📁 Hardware

Questions are emerging regarding the timing of Nvidia's initiative to adopt 800V power systems in data centers. While the company is pushing for this technology, suppliers report a lack of clarity on release plans, creating uncertainty for those designing high-density infrastructures for AI workloads.

AMD has unveiled its first estimated benchmark results for the upcoming 256-core EPYC Zen 6 'Venice' processor. The company claims this CPU delivers 3.3 times higher rack-level performance compared to Nvidia's Vera platform. These preliminary figures mark a significant move by AMD in the data center market, aiming to strengthen its position against key rival Nvidia with solutions optimized for efficiency and compute density.

Google has reportedly entered an agreement with Intel for the packaging of over 3 million TPU units by 2028. This strategic move highlights the increasing complexity of the AI hardware supply chain and the critical role of advanced packaging technologies, such as EMIB, for high-performance HBM integration. The collaboration underscores the need for diversified suppliers and specialized manufacturing capabilities to meet the demand for AI accelerators.

AMD is developing new code for its AMDGPU Linux kernel driver to support HDMI 2.1 compliance testing. This effort is part of the company's commitment to providing a fully open-source HDMI 2.1 driver implementation, including features like FRL and Display Stream Compression. A mature, open-source driver is crucial for enterprise adoption and on-premise deployment strategies, ensuring control and transparency.

The Mesa Radeon Vulkan driver (RADV) now leverages the INST_PREF_SIZE feature in AMD's RDNA3 and RDNA4 GPUs. This optimization enhances instruction prefetching, a critical aspect for GPU efficiency. For CTOs and infrastructure architects deploying on-premise AI workloads, this development is significant for maximizing hardware performance and optimizing Total Cost of Ownership (TCO).

Samsung is considering building a new chip packaging plant in Gwangju. This strategic move is driven by growing energy supply difficulties that are limiting the expansion of its semiconductor production operations in the Seoul area. The decision highlights the infrastructural complexities and energy constraints influencing deployment and production strategies in the tech sector, with implications for AI infrastructure as well.

Taiwan is outlining an ambitious industrial strategy for artificial intelligence, identifying silicon photonics as a key element to solidify its competitive advantage. This strategic move aims to strengthen the island's position in the global AI supply chain, focusing on advanced interconnection technologies essential for demanding workloads.

NVIDIA has listed its RTX PRO 6000 Blackwell Workstation Edition at $13,250 on its official marketplace. This pricing highlights the significant investment required for dedicated on-premise AI hardware solutions, offering professionals total control over workloads and data sovereignty, despite a high CapEx. The GPU targets those seeking high performance and autonomy for Large Language Model development and inference.

Elon Musk has unveiled details of his first orbital data center, the AI1 Satellite. This platform, wider than a Boeing 747, is designed to host a 120 kW compute payload, peaking at 150 kW, and integrates an interchangeable chip system. The initiative marks a step towards advanced data processing in space, offering new perspectives for AI and LLM workloads.

Custom NVIDIA V100 cards have emerged from China, featuring a single-slot, half-height design with NVLink. These GPUs, available in 16GB and 32GB VRAM versions, offer full performance with flexible power options (75W or 300W). With an estimated price below $220, they represent an intriguing solution for compact, low-cost on-premise deployments, especially for LLM inference workloads.

Tencent is adopting a "dual-track" approach to AI chip development, combining its proprietary Canghai V2 processor with strategic domestic partnerships. This strategy aims to strengthen supply chain control and optimize performance for artificial intelligence workloads, reflecting a growing emphasis on technological sovereignty and operational efficiency for large-scale deployments.

A recent analysis highlights a significant leap in RISC-V CPU performance, with improvements of up to eight times over five years. The comparison between the new SpacemiT K3 SoC, a first-to-market RISC-V RVA23, and the five-year-old SiFive HiFive Unmatched board, reveals the rapid evolution of RISC-V hardware. This progress opens new perspectives for on-premise deployments and edge solutions, offering increasingly competitive alternatives.

Computex reaffirmed Nvidia's dominant position in the artificial intelligence hardware landscape. The event highlighted how the silicon giant's solutions have become a cornerstone for the development and deployment of Large Language Models, profoundly influencing infrastructure strategies, especially for those evaluating self-hosted options and data sovereignty.

SuperAlloy is focusing its strategy on the semiconductor supply chain, promoting the use of recycled aluminum. This initiative aims to integrate sustainability into a key sector for technological innovation, responding to the growing demand for supply chain resilience and responsible resource management. The adoption of recycled materials can positively impact TCO and hardware stability for on-premise AI infrastructures.

A user has repurposed an NVIDIA Jetson Orin NX for on-premise Large Language Model (LLM) inference, transforming it from a bulky server into a compact, silent solution. The goal was to exceed 10 tokens/s and support a 65K context window for Hermes Agent, with a 40W power consumption. Tests with Gemma 4 26B A4B UD Q2_K_XL confirmed a 66K context window and performance of 14.65 tokens/s at 8K context, dropping to 10.21 tokens/s at 60K, highlighting the potential of LLMs on edge hardware.

Researchers at Georgia Tech have released Vortex 3.0, a new version of their fully Open Source RISC-V GPGPU. This OpenCL-compatible implementation introduces a 3D pipeline, expanding its capabilities beyond general-purpose computing. The initiative highlights the growing interest in open hardware solutions, offering new perspectives for on-premise deployments and control over the technology stack.

Linux developers are leveraging AI-assisted development tools, such as GitHub Copilot, to modernize and optimize drivers for vintage AMD GPUs. This approach has enabled the cleanup of the R600 driver, extending the lifespan of graphics cards from the HD 2000 to HD 6000 series. It's a concrete example of how AI can contribute to sustainability and efficiency in managing legacy hardware, with positive implications for on-premise deployments and TCO.

China's National Medical Products Administration has greenlit NEO, a coin-sized brain-computer interface (BCI) developed by NeuraMatrix and Tsinghua University. Aimed at patients with spinal cord injuries, this implant marks the entry of a BCI product into the commercial market, transforming the global competition in neurotechnology from theoretical to concrete.

A Chinese startup has announced an innovation in photonic chip production, bypassing expensive DUV lithography. Using a nanoimprint process, the company claims to cut production costs by up to 90% for 8-inch wafers, promising a significant impact on the semiconductor industry and the accessibility of AI hardware.

Apple is enhancing its Photos app with new artificial intelligence-driven editing capabilities. Among these, "Reframe" stands out as a spatial feature enabling users to adjust image perspectives directly on their device. This innovation highlights the increasing integration of AI at the edge computing level, a trend that raises questions about on-device processing capabilities and data management, key topics for those evaluating AI deployments.