Hardware News

JEDEC Approved the SPHBM4 Standard — HBM4 Will Become Cheaper Without Losing Speed The JEDEC organization, responsible for defining specifications for standard memory types, has published the JESD330-4 document, which describes the final specification of the SPHBM4 (Standard Package High Bandwidth Memory) standard. The new memory standard is designed to provide the same total bandwidth as HBM4 memory, with greater capacity and lower integration costs due to compatibility with traditional PCB (organic) substrates. JEDEC announced work on the final specification of the SPHBM4 standard in December 2025. In the final version of the SPHBM4 standard specification, the same HBM4 DRAM memory chips and base memory layers are retained, but a different base chip interface is presented. While the regular HBM4 interface uses 2048 data signals, SPHBM4 reduces this number to 512 signals, compensating for the narrower connection with higher interface frequencies and a 4:1 transmission sequence. “The JESD330-4 SPHBM memory is designed to operate with the same aggregate bandwidth as HBM4, using fewer contacts while operating at a higher frequency. If the HBM4 interface contains 2048 data signals, SPHBM4 will define 512 data signals with a 4:1 sequence to achieve the same bandwidth. This change allows for a reduction in the bump height required for connection to organic substrates,” stated JEDEC. JEDEC explained that the 4:1 sequence means that each data transmission connection in SPHBM4 memory transmits four times more data over time. This allows the interface to maintain the overall bandwidth at the HBM4 level while using a greater distance between the centers of two adjacent connecting microcontacts when combined with organic substrates. The memory volume available for each stack remains unchanged, as SPHBM4 uses standard HBM4 memory layers. Routing on the organic substrate also supports longer connections between the processor and memory, which may allow accelerator developers to place more memory stacks around the SoC. SPHBM4 memory is not intended to replace standard HBM4. It merely changes the way high-performance memory stacks interact with the processor and how they are integrated into a single package. JEDEC has not provided any information about products that will use SPHBM4 memory or the timeline for the implementation of the JESD330-4 standard. If you value what we do, you can support our projects on Boosty. With love, NSWTL.

Chinese Company Unveils Smartphone That No Longer Needs Apps — Everything is Done by an AI Agent Artificial intelligence functions in modern smartphones are scattered across different apps and deep within the operating system, making them not always easy to use. The Chinese company StepFun has introduced the StepX Neo, which it claims is the world's first mass-market smartphone with an AI agent that independently handles multi-step tasks from start to finish, according to Gizmochina. Repeated requests and opening apps are not required. The interface of the StepX Neo is based on the AI agent Step Amoo, developed by the manufacturer and integrated into the Step AOS operating system, which is based on an open version of Android. Users can interact with Amoo like a chatbot, operating in a fully autonomous mode without requiring constant internet connectivity. The agent accepts natural language requests and performs tasks across various apps, web services, and built-in device tools until a complete result is achieved. Over time, the system learns the owner's preferences, such as prioritizing flights that allow pets and pre-booking a table at a favorite restaurant if needed. While traveling abroad, the StepX Neo can translate conversations, text messages, calls, road signs, restaurant menus, and museum posters in real-time — supporting 32 languages. The smartphone will display public transport schedules, find nearby events, or analyze routes, even if the user is in a remote area without mobile connectivity. It will help fill out customs forms, note visa requirements, and remind users of flight check-in times. The architecture of Step AOS is based on Android, Linux, and RTOS components and is divided into four function groups using the MCP protocol: communication, applications, files, and system tools. The AI agent can freely combine them to perform tasks. There is also a voice and visual input component, NUI, which learns user habits: users do not need to manage the device in the traditional sense — they just need to indicate what they want to do. The device's memory is divided into user and agent-owned sections — the latter moves data through the "capture — understand — store" pipeline. A multi-level decision-making mechanism is provided: simple requests are directed to small, fast models, while complex ones go to resource-intensive ones. The security system is based on trusted execution, complete audit logs, permissions provided only when necessary and revoked immediately after task completion; erroneous actions can also be undone with a single touch. To cover a wide range of functions, StepFun intends to ensure integration with Chinese services like Ctrip, Alipay, Didi, Meituan, WPS, and CapCut — allowing tasks related to travel, payments, taxi services, local services, office work, and content creation to be solved. The release date and cost of the StepX Neo smartphone have not been specified. If you value what we do, you can support our projects on Boosty. With love, NSWTL.

Google Pixel 11 may surpass iPhone 18 and become the first smartphone with a 2nm chip The Google Pixel 11 series smartphones will be powered by the Tensor G6 processor—its technical specifications are not yet known, but according to unofficial information, it will feature a more powerful central processor but a less efficient graphics processor. It will be the first mobile chip on the market manufactured using 2nm technology. The world's largest semiconductor contractor has started mass production using the 2nm process; it will also be used in Google Tensor G6 chips, as reported by IThome. Typically, Apple is the first to access new TSMC solutions, but this year Google seems to have outpaced the iPhone manufacturer. As a result, the Pixel 11 series smartphones will be the first major flagships with processors based on TSMC's new process, and a month later, the iPhone 18 series with Apple A20 chips will be released. The 2nm process, like all new solutions, will help improve the performance and energy efficiency of the processors in the Pixel 11. Interestingly, mentions of MediaTek algorithms were found in the U.S. Federal Communications Commission (FCC) database during a radio frequency component test—instead of the traditional Samsung modem, the new Tensor G6 will work with a modem from the Taiwanese manufacturer. The official announcement of the Google Pixel 11 smartphones is scheduled for August 12. If you value what we do, you can support our projects on Boosty. With love, NSWTL.

New York Becomes the First in the US to Implement a Year-Long Ban on Building Large Data Centers New York Governor Kathy Hochul signed the first moratorium in the US on the construction of data centers with a capacity of more than 50 MW in the state for a year. Meanwhile, a bill passed by the state legislature, which could block even more data center construction projects, is still awaiting her signature, noted The Verge. According to the governor's order, the issuance of environmental permits for the construction of large data centers is halted for a year, which, as explained by the governor's office, will give the state time to develop appropriate measures to protect residents from rising energy costs and environmental impact. The maximum data center capacity specified in the order is higher than the 20 MW threshold approved by state legislators. According to state leadership, this limit will help prevent disruptions in the operation of smaller data centers used by institutions such as hospitals. It is currently unknown whether the governor will sign the moratorium passed by the legislature, but her executive order allows state authorities to pause while the bill is under consideration. "As the development of data centers threatens to increase utility bills, deplete our natural resources, and create uncertainty for New York residents, it is my duty to take action and lead this process," Hochul stated in her announcement. The moratorium could last up to a year while the Department of Public Service (DPS) develops standards to assess the expected environmental impact of the construction and operation of data centers, including water consumption and air quality. Hochul suggested that the DPS consider creating a mechanism that provides for data centers to invest in the state's energy infrastructure and instructed the state development agency to develop a framework to help local communities negotiate benefits if data centers are still placed in the city. Maine could have been the first state to introduce a moratorium on data center construction, but its governor vetoed the bill in April this year. If you value what we do, you can support our projects on Boosty. With love, NSWTL.

Nintendo will soon release an updated Switch 2 with an OLED display, but it's not certain This fall, Nintendo will release an updated version of the Switch 2 console with a replaceable battery, but the company has other plans for the future. The Japanese manufacturer is also considering releasing a version of the Switch 2 with an OLED display with Full HD resolution (1920 × 1080 pixels). A major update to the Nintendo Switch 2 may involve the use of rigid OLED panels; Samsung Display, which supplied OLED screens for the first Nintendo Switch, is eager to secure the contract for their production, reports ZDNet Korea. Of course, this won't happen quickly. Amid the rapid rise in the cost of memory and other important semiconductor components, the main obstacle remains the price difference between LCD and OLED displays. If the Japanese manufacturer decides to release the Switch 2 with an OLED screen, the development of the updated model may not begin until the end of this year. This means that mass production will start at the end of 2027 or the beginning of 2028. Nintendo traditionally updates its products in stages to extend their life cycle. The first-generation Switch with an LCD display was released in 2017, the Switch Lite version in 2019, and the Switch with an OLED display appeared only in 2021. All three versions of the first-generation Switch were equipped with screens with a rather modest resolution of 720p, while the Switch 2, debuting in 2025, received a Full HD screen. The transition to the OLED variant around 2027 aligns with the usual product update schedule of the Japanese brand. Meanwhile, the company will have to find a balance between improving product quality for players and the harsh reality of rising production costs. If you value what we do, you can support our projects on Boosty. With love, NSWTL.

MSI confirmed compatibility of Chinese CXMT DDR5-8000+ memory with Intel 800 series boards MSI announced successful testing of new BIOS versions for Intel 800 series motherboards with full optimization for DDR5 memory modules based on chips from the Chinese manufacturer CXMT (ChangXin Memory Technologies). Earlier this month, the company also announced the compatibility of CXMT DDR5-8000+ memory with AMD AM5 motherboards. Compatibility tests for Intel 800 series boards were conducted on the MEG Z890 Unify-X motherboard with two DIMM slots and the Pro Z890-S WIFI model with four DIMM slots. On the MEG Z890 Unify-X model, a dual-channel memory kit based on CXMT chips with a total capacity of 48 GB (2 × 24 GB) was tested. The memory operated in DDR5-8600 mode with timings of CL46-56-56-134. An Intel Core Ultra 5 245K processor was also installed on the board. The memory successfully passed the MemTest check with over 100% coverage. Memory testing on the Pro Z890-S WIFI motherboard was also conducted in dual-channel mode. For this, the manufacturer used a memory kit with a total capacity of 32 GB (2 × 16 GB). The memory operated in DDR5-8200 mode with timings of CL44-56-56-132. An Intel Core Ultra 5 250K Plus processor was also used in the testing. The memory successfully passed the MemTest check with over 100% coverage. MSI notes that thanks to optimization implemented on both Intel and AMD platforms, users can expect stable support for CXMT memory on its motherboards. Optimizations for CXMT memory are also integrated into the proprietary MSI Memory Try It! quick RAM overclocking feature. If you value what we do, you can support our projects on Boosty. With love, NSWTL.

AI Boom Allowed China to Nearly Double Chip Exports in the First Half of the Year Amid the global boom in artificial intelligence technologies as a key driver of economic growth, China nearly doubled its chip exports in the first half of 2026 compared to the same period the previous year. This was reported by the South China Morning Post citing customs data. In the first six months of 2026, China exported 179.44 billion integrated circuits totaling $177.28 billion, which is more than 96% higher than the figure for the same period last year. The sharp increase in chip exports was one of the main factors ensuring the country’s double-digit export growth during this period—industrial robots and other high-tech products are also in high demand. Beijing is actively promoting the adoption of domestically developed chips, aiming to strengthen the country's independence in the semiconductor industry. However, authorities were forced to allow several Chinese companies to purchase American AI accelerators Nvidia H200. One of the main factors in the growth of foreign trade was AI technologies and their applications: exports of data processing equipment and related components (PCs, servers, memory modules, and other components) in the first half of the year grew by 41.3% to $138.08 billion. Chinese manufacturers also increased exports of AI consumer goods: glasses, translators, and robotic exoskeletons. The export of Chinese robotics shows positive dynamics: by the end of the first half of the year, exports of industrial robots grew by 18.6% to 6.29 billion yuan ($927.7 million); last year, China became a net exporter of this equipment for the first time. Significant growth was also seen in the export of Chinese surgical robots and cleaning robots. If you value what we do, you can support our projects on Boosty. With love, NSWTL.

Optimization Wonders: Chinese Researchers Accelerate Neural Networks by 149 Times Without Increasing Chip Performance It's no secret that in the race for anything by any means, process optimization is the first to suffer. All the issues with resource-intensive AI are precisely about this. Optimization is only considered when the resource for continuing the race runs out or is no longer sufficient for everyone. With their sanctions, the US has put Chinese developers in a position where optimization becomes the only way out of the resource shortage crisis. But this is not just a way out — it's a chance for a breakthrough. Specifically, researchers from Peking University created an experimental computing system in which several ordinary electronic processors exchanged data through silicon photonic transmitters and an optical switch. During experiments, the platform performed a neural network noise reduction task almost 149 times faster than a graphics processor, although its nominal computing performance was only about 11.6% of the GPU's performance: about 1.97 versus 16.96 Tflops. The system consisted of five field-programmable gate arrays (FPGAs), each hosting one layer of a five-layer convolutional neural network. Silicon photonic transceivers with a speed of 400 Gbps and a 16 × 16 optical switch were installed between the boards. Four data transmission channels of 100 Gbps each were transmitted over a common fiber at four wavelengths. The switch had losses of less than 5 dB and could theoretically connect up to 16 computing chips with a combined bandwidth of 6.4 Tbps. The main advantage of the solution was pipeline processing. Each FPGA continuously processed its neural network layer and immediately transmitted the intermediate result to the next chip via the optical channel. This eliminated the need to save data to external memory after processing each layer and then reload it — such operations create the so-called "memory wall" and slow down GPU performance. A thousand images sized 32 × 32 pixels were processed in 105.16 microseconds, while the control GPU took 15.643 milliseconds for the same operations, and the efficiency of the FPGA computing blocks reached an impressive 94.7%. For the experiment, scientists used a simple five-layer neural network with 5 × 5 kernels and the Fashion-MNIST dataset. It would be tempting to see similar experiments with larger language or other modern generative models. This hasn't happened yet, but the experience already provides a clear idea of the enormous potential for AI acceleration with proper optimization. "Specific tasks can be implemented with limited computing resources when algorithms, processor microarchitectures, and chip-level interconnections are jointly developed," the authors of the work reasonably note. "This structure can also reduce irrational energy consumption in data centers and optimize latency or consumption in advanced computing scenarios." If you value what we do, you can support our projects on Boosty. With love, NSWTL.

Amazon Reveals Colors of Upcoming Google Pixel 11 Smartphones Less than a month remains until the official announcement of the Google Pixel 11 series smartphones. Ahead of this event, pages for three upcoming models briefly appeared on Amazon, revealing their color options, as noted by 9to5Google. Three products briefly appeared on the Amazon platform as items from the Google Store lineup. The general list of colors for the base Pixel 11 includes "Obsidian," "Hibiscus," and "Pistachio," but the description pages mention previously leaked and corresponding "Midnight," "Fuchsia," and "Moss." The model index 4CS4, which appeared alongside Pixel 11 in previous leaks, is also mentioned. Subsequently, pages for the Pixel 11 Pro were found in the colors "Dune," "Light Fog," "Pine," and "Sterling." For the foldable Pixel 11 Pro Fold, the colors "Pine" and "Midnight" are indicated. A detailed examination of the Pixel 11 pages revealed a mention of a 6.3-inch display with a resolution of 1080 × 2424 pixels, 12 GB of RAM, and 256 GB of internal storage, priced at $899. Thus, Google likely does not intend to reduce the smartphone's memory to 8 GB. Interestingly, many descriptions mention a compatible gadget, the Google Pixel Tag, although such a tracker does not yet exist. The battery capacity for the Pixel 11 is listed as 4985 mAh, for the Pixel 11 Pro XL as 5115 mAh, and for the Pixel 11 Pro Fold as 4750 mAh. The official announcement is scheduled for August 12. If you value what we do, you can support our projects on Boosty. With love, NSWTL.

Founder of DeepSeek Becomes the Wealthiest AI Model Developer in History The fortune of Liang Wenfeng, the founder of the Chinese company DeepSeek, more than doubled after the latest investment round—from approximately $16.7 billion to $36 billion. According to the Bloomberg Billionaires Index, this made him the wealthiest entrepreneur among founders of companies specializing directly in the creation of large artificial intelligence models. The Chinese entrepreneur significantly surpassed Anthropic co-founder Dario Amodei and OpenAI co-founder Greg Brockman. The sharp increase in his fortune is linked to the June funding round for DeepSeek, which reached over $7.4 billion. Investors valued the Chinese startup at approximately $50 billion, whereas in April, its estimated value was around $10 billion. Liang himself invested about $3 billion in the round and, according to Bloomberg's calculations, retained about 78% of the company after the placement. Thus, the entrepreneur's fortune remains predominantly "on paper": it is determined by the estimated value of his stake in the private part of DeepSeek, whose shares are not publicly traded. In the next funding round or with a change in the company's valuation, Liang's fortune could change significantly. His high position in the wealth rankings is explained not only by DeepSeek's growth but also by the fact that he retained a much larger share of the business than many American founders who raised capital over several rounds. In the overall billionaire rankings, Liang Wenfeng holds the 63rd position. In China, he is approximately in eighth place by this parameter. It should be noted that Bloomberg compiled the ranking with the founder of DeepSeek in a high position without considering the entire spectrum of AI business (from chip development to data centers). Otherwise, Elon Musk, with his trillions, including for AI models, would claim the top spot. Despite the partly "on paper" basis of DeepSeek's founder's success, the company still has room for growth, which relies on very real "hardware"—according to rumors, the company is developing its own chip for AI inference and intends to move away from Nvidia and Huawei equipment. As for the leaders of Anthropic and OpenAI, they should more frequently withdraw their advanced AI models from general use—this is a "sure" path to success. If you value what we do, you can support our projects on Boosty. With love, NSWTL.