to assemble compact and efficient multi-chip GPU for the mass market.
AMD already has experience working with such architectures in the series Instinct MI200/MI350 and is going to unite gaming and AI solutions under the flag UDNA. This will simplify support for drivers, compilers and software libraries, accelerating the release of new products.
Although RDNA 3 already used MCD chiplets, it was delays between components have become a bottleneck. Now AMD hopes to remove that barrier with new type of switch and shared cache. It is not yet clear in which generation such architecture will appear, but UDNA 5 — the most likely candidate.
Company Colorful is preparing to release an unusual video card with two M.2 slots for SSD, located directly on the board. The new product was spotted at the exhibition Bilibili World 2025, where the company introduced a device called iGames Ultra Duo White SSD. Although the exact model of the graphics processor has not been disclosed, judging by the design and dimensions, we are talking about mid-level solution, presumably based on GPU with interface PCIe x8, such as the RTX 4060 or the upcoming RTX 5050.
On the back panel of the video card you can see two full-size M.2 slots, designed for installing NVMe drives. This solution looks unconventional, but is very practical: users will be able to add SSD without having to use motherboard slots, and also to simplify data replacement and transfer between systems. Unlike existing solutions (for example, the ASUS RTX 4060 Ti model with one M.2), Colorful offers double functionality.
It is important to note that such a decision does not require additional cables or controllers — the video card uses the available PCIe lines directly, without any loss of bandwidth. The interface operates at full x16 speed, and the drives themselves are connected via the provided power and bus wiring inside the PCB.
Perhaps such an innovation may affect the appearance of the assembly: installed SSDs, especially with radiators, can ruin the aesthetics of the case. However, for many users, the convenience and expandability of the system are more important.
Release date and price iGames Ultra Duo White SSD have not yet been announced, but Colorful is clearly hoping to attract enthusiasts who care about high density data storage right in GPU.
Corporation Intel officially announced the spin-off of its division RealSense, which developed sensors and cameras for robots. A new independent company called RapidSense has already attracted $ 50 million as part of a Series A funding round. The decision comes as part of a larger optimization strategy led by the new CEO Lip-Bu Thanom, which aims to focus Intel's efforts on semiconductor manufacturing.
RealSense is known for creating specialized sensors used in robotics, from industrial machines to autonomous systems. However, despite the rapidly growing interest in humanoids, Intel has decided to abandon this niche. Analysts Morgan Stanley estimate the future market for humanoid robots at $5 trillion by 2050, Ilon Mask already calls robots Tesla's priority over cars.
However, Intel is staying the course Accelerating production of our own chips, as well as the reorganization of the business structure. According to insiders, the corporation may even refuse to introduce its 18A solutions to the external market within the framework of Intel Foundry, switching to the less risky node 14A.
Previously, former Intel CEO Pat Gelsinger admitted that the company missed the mark with AI products and lost its leadership NVIDIA. The latter is actively developing not only AI accelerators, but also solutions for robotics. Against this background, Intel's departure from the RealSense segment looks like a strategic refusal to participate in the race of the future - at least in its hardware part.
Overclocker bl4ckdot has set a new world record for DDR5 memory overclocking, reaching speeds of 12872 megatransfers per second (MT/s) on the platform Intel Z890. The record was set using modules G.Skill Trident Z5, processor Intel Core Ultra 9 285K and the motherboard ASUS ROG Maximus Z890 Apex.
This result significantly surpassed the previous record set by saltycroissant (12843 MT/s), which lasted for more than a month - quite a rare case in the rapidly developing field of extreme overclocking. To achieve this frequency, bl4ckdot used a liquid nitrogen for memory cooling. The system configuration with timings CL68-127-127-127-2 has been officially validated in the database CPU-Z and entered into the HWBOT table.
G.Skill noted that the overclocking was performed on the basis of SK Hynix chips, which form the basis of the Trident Z5 series. It is this memory that is now most often used in record sessions due to its good compatibility with the Intel platform and high frequency stability.
At this point in Memory Frequency Hall of Fame table bl4ckdot takes the top spot, followed by saltycroissant and Seby. Only two overclockers, bl4ckdot and saltycroissant, managed to break the 6400 MHz mark (effectively 12800 MT/s) with DDR5, but others like Hicookie are also getting close to those numbers.
According to Bloomberg, large-scale construction is underway in Yiwu County in China's Xinjiang Uyghur Autonomous Region to create new data centers, in which it is planned to use over 115 Nvidia graphics acceleratorsThese chips are necessary for processing large amounts of data and training LLM models, similar to GPT or DeepSeek R1.
The Yiwu complex will become part of "Economic zone for demonstration of computing power" and will receive energy from solar and wind power plants, located in the Gobi Desert. The projects are funded by regional authorities and are aimed at supporting the development of Chinese AI services. However, there are no indications that How exactly do Chinese companies plan to obtain banned chips?, not in official documents.
Since 2022, the United States has imposed restrictions on the supply of advanced GPU from Nvidia to China, including H100, H200 and Grace Blackwell. In this case, US authorities admit existence of illegal imports, but according to White House estimates, there are currently no more than 25 such chips, and there is no data on the centralized use of such equipment. Nvidia also denies supporting data centers with contraband accelerators and says such operations do not make economic sense.
Nevertheless, Construction in Xinjiang continues, and local authorities are showing a strong commitment to comply with the national strategy for the development of artificial intelligence. Even if not all 39 declared objects are completed, the very fact of planning on such a scale is causing concern in the US, where they believe that access to advanced AI technologies could give Beijing a military-technological advantage.