➀ NVIDIA and AMD plan to launch scaled-back AI GPUs for China in Q3 2025, complying with U.S. export restrictions;

➁ These GPUs are tailored to support Chinese AI frameworks like DeepSeek while adhering to new regulations;

➂ The move reflects efforts to maintain market presence amid geopolitical trade constraints.

➀ The RTX 5070 and RX 9070 are closely matched in 1080p gaming, but AMD gains a notable edge at 1440p and 4K resolutions thanks to its RDNA4 architecture and Infinity Cache design.

➁ NVIDIA maintains advantages in ray tracing (especially path-tracing), software ecosystem (DLSS 4, RTX tools), and current market availability, while AMD offers superior raw performance per dollar at higher resolutions.

➂ With both GPUs priced at $549 MSRP, real-world pricing and availability significantly impact value, creating a rare near-tie in the mid-range GPU market.

➀ Q1 2025 semiconductor revenues reached $167.7 billion, up 18.8% YoY but down 2.8% QoQ, with mixed performance among major companies—NVIDIA rose 12% while STMicroelectronics and Kioxia saw >20% declines;

➁ Economic uncertainty and tariffs are dampening outlooks, though AI demand boosts NVIDIA and memory firms (e.g., SK Hynix expects 14.6% Q2 growth), while automotive and smartphone markets slow;

➂ Semiconductor Intelligence forecasts 7% 2025 growth, citing tariff risks potentially lowering TechInsights' projection from 14% to 2%, with weakness likely extending to 2026.

➀ NVIDIA unveiled its new MGX PCIe Switch Board with ConnectX-8, designed for 8x PCIe GPU servers, marking a significant architectural shift in AI server infrastructure;

➁ The board integrates four PCIe Gen6 ConnectX-8 NICs, enabling 800Gbps networking and direct PCIe Gen6 x16 connectivity between GPU pairs, bypassing CPU limitations;

➂ This innovation allows NVIDIA to deploy PCIe Gen6 and high-speed networking capabilities independently of next-gen CPU platforms, accelerating AI factory scalability.

➀ NVIDIA develops 800V HVDC architecture to support 1 MW server racks by 2027, addressing existing 54V DC system limitations;

➀ Current 54V DC systems struggle beyond 200kW, requiring impractical copper quantities for gigawatt data centers;

➂ New 800V HVDC integrates with 13.8kV AC sources, reduces copper use by 45%, and eliminates inefficient AC-DC conversions.

➀ NVIDIA's suppliers, including Foxconn and Dell, resolved technical challenges such as liquid cooling leaks and overheating, enabling accelerated production of the GB200 AI server racks;

➁ Delays in GB200 shipments were due to software bugs and inter-chip connectivity issues, with suppliers collaborating closely with NVIDIA for months;

➂ NVIDIA abandoned the Cordelia design for GB300, reverting to the older Bianca layout, which limits GPU replacement flexibility but expedites production.

➀ U.S. semiconductor sanctions aimed at curbing China's growth may inadvertently boost its domestic industry, fostering a resilient ecosystem;

➀ Taiwanese IC substrate firms face divergent outcomes: Unimicron struggles with sanctions, while Zhen Ding thrives via a 'China for China' strategy;

➂ Nvidia CEO Jensen Huang criticizes U.S. export bans as ineffective, noting Chinese competitors gain ground due to restricted access to advanced chips.

➀ NVIDIA collaborates with Texas Instruments, Infineon, STMicroelectronics, and others to advance 800Vdc power architecture for AI data centers;

➁ 800Vdc distribution aims to improve power density by 85% and efficiency via dc-dc conversion, supporting future 1MW AI server racks;

➂ The project involves power component leaders like Delta, Schneider Electric, and Navitas to optimize energy infrastructure.

➀ The article compares the performance and cost efficiency of AMD and NVIDIA GPUs for various AI tasks such as chat, translation, reasoning, and summarization.

➁ It highlights the MI325X and MI300X as cost-effective options for Llama3 70B chat and translation tasks.

➂ The analysis reveals that AMD GPUs are less cost-effective in rental scenarios due to limited availability and higher prices.

➃ The article discusses the need for better inference benchmarks and explores the features and capabilities of NVIDIA's Dynamo framework.

➀ NVIDIA plans to launch the Blackwell-based RTX PRO 6000D (B40) in China as a cheaper alternative to the banned H20 GPU, priced between $6,500–$8,000;

➁ The B40 uses GDDR7 memory, avoids TSMC’s CoWoS packaging, and lacks NVLink support, limiting multi-GPU scalability;

➂ Deliveries could begin by Q3 2025, relying on PCIe 6.0 switches and Spectrum-X networking for server integration.

➀ Intel Xeon 6 processors with High-Priority Cores (P-cores) are positioned as a key feature in NVIDIA DGX B300 AI servers, using the 64-core Xeon 6776P CPU for enhanced workload distribution;

➁ Intel markets its Xeon 6700P's 2DPC memory speed advantage (5,200 MT/s) over AMD EPYC 9005 (4,000 MT/s), though AMD offers more memory channels and capacity;

➂ NVIDIA's preference for Intel CPUs in GPU servers stems from reduced competitive overlap in AI accelerators, benefiting from Intel's canceled Rialto Bridge and Falcon Shores projects.

➀ Intel launched three new Xeon 6 P-Core CPUs (6732P, 6774P, 6776P) optimized for GPU-powered AI workloads, featuring dynamic core prioritization and up to 8TB memory support;

➁ The processors debut in Nvidia's DGX B300 AI systems with 30% faster memory speeds vs. AMD EPYC and 2.3x higher bandwidth over prior-gen Xeon;

➂ The 350W TDP chips include Intel AMX with FP16 support for AI tasks, alongside a lower-power 235W 40-core B-variant model.

➀ Nvidia announced the Grace Blackwell Ultra AI chip (GB300) shipping in Q3 and a 4x performance boost in its RTX Pro Server for AI models;

➁ Nvidia and Infineon collaborated on centralized 800V HVDC power architecture for AI data centers to optimize space and efficiency;

➂ Wolfspeed faces bankruptcy due to $6.5B debt and loss of key clients like Tesla and Volkswagen, reducing its SiC market share from 32% to 19%.

➀ Nvidia and Infineon collaborate to develop a centralized 800V HVDC power architecture for AI data centers, aiming to replace decentralized AC/DC systems and reduce energy losses;

➀ The new approach minimizes power conversion stages, improves efficiency, and supports higher power densities (projected over 1MW per rack by 2030), critical for large-scale AI workloads involving 100,000+ GPUs;

➂ Centralized HVDC enhances scalability, reliability, and sustainability while meeting the performance demands of next-gen AI infrastructure.

➀ AMD announces Radeon RX 9060 XT with 8GB ($299) and 16GB ($349) variants, targeting competition against NVIDIA's RTX 5060 Ti.

➁ Featuring Navi 44 silicon on TSMC’s 4nm process, the GPU upgrades to PCIe 5.0 x16, offering 14% higher clock speeds and 11% more memory bandwidth than its predecessor.

➂ AMD highlights 6% performance lead over NVIDIA’s RTX 5060 Ti 8GB at 1440p, but 8GB VRAM model faces criticism for potential obsolescence.

➀ AMD's Computex 2025 keynote focuses on client and workstation products, led by Senior VP Jack Huynh;

➁ Anticipated announcements include Zen 5-based Threadripper CPUs, mainstream RDNA 4 Radeon GPUs (response to NVIDIA's RTX 5060), and promotion of Ryzen AI MAX laptops;

➂ Emphasis on gaming, workstation hardware, and competition in the mainstream GPU market.

➀ China aims to capture 40% of its domestic AI chip market in 2024, led by local players like Huawei's Ascend series, Cambricon, Alibaba's Hanguang 800, and Baidu's Kunlun chips;

➁ Global competitors such as Intel, AMD, Google (Trillium TPU), AWS (Trainium), and Meta (MTIA) are advancing their own AI processors, intensifying pressure on Nvidia;

➂ Imec's proposal for reconfigurable AI processors could disrupt current development paradigms, potentially challenging the strategies of major AI chip designers.

➀ Razer unveils the thinnest Blade 14 laptop at Computex 2025, featuring Nvidia RTX 50 series GPUs (5060/5070) and AMD Ryzen AI 9 365 processor.

➀ The laptop boasts a redesigned thermal system with a 'Thermal Hood' and vapor chamber, a 14-inch 3K OLED 120Hz display, and up to 11-hour battery life.

➂ Priced from $2,299, it aims to challenge Asus TUF Gaming A14 as the top 14-inch gaming laptop.

➀ Intel unveiled its Arc Pro B50 and B60 GPUs at Computex 2025, targeting the workstation and AI market with 16GB (B50) and 24GB (B60) memory capacities, emphasizing cost efficiency and AI inference capabilities;

➁ The B50 features a 70W bus-powered design, while the B60 supports up to 200W, with a focus on Linux-based virtualization (SR-IOV) and scalable configurations like dual-GPU systems for AI workloads;

➂ Intel also showcased its Panther Lake 18A processors, aiming to challenge TSMC's process leadership, with GPUs expected to launch in Q3 2025 to address the underserved low-end AI inference and server market.