Latest news with #NVLink
AllAfrica
5 days ago
- Business
- AllAfrica
Chinese worry Nvidia H20 chips are poisoned wine for AI industry
The relaxation of the United States' export controls for Nvidia's H20 chips won brief applause in China – while raising longer-term concerns about whether Chinese firms will over-rely on foreign artificial intelligence chips. During a trip to China on July 15, Nvidia Chief Executive Jensen Huang said in a press conference that the company will resume H20 chip sales to China, now that the US government has indicated it will soon grant export licenses. 'I hope to get more advanced chips into China than the H20,' Huang said. 'Technology is always moving on. Today, Hopper's terrific, but some years from now, we will have more and more and better and better technology, and I think it's sensible that whatever we're allowed to sell in China will continue to get better and better over time as well,' he said. He also took the opportunity to praise Huawei's achievements in making AI chips. However, some Chinese commentators viewed the development as unfavorable for China's chip-making sector. 'This is not a simple lifting of the export restrictions, but a carefully designed measure for the United States to maintain its technological blockade against China,' a Guangdong-based columnist says in an article. 'The H20's FP16 computing power is only 15% of H100, while its NVLink bandwidth is reduced from 900GB/s to 400GB/s. The chip's transformer engine (TE) is completely deleted,' he says. 'Such a design ensures the chip's AI inference ability and reduces its AI training capability, perfectly implementing the United States' strategy of blocking high-end chips, but not mid-end ones, to China.' 'By limiting the key performance of H20, the US can maintain its blockade of high-end computing power while handing Chinese companies a glass of 'poisoned wine,'' he says. In Chinese idiom, a person who 'drinks poisonous liquor to quench thirst' knows that it will kill him, in the long run, but he can't do anything to change the situation. Applied to Chinese technology companies, this means that the domestic chipmakers can benefit from foreign AI chips in the short term but will miss an opportunity to grow and establish an ecosystem. An AI firm needs tens of thousands of Nvidia's high-end chips, such as A100 or H100, to train a large language model (LLM) like ChatGPT. Once an LLM is developed, the company can use slower graphics processing units for inference tasks. In an article published by a columnist using the pseudonym 'Silicon Rabbit' says that Nvidia's Huang made a subtle and cunning move to help the US curb China's chip sector. 'Imagine that someone sold you a Ferrari with a powerful V12 engine, but downgraded its gas pipe, gearbox and wheels. This car can run normally on straight roads, but it faces limitations when continuously speeding up or making sharp turns,' he says, attributing the metaphor to an unnamed senior software engineer who had participated in the Hopper architecture's performance optimization project. He says that the computing power of a single H20 chip is far below that of the H100, while a reduced interconnect bandwidth means a significant reduction in AI training capability. 'AI training is similar to having tens of thousands of people work together, which requires fast information exchanges,' he says. 'A low interconnect bandwidth means that people communicate slowly, resulting in a low thinking efficiency.' He says that the H20 chip cannot be used to train trillion-parameter LLMs. 'The H20 generously offers 96 gigabytes of the third-generation high bandwidth memory (HBM3) – higher than the H100's 80GB HBM3e. However, the H20's memory bandwidth is only 4.0 terabytes per second (TB/s), lower than the H100's 4.8 TB/s,' he says. 'It is like someone giving you a bigger table to read more books, but making it harder for you to take books from the shelves.' The writer says the relaxation of the export rules for the H20 is aimed at permitting the US to control the pace of China's AI development precisely. Reuters, citing sources familiar with the situation, reported that Chinese internet giants, including ByteDance and Tencent, are submitting applications for the H20 chip. ByteDance denied the report. Tencent did not respond to Reuters' request for comment. On April 9, 2025, according to Nvidia, the US government informed the company that a license is required for exporting its H20 products into the Chinese market. The new curb was a part of Washington's countermeasures after China retaliated against the Trump administration's reciprocal tariffs. In announcing results for the three months ended April 27, Nvidia said sales of H20 products were US$4.6 billion before the new export licensing requirements took effect. It said it could not ship an additional $2.5 billion of H20 revenue. After US and Chinese officials held meetings in London on June 9, both sides agreed to de-escalate the trade war. Beijing decided to ease export controls on niche metals to the US. In return, the US would allow Chinese firms to use its chip-making software and export parts for China's C919 flight engines. And now, the US will enable Nvidia to ship the H20 chips to China. A Guangdong-based columnist says Nvidia's H20 chips will enjoy an advantage in the Chinese market, although Huawei's Ascend 910B chips perform better in many aspects of AI training. He says Nvidia's CUDA platform is more advanced than Huawei's MindSpore framework, making customers reluctant to use non-Nvidia chips. For example, he says that Alibaba prefers to use the H20 chips to migrate its existing AI system, while the Ascend 910B chips may target state-owned enterprises. A Beijing-based writer expects Nvidia's CUDA platform to continue enjoying an 80% market share in China, as it would be expensive for companies to switch to new platforms. On July 18, a spokesperson for the Chinese Ministry of Commerce said the US should abandon its 'zero-sum mentality' and further remove a series of trade restrictions targeting Chinese companies that the ministry considered unreasonable. The spokesperson stated that in May, the US unveiled export control measures targeting Huawei's Ascend chips, tightened restrictions on Chinese chip products following unfounded accusations, and intervened in fair market competition with administrative measures. The spokesperson urged the US to work with China to correct erroneous practices through equal consultation. Meanwhile, the Trump administration increased its efforts to prevent China from obtaining Nvidia's high-end chips. It urged Malaysia and Thailand to curb transhipments of Nvidia's AI chips to China. On July 14, the Malaysian government announced that export, transshipment or transit of high-performance AI chips of US origin will require a trade permit. Companies must notify the government at least 30 days before shipping Nvidia's high-end chips elsewhere. Read: US plans to tighten AI chip export rules for Malaysia, Thailand
Business Wire
5 days ago
- Business
- Business Wire
Panmnesia Introduces Today's and Tomorrow's AI Infrastructure, Including a Supercluster Architecture That Integrates NVLink, UALink, and HBM via CXL
DAEJEON, South Korea--(BUSINESS WIRE)--Panmnesia has released a technical report titled 'Compute Can't Handle the Truth: Why Communication Tax Prioritizes Memory and Interconnects in Modern AI Infrastructure.' In this report, Panmnesia outlines the trends in modern AI models, the limitations of current AI infrastructure in handling them, and how emerging memory and interconnect technologies—including Compute Express Link (CXL), NVLink, Ultra Accelerator Link (UALink), and High Bandwidth Memory (HBM)—can be leveraged to improve AI infrastructure. Panmnesia aims to address the current challenges in AI infrastructure, by building flexible, scalable, and communication-efficient architecture using diverse interconnect technologies, instead of fixed GPU-based configurations. Panmnesia's CEO, Dr. Myoungsoo Jung, explained, 'This technical report was written to more clearly and accessibly share the ideas on AI infrastructure that we presented during a keynote last August. We aimed to explain AI and large language models (LLMs) in a way that even readers without deep technical backgrounds could understand. We also explored how AI infrastructure may evolve in the future, considering the unique characteristics of AI services.' He added, 'We hope this report proves helpful to those interested in the field.' Overview of the Technical Report Panmnesia's technical report is divided into three main parts: Trends in AI and Modern Data Center Architectures for AI Workloads CXL Composable Architectures: Improving Data Center Architecture using CXL and Acceleration Case Studies Beyond CXL: Optimizing AI Resource Connectivity in Data Center via Hybrid Link Architectures (CXL-over-XLink Supercluster) 1. Trends in AI and Modern Data Center Architectures for AI Workloads1 AI applications based on sequence models—such as chatbots, image generation, and video processing—are now widely integrated into everyday life. This technical report begins with an overview of sequence models, their underlying mechanisms, and the evolution from recurrent neural networks (RNNs) to large language models (LLMs). It then explains how current AI infrastructures handle these models and discusses their limitations. In particular, Panmnesia identifies two major challenges in modern AI infrastructures: (1) communication overhead during synchronization and (2) low resource utilization resulting from rigid, GPU-centric architectures. 2. CXL Composable Architectures: Improving Data Center Architecture Using CXL and Acceleration Case Studies2 To address the aforementioned challenges, Panmnesia proposes a solution built on CXL, an emerging interconnect technology. The report offers a thorough explanation of CXL's core concepts and features, emphasizing how it can minimize unnecessary communication through automatic cache coherence management and enables flexible resource expansion—ultimately addressing key challenges of conventional AI infrastructure. Panmnesia also introduces its CXL 3.0-compliant real-system prototype developed using its core technologies, including CXL IPs and CXL Switches. The report then shows how this prototype has been applied to accelerate real-world AI applications—such as RAG and deep learning recommendation models (DLRM)—demonstrating the practicality and effectiveness of CXL-based infrastructure. 3. Beyond CXL: Optimizing AI Resource Connectivity in Data Center via Hybrid Link Architectures (CXL-over-XLink Supercluster)3 This technical report is not limited to CXL alone. Panmnesia goes further by proposing methods to build more advanced AI infrastructure through the integration of diverse interconnect technologies alongside CXL. At the core of this approach is the CXL-over-XLink supercluster architecture, which uses CXL to enhance scalability, compatibility, and communication efficiency across clusters connected via accelerator-centric interconnects—collectively referred to as XLink—including UALink, NVLink, and NVLink Fusion. The report explains how the integration of these interconnect technologies enables an architecture that combines the advantages of each. It then concludes with a discussion on the practical application of emerging technologies such as HBM and silicon photonics. Conclusion With the release of this technical report, Panmnesia reinforces its leadership in next-generation interconnect technologies such as CXL and UALink. In parallel, the company continues to actively participate in various consortia related to AI infrastructure, including the CXL Consortium, UALink Consortium, PCI-SIG, and the Open Compute Project. Recently, Panmnesia also unveiled its 'link solution' product lineup, designed to realize its vision for next-generation AI infrastructure and further strengthen its brand identity. Dr. Myoungsoo Jung, CEO of Panmnesia, stated, 'We will continue to lead efforts to build better AI infrastructure by developing diverse link solutions and sharing our insights openly.' The full technical report on AI infrastructure is available on Panmnesia's website: 1 This corresponds to Sections 2 and 3 of the technical report. 2 This corresponds to Sections 4 and 5 of the technical report. 3 This corresponds to Section 6 of the technical report. Expand
Business Wire
5 days ago
- Business
- Business Wire
Panmnesia Introduces Today's and Tomorrow's AI Infrastructure,
BUSINESS WIRE)--Panmnesia has released a technical report titled 'Compute Can't Handle the Truth: Why Communication Tax Prioritizes Memory and Interconnects in Modern AI Infrastructure.' In this report, Panmnesia outlines the trends in modern AI models, the limitations of current AI infrastructure in handling them, and how emerging memory and interconnect technologies—including Compute Express Link (CXL), NVLink, Ultra Accelerator Link (UALink), and High Bandwidth Memory (HBM)—can be leveraged to improve AI infrastructure. Panmnesia aims to address the current challenges in AI infrastructure, by building flexible, scalable, and communication-efficient architecture using diverse interconnect technologies, instead of fixed GPU-based configurations. Panmnesia's CEO, Dr. Myoungsoo Jung, explained, 'This technical report was written to more clearly and accessibly share the ideas on AI infrastructure that we presented during a keynote last August. We aimed to explain AI and large language models (LLMs) in a way that even readers without deep technical backgrounds could understand. We also explored how AI infrastructure may evolve in the future, considering the unique characteristics of AI services.' He added, 'We hope this report proves helpful to those interested in the field.' Overview of the Technical Report Panmnesia's technical report is divided into three main parts: Trends in AI and Modern Data Center Architectures for AI Workloads CXL Composable Architectures: Improving Data Center Architecture using CXL and Acceleration Case Studies Beyond CXL: Optimizing AI Resource Connectivity in Data Center via Hybrid Link Architectures (CXL-over-XLink Supercluster) 1. Trends in AI and Modern Data Center Architectures for AI Workloads 1 AI applications based on sequence models—such as chatbots, image generation, and video processing—are now widely integrated into everyday life. This technical report begins with an overview of sequence models, their underlying mechanisms, and the evolution from recurrent neural networks (RNNs) to large language models (LLMs). It then explains how current AI infrastructures handle these models and discusses their limitations. In particular, Panmnesia identifies two major challenges in modern AI infrastructures: (1) communication overhead during synchronization and (2) low resource utilization resulting from rigid, GPU-centric architectures. 2. CXL Composable Architectures: Improving Data Center Architecture Using CXL and Acceleration Case Studies 2 To address the aforementioned challenges, Panmnesia proposes a solution built on CXL, an emerging interconnect technology. The report offers a thorough explanation of CXL's core concepts and features, emphasizing how it can minimize unnecessary communication through automatic cache coherence management and enables flexible resource expansion—ultimately addressing key challenges of conventional AI infrastructure. Panmnesia also introduces its CXL 3.0-compliant real-system prototype developed using its core technologies, including CXL IPs and CXL Switches. The report then shows how this prototype has been applied to accelerate real-world AI applications—such as RAG and deep learning recommendation models (DLRM)—demonstrating the practicality and effectiveness of CXL-based infrastructure. 3. Beyond CXL: Optimizing AI Resource Connectivity in Data Center via Hybrid Link Architectures (CXL-over-XLink Supercluster) 3 This technical report is not limited to CXL alone. Panmnesia goes further by proposing methods to build more advanced AI infrastructure through the integration of diverse interconnect technologies alongside CXL. At the core of this approach is the CXL-over-XLink supercluster architecture, which uses CXL to enhance scalability, compatibility, and communication efficiency across clusters connected via accelerator-centric interconnects—collectively referred to as XLink—including UALink, NVLink, and NVLink Fusion. The report explains how the integration of these interconnect technologies enables an architecture that combines the advantages of each. It then concludes with a discussion on the practical application of emerging technologies such as HBM and silicon photonics. Conclusion With the release of this technical report, Panmnesia reinforces its leadership in next-generation interconnect technologies such as CXL and UALink. In parallel, the company continues to actively participate in various consortia related to AI infrastructure, including the CXL Consortium, UALink Consortium, PCI-SIG, and the Open Compute Project. Recently, Panmnesia also unveiled its 'link solution' product lineup, designed to realize its vision for next-generation AI infrastructure and further strengthen its brand identity. Dr. Myoungsoo Jung, CEO of Panmnesia, stated, 'We will continue to lead efforts to build better AI infrastructure by developing diverse link solutions and sharing our insights openly.' The full technical report on AI infrastructure is available on Panmnesia's website:
Yahoo
15-07-2025
- Business
- Yahoo
Can NVIDIA's End-to-End Stack Keep Driving Networking Revenues?
NVIDIA Corporation's NVDA networking business is gaining strong momentum, thanks to its full-stack approach. In the first quarter of fiscal 2026, networking revenues surged 64% sequentially to approximately $5 billion. The robust growth was driven by the rise in AI factory buildouts and the growing demand from hyperscalers and enterprises building large AI clusters. NVIDIA's networking business focuses on providing high-performance connectivity solutions, including chips, switches, interconnects and software for data centers, which help customers scale their AI workloads more efficiently. The company's NVLink interconnect offers 14 times the bandwidth of PCIe Gen 5 and supports massive data throughput in a single rack. New offerings like NVLink Fusion are attracting custom chipmakers and CPU vendors, adding more partners to the NVIDIA ecosystem. In the last reported quarter, NVLink shipments exceeded $1 billion, showing strong early adoption. NVIDIA's Ethernet networking platform, Spectrum-X, is also expanding quickly. It is now on pace to generate more than $8 billion in annualized revenues. Companies like Microsoft, Meta, Oracle and Google Cloud are already deploying it to handle AI network traffic with low latency and high utilization. As AI workloads grow more complex and interconnected, NVIDIA's end-to-end stack makes it easier and cheaper to scale. If this demand trend holds, networking could become a larger and more stable revenue driver over time. Per our model, the company's revenues from the networking business are estimated to grow 57.7% year over year to $20.5 billion in fiscal 2026. Broadcom AVGO and Marvell Technology MRVL are two major companies that are continuously expanding in the AI networking space. Broadcom offers custom networking chips and high-speed interconnects to major cloud providers. Its strong position in Ethernet switches and application-specific integrated circuits (ASICs) gives it a foothold in the AI infrastructure space. However, Broadcom doesn't offer a complete end-to-end solution like NVIDIA. The company relies more on partnerships with cloud customers rather than offering a unified platform. Marvell Technology is also focusing on data center connectivity and recently launched products tailored for AI and cloud workloads. The company works with major hyperscalers and has a growing presence in optical and Ethernet switching. But like Broadcom, Marvell lacks NVIDIA's software and system-level integration, which limits its ability to capture the full AI networking stack. Shares of NVIDIA have risen around 22.2% year to date against the Zacks Computer and Technology sector's gain of 7.4%. Image Source: Zacks Investment Research From a valuation standpoint, NVDA trades at a forward price-to-earnings ratio of 33.81, higher than the sector's average of 27.39. Image Source: Zacks Investment Research The Zacks Consensus Estimate for NVIDIA's fiscal 2026 and 2027 earnings implies a year-over-year increase of approximately 41.8% and 31.9%, respectively. Estimates for fiscal 2026 have been revised downward over the past 30 days, while those for fiscal 2027 have been revised upward in the past 30 days. Image Source: Zacks Investment Research NVIDIA currently carries a Zacks Rank #3 (Hold). You can see the complete list of today's Zacks #1 Rank (Strong Buy) stocks here. Want the latest recommendations from Zacks Investment Research? Today, you can download 7 Best Stocks for the Next 30 Days. Click to get this free report NVIDIA Corporation (NVDA) : Free Stock Analysis Report Marvell Technology, Inc. (MRVL) : Free Stock Analysis Report Broadcom Inc. (AVGO) : Free Stock Analysis Report This article originally published on Zacks Investment Research ( Zacks Investment Research Sign in to access your portfolio
Business Insider
11-07-2025
- Business
- Business Insider
‘Stop Crying Bubble,' Says Investor About Nvidia Stock
Nvidia (NASDAQ:NVDA) and records are no strangers, and the chip giant notched another on Wednesday, when the stock crossed the $4 trillion market cap threshold, becoming the first company ever to hit that milestone. Don't Miss TipRanks' Half-Year Sale Take advantage of TipRanks Premium at 50% off! Unlock powerful investing tools, advanced data, and expert analyst insights to help you invest with confidence. Make smarter investment decisions with TipRanks' Smart Investor Picks, delivered to your inbox every week. That saw Nvidia pull further away from closest peers Microsoft and Apple, but if you think the stock might be looking pricey at such an elevated level, then think again. That at least is the opinion of one investor, known by the pseudonym the Agar Capital (AC), who argues that Nvidia's valuation is 'justified by its dominant AI infrastructure, explosive earnings growth, and unmatched margins, not by speculative hype or bubble dynamics.' Indeed, the investor pushes back against the chorus of skeptics who tend to sound the bubble alarm each time Nvidia sets a new high. In AC's view, Nvidia isn't just a 'buzzword' – it's the engine room of a technological revolution, consistently delivering profits and cash flow on a scale that rivals can't match. Far from being 'overvalued,' its current valuation is a reflection of the outsized influence it wields over the future of artificial intelligence. Skeptics, of course, have been voicing the same concerns since Nvidia was worth $500 billion, then $1 trillion, and later $2 trillion. Now, at $4 trillion, those doubts persist. Yet, as AC points out, 'The price isn't the problem. It's the market's incredulity.' What's more, focusing solely on Nvidia's share price overlooks the broader narrative. As AC emphasizes, this is arguably the most 'powerful margin machine' on the planet, with the company raking in $44 billion in revenue and $26 billion in cash in just the first quarter of FY26. About 90% of that revenue now comes from data centers, a segment where margins are still expanding, driven by the rapid adoption of new Blackwell chips and rising demand for Nvidia's proprietary networking technologies, such as NVLink and Spectrum-X. While it's true the stock isn't cheap, labeling it as expensive is also misleading. Nvidia's P/E ratio continues to fall as EPS surges, and the PEG ratio hovers around 1, signaling that growth is keeping pace with valuation. Free cash flow yield remains 'positive and solid.' And compared to peers like AMD, Intel, and Broadcom, Nvidia clearly stands out for its superior profitability, high-quality earnings, and robust returns on capital. Over the years, it has become the backbone of the AI-driven economy, and as AC concludes, 'This is not a bubble. It is a dominant position monetized to the highest degree.' With all of that in mind, it's hardly surprising that Agar Capital rates NVDA shares a Buy. (To watch Agar Capital's track record, click here) Wall Street broadly agrees; based on a mix of 37 Buys, 4 Holds and a single Sell, Nvidia boasts a Strong Buy consensus rating. (See NVDA stock forecast) To find good ideas for stocks trading at attractive valuations, visit TipRanks' Best Stocks to Buy, a tool that unites all of TipRanks' equity insights.
