NVIDIA has encountered unexpected problems in developing its first ARM processors for Windows-PC, which forced the company postpone the release until 2026According to SemiAccurate, the cause was critical hardware defects in the N1 and N1X chips, which were originally planned for release in 2025.
Let us recall that NVIDIA’s plans to enter the consumer PC processor market were first announced back in 2023. However, at the March GTC 2025, the company presented only GB10 Superchip — a powerful SoC developed jointly with MediaTek for workstations. No details about the versions for ordinary users. Similarly, at Computex in June The N1/N1X chips were not shown, which indirectly confirmed the postponement of the deadlines.
Rumor has it that the first bug has been fixed. without new process technology, only due to microcode and firmware. This temporarily allowed NVIDIA to move the deadlines forward to early 2026. However, second defect turned out to be more serious: now its elimination requires full-fledged “respin” - new photo template and redesign of the chip.
Respin process on TSMC 4nm process technology may take from three to six months, depending on the complexity of the changes. If the defect affects low metallization levels or logic, a complete redesign and verification will be required. In addition, even with a quick redesign, NVIDIA will have to retest all the functions and parameters of the chip, which will also increase latency.
NVIDIA and MediaTek have not officially commented on the situation. But given the current difficulties with the architecture and production cycle, full-fledged consumer NVIDIA ARM Processors for Windows-PC should not be expected before the second half of 2026.
After a six-year hiatus, Razer has unveiled a new external enclosure for graphics cards — Core X V2, based on Thunderbolt 5. A separate one is also published along with it. Thunderbolt 5 dock, designed to compensate for the missing functions. The new eGPU- the case supports full-size PCIe 4.0 video cards, including modern models GeForce RTX 50 и Radeon RX 9000, and also provides up to 140W of charging via USB PD. However, despite its advanced features, the device has lost a number of familiar functions.
Unlike previous models in the series, The Core X V2 does not have a built-in power supply. - users must install an external ATX PSU on its own. It also lacks USB ports, Ethernet, and RGB lighting, making it a strictly functional solution. Thunderbolt 5 delivers up to 120Gbps one-wayBut PCIe 4.0 x4 limitation (64 Gbps) remains a bottleneck for video cards - roughly at the level of the cheaper and simpler OCuLink.
It is also worth noting the complete lack of support MacOS — after switching to Apple Silicon, Apple officially stopped supporting eGPU. Core X V2 is only compatible with Windows and a limited number of Thunderbolt 5-enabled laptops.
Thunderbolt 5 dock completes the picture: there are 3x 4K@120Hz, USB-A/C 10Gbps, 140W charging, Gigabit Ethernet, a slot for M.2 SSD and Thunderbolt Share feature for transferring files between two PCs. However, the price starts at $389.99, and the Chroma RGB version is $399.99. The Core X V2 case itself will cost $350 - excluding the PSU and graphics card.
AMD has revealed details about the processor Ryzen AI 5 330, intended for Ultrabooks and compact PCs, offering AI-powered performance Krackan Point architecture. It quad core processor с 12 streams, consisting of 3 full Zen 5 cores and 1 power-efficient Zen 5C, with a maximum frequency of up to 4.5 GHz i podderjkoy SMT.
The chip is built using the process technology TSMC 4nm FinFET, It has 8 MB L3 и 4 MB L2 cacheAt TDP 28 W (adjustable from 15 to 28 watts). It has integrated graphics Radeon 820M with two RDNA cores and a frequency of up to 2.8 GHz, and also supports up to 256GB of LPDDR5X-8000 and DDR5-5600 memory on a dual-channel architecture.
The main feature of the Ryzen AI 5 330 is the integrated neural processing unit (NPU) with performance up to 50 TOPS, which makes it attractive to local AI acceleration in consumer and office tasks.
Support PCIe 4.0, USB4, HDMI 2.1, DisplayPort 2.1 with Adaptive Sync, output to 4 displays, AV1/HEVC/H.264 decoding up to 8K60, and built-in Miracast wireless transmission make the chip a versatile solution for multimedia and AI applications. no ECC and AMD PRO support, which emphasizes its consumer focus.
Professional video card AMD Radeon AI Pro R9700 appeared in online stores two weeks before the official launch, according to a leak from an insider MelodicWarrior, two versions from Sapphire - VCX 32358-01-20G и VCX 32358-01-21G — appeared on sale at prices $ 1244 and $ 1277 respectively. Both models are already marked as "Out of Stock", but are available for order with the note "Backordered", which indicates fast demand or early listing.
MelodicWarrior also reported that MSRP will be "reasonable", considering the professional focus of the card. However, current retail prices, although not prohibitive for this class of devices, still exceed regular gaming solutions at the same level of performance. Both versions of the card are equipped with 32 GB GDDR6 memory, 256-bit bus and, judging by the markings, aimed at the retail segment with AI processing capabilities.
Radeon AI Pro R9700 Could Be AMD's Answer to Demand for export-safe neural network accelerators in countries with limited access to NVIDIA's HBM chips and Blackwell architecture. AMD was previously reported to be working on several GPU, which comply with new export control requirements, including for the Chinese market.
After the US eased export restrictions NVIDIA May Cancel RTX 5090D V2 Release — a cut-down version of a graphics card previously designed specifically for the Chinese market. According to insiders and supply chain sources, the company is now leaning toward resuming sales of the regular RTX 5090D in China, thereby calling into question the need for a simplified modification.
The RTX 5090D was initially sold in China, but its shipments stopped due to strict restrictions. In response, NVIDIA began developing the RTX 5090D V2, an even more stripped-down model that would meet the new requirements. However, with the resumption of sales of H20 accelerators and the possible lifting of a number of other restrictions, the decision may be revised: NVIDIA has RTX 5090D stock and may bring it back to market without going through the V2 version.
While RTX 5090D V2 shipments still haven't started, and AIB partners did not receive chips. At the same time, rumors that V2 will be too weak and impractical against the backdrop of regulatory changes are only growing stronger. In conditions when The recovery of sales of the original RTX 5090D looks more profitable, and the H20 is already ready for mass delivery, the V2 project may be cancelled before the announcement.
After receiving export approval GPU Nvidia H20 in China, an analytical company TrendForce reported that the Chinese market has seen a sharp increase in demand for these accelerators. Amid months of sanctions from the United States, deferred demand among Chinese cloud service providers and OEMs can provide explosive volume of purchases already in the second half of 2025.
TrendForce notes that the share of foreign AI chips in the Chinese market will grow to 49%, vs 42% earlier, despite the support of local players like Huawei и Cambricon. The reason is restrictions on access to advanced technological processes, including the inability to order chips from TSMC. H20 is built on the architecture Hopper and uses HBM3 8hi, but in 2025 the transition to HBM3e, which will further increase interest in NVIDIA chips, especially against the backdrop of a shortage of high-speed memory in China.
Although the H20 is limited in performance to meet US requirements, Jensen Huang emphasizes that legal export and control is better than encouraging local competitors to develop. However, TrendForce warns — China will continue to actively develop its own ecosystem, as the market remains vulnerable to geopolitical risks, NVIDIA dependence is a temporary advantage.
NVIDIA plans to significantly increase production of memory modules SOCAMM (System-on-Chip Advanced Memory Module) for its AI products. This modular memory is designed to simplify memory upgrades in devices, while providing high performance and energy efficiency.
SOCAMM memory presented at the event NVIDIA GTC, differs from traditional HBM and LPDDR5X used in servers and mobile platforms. It is based on LPDDR DRAM, but, unlike soldered solutions, SOCAMM is modular and can be replaced (fastened with three screws).
According to the Korean publication ETNews, NVIDIA intends to produce 600 to 000 SOCAMM modules this year. This significant increase in production is intended to be implemented into NVIDIA's AI product stack. One of the first platforms to use SOCAMM is the latest GB300 Blackwell. Although current volumes are below HBM supply, SOCAMM production is expected to scale up next year, especially with the advent of SOCAMM 2.
SOCAMM offers a compact and modular form factor that significantly more energy efficientthan RDIMM. It also provides higher bandwidth than RDIMM, LPDDR5X, and LPCAMM while offering up to 150-250 GB/s bandwidth memory. The ability to replace makes SOCAMM a universal solution for AI PCs and AI servers.
Micron is currently NVIDIA's primary SOCAMM manufacturer, but Samsung and SK Hynix are also in talks to produce the modules.
Technology Neural Texture Compression (NTC) from NVIDIA, which uses neural networks to compress and decompress game textures with minimal loss of quality, shows significant improvements when combined with DirectX Raytracing 1.2 from Microsoft. The latest tests show a significant reduction in video memory (VRAM) consumption and a noticeable increase in performance.
User @opinali ran tests where the combination of NTC and DirectX Raytracing 1.2 resulted in massive reduction in VRAM usage. This became possible thanks to a new feature Cooperative Vectors in DXR 1.2, which allows shaders GPU more efficiently work with matrix and vector operations. As a result, the NTC compression/decompression mechanism works efficiently in standard game shaders via DX12, reducing VRAM consumption.
Testing has shown that with Cooperative Vectors and NTC enabled in "Default" mode, texture rendering reaches 2350 FPS. Disabling these features (DP4A mode) reduces performance to 1030 FPS, which accounts for almost 80% of the difference.
At the moment NTC technology is only available for GPU from NVIDIA, as Intel and AMD's similar neural rendering kits are not yet complete. NTC works with the latest 590.26 preview drivers, which also include NVIDIA Smooth Motion technology. These software improvements are expected to have a significant impact on the performance of next-generation GPUs.
GPD continues to develop its next-generation portable gaming device, with a Geekbench leak showing the system codenamed GPD-G1618-05, which is presumably GPD win 5. This prototype became the first known portable device equipped with the most powerful processor with integrated graphics to date - AMD Strix Halo APU.
Geekbench leak confirms flagship usage 16-core Ryzen AI MAX+ 395 Halo APU. The data shows that the processor can reach the clock speed over 5.1 GHz.
The system memory attracts special attention. The device uses DDR5 memory, which is an unusual choice, as LPDDR5 would save space on the PCB. Also, the memory capacity is 24GB, but Notebookcheck speculates that this could be 32GB system with 8GB reserved for VRAMIf so, GPD Win 5 will outperform devices like ROG Ally or MSI Claw, which have 24GB of memory.
Despite the powerful configuration, the cost Strix Halo APU remains high, which raises concerns about the final price of the device (similar mini-PCs based on it cost more than $2000). There are also suggestions that the prototype can work without a built-in battery, which requires confirmation.
These are updates for the driver R600 Gallium3D, the Mesa component responsible for supporting the specified GPU, courtesy of developer Patrick Lerda. One of the key changes was a fix to the Depth Compare function, which is now correctly set to "NEVER" when depth comparison is disabled. This solves issues with OpenGL tests on Radeon HD 2000/3000/4000, increasing Compatibility with 3D applications and games.
Also included is a fix for Border Color handling for OpenGL on GPU RV770-based GPUs (e.g. Radeon HD 4850 and HD 4870), which fixes about 120 bugs in OpenGL tests. This improves OpenGL compliance for the entire R600 family.
Several additional fixes are still pending and will likely be added to the stable release of Mesa 25.1. These patches are aimed at fixing bugs rather than adding new features. Overall, the open source community has shown a high commitment to supporting older hardware, allowing these GPU for retro games or extend the life of old systems.
