Tag: Chips

  • Huawei 950PR – New Nvidia rival conquers China’s AI market

    Huawei 950PR – New Nvidia rival conquers China’s AI market

    For years, Nvidia’s dominance of China’s artificial intelligence market seemed inviolable, protected not only by hardware performance, but more importantly by its CUDA software ecosystem. However, the latest reports from Shenzhen suggest that Huawei has finally found the key to the closed doors of the biggest private tech giants. The new Ascend 950PR chip has passed tests with players such as ByteDance and Alibaba, heralding a new chapter in the competition for the heart of China’s AI sector.

    The success of the 950PR is not solely due to raw computing power, which has increased only marginally over its predecessor (the 910C). The breakthrough came in the realm of usability. Huawei, hitherto rigorously pushing its own CANN software, has opted for greater compatibility with the Nvidia ecosystem. For engineers in Beijing or Hangzhou, this means a drastic reduction in model migration costs and an easier transition to a domestic architecture without having to write code from scratch.

    Strategic chip positioning hits a turning point in the Chinese market. While the world continues to be fascinated by training ever-larger models, the Chinese technology sector is shifting its focus towards inference – that is, the practical implementation and running of ready-made algorithms. This is where 950PR is expected to offer the highest operating culture and responsiveness. The rise in popularity of open-source solutions, such as OpenClaw, is further driving demand for units optimised for specific business tasks, rather than just theoretical performance records.

    The economic aspect also speaks in Huawei‘s favour. With a starting price hovering around $6,900 per card, the 950PR becomes an attractive alternative to US products whose availability is regularly restricted by Washington’s sanctions. Even if Nvidia is given the green light to supply chips such as the H200, regulatory uncertainty is prompting Chinese companies to diversify their suppliers.

    With mass production planned for next month and an ambitious target of shipping 750,000 units later this year, it shows that Huawei is coming out of its defensive phase. If the 950PR maintains the positive feedback from the tests, the Chinese giant may cease to be merely the beneficiary of patriotic appeals from the government and become a viable commercial choice for companies that are primarily looking for cost efficiency and operational stability in technology.

  • AI chips for investment in the US. Trump’s administration is changing the rules

    AI chips for investment in the US. Trump’s administration is changing the rules

    Washington is preparing to fundamentally redefine the rules of the global artificial intelligence arms race. According to documents, Donald Trump’ s administration is considering a new restrictive export framework for the most advanced AI systems. Instead of the previous policy based primarily on national security, the new model shifts the centre of gravity towards tough business negotiations and a requirement for direct investment within the US.

    The proposed legislation stipulates that access to large batches of processors – in excess of 200,000 units – will be conditional not only on security guarantees, but primarily on capital commitment to US infrastructure. This is a radical departure from the previous administration’s approach, which rewarded close allies with broad exemptions from export controls. The current vision is more transactional: you want to build AI power at home, you must first help build it in America.

    The control mechanism is to be multi-level and extremely precise. The documents suggest that even relatively small installations of less than a thousand chips may require special licences. Crucially, the burden of oversight is to fall directly on manufacturers such as Nvidia and AMD. These companies would be required to monitor how the hardware is used, and recipients would have to implement software to prevent processors from combining into massive computing clusters without Washington’s explicit approval.

    The model for the new regulations is to be the agreements recently concluded with Saudi Arabia and the United Arab Emirates. The Department of Commerce, cutting out the – what it called – ‘onerous’ rules of its predecessors, is betting on a model in which technology transfer is a reward for economic loyalty. For global technology players, this means they need to review their data centre location strategies.

    While the new rules do not directly hit fully embargoed countries like Russia, they cast a shadow over relations with existing partners in Europe and Asia. Washington is gaining powerful leverage: AI chips are becoming the most important currency in economic diplomacy. Companies hoping to maintain their growth momentum must now factor in the cost of investing in the US, which will become a kind of ‘ticket of entry’ to the world’s most advanced silicon architecture.

  • Physics versus marketing. What do you really gain by investing in 1.8nm and 3nm processors?

    Physics versus marketing. What do you really gain by investing in 1.8nm and 3nm processors?

    Intel is bringing out the heavy guns in the form of third-generation Core Ultra processors, known as Panther Lake, which are based on 18A, or 1.8 nanometre, technology. On the other side of the market barricade is AMD with its Ryzen chips, baked in TSMC ‘s Taiwanese factories using a 3nm process. On paper, Intel’s advantage seems crushing, suggesting a technology almost half the size and more modern. However, in the CFO’s portfolio, this difference may prove to be a statistical error. In a world where ‘nanometre’ has become a brand rather than a measurement, business must learn to look at what really drives performance, ignoring the labels on the boxes.

    When IT managers look at the specifications of new laptops or servers, their gaze naturally goes to the numbers, because in the technology industry, smaller usually means better, faster and more economical. Manufacturers are well aware of this, which is why the arms race in the semiconductor sector has moved from the physics labs to the marketing departments. To make an informed purchasing decision for 2025-2026, you need to understand where the engineering ends and the wordplay begins.

    The grand illusion of the nanometre

    For decades, the IT industry has operated with a simple and understandable currency. Back in 1995, when we talked about the 350 nm technology process, it meant that the gate of a transistor on a silicon wafer was actually 350 nanometres long. The engineer and the salesman spoke the same language, and the node name was a direct reflection of physical reality. However, this order broke down in the late 1990s with the introduction of new technologies for building microtransistors, which broke the direct link between the node name and the physical dimension of the components.

    Today, names such as ‘Intel 4′, ’18A’ meaning 18 Angstroms, or ‘TSMC N3’ are predominantly trade names. Treating them as a technical measure of length is a mistake that can lead to misleading business conclusions. It is a situation analogous to the automotive market, where the model designation of a car, for example the BMW 330, no longer necessarily denotes a three-litre engine. The number now serves to position the product in the range, rather than to describe its technical parameters precisely.

    For business, this means that the approach to analysing offerings needs to change. The fact that one processor is labelled ‘1.8 nm’ and another ‘3 nm’ does not automatically mean that the former is physically much smaller. In fact, the differences may be minimal and, in extreme cases, the packing density relationship may even be the opposite of what the numbers suggest.

    The hard currency of silicon

    Since nanometres are conventional, an informed investor or IT manager should look at other metrics. If we look under the hood of Panther Lake processors or the latest Ryzen processors, we find objective parameters that PR departments are reluctant to talk about, but which are crucial for engineers. These are, first and foremost, Gate Pitch, which is the minimum distance between individual transistors, and Metal Pitch, denoting the minimum distance between the copper paths connecting these components.

    Analysis of this hard data leads to surprising conclusions. Comparing the current generation of processes, it appears that the Intel 4 technology and the competing TSMC N4 have almost identical physical characteristics, with a gate pitch oscillating between 50 and 51 nanometres. Despite the different trade names, the packing density of the technologies is very similar. The future looks even more interesting, with Intel promoting an 18A process suggesting 1.8 nm, while TSMC is preparing to implement a 2 nm process. Paradoxically, according to many technical analyses, it is the Taiwanese ‘2 nm’ that may offer higher transistor density than the US solution. Intel is making up for it with marketing, suggesting leadership, but in practice the two giants are going head to head and their nodes will meet each other halfway in terms of real-world performance.

    Physics translates into costs

    Although the labels are confusing, the technological advances are real and central to the cost of doing business, or TCO. Regardless of the nomenclature, the drive towards denser transistor packing is driven by the inexorable laws of physics, as a smaller transistor with a shorter path between source and drain requires a lower voltage to switch its logic state. For the company, this translates directly into energy efficiency and thermal performance.

    The chip, made using a newer, denser process, uses less power for the same load. On the scale of a single laptop, this means an extra hour of battery life during a business trip, while on the scale of a data centre, it translates into thousands of zlotys of savings on electricity bills. The thermal aspect is equally important, as less power consumption means less heat generated. This allows the processors to run at higher frequencies without the risk of thermal throttling, ensuring more stable operation of demanding applications. Therefore, Intel Panther Lake will be inherently better than its predecessor not because of the name ’18A’, but because the engineers have actually improved the physical structure of the chip, which is also true for AMD using TSMC improvements.

    The strategic trap of the single supplier

    There is another element of business risk in this technological jigsaw puzzle, related to incompatibility. Intel’s, TSMC’s and Samsung’s manufacturing processes have diverged dramatically, with each giant using different chip production methods, deploying technologies such as FinFET or RibbonFET at different times. This means that chip designers such as AMD and NVIDIA are firmly tied to their chosen factory and cannot move production to a competitor overnight. Adapting a design to another factory is a process that takes up to a year and incurs huge costs. When choosing a hardware platform for a company, decision makers are therefore choosing not just a processor, but the entire supply chain, where the stability of the manufacturing partner becomes a strategic factor, more important than the marketing name of a nanometre.

    We are approaching the point where comparing processors solely on the basis of lithography becomes pointless. Intel Panther Lake and the upcoming Ryzen generations will be powerful chips, but their value to business is not based on the labels on the box. When planning infrastructure purchases, the key indicator should be the performance-per-watt ratio. It is this parameter that determines whether an investment in new hardware will translate into real productivity gains and reduced operating costs for the business.

  • Intel takes another approach to AI. “Not everything for everyone”.

    Intel takes another approach to AI. “Not everything for everyone”.

    Intel is back in the game for the AI market, announcing a new data centre GPU, Crescent Island, at the Open Compute Summit. The chip is due to debut next year and, according to the company’s statement, will be optimised not for impressive benchmarks, but for operating economics: cost per token processing, energy efficiency and AI model inference.

    This is a clear signal of a change in strategy. After years of failed approaches – from the abandoned Gaudi line to the frozen Falcon Shores project – Intel is betting on pragmatism. “We don’t want to build everything for everyone. We are focusing on inference,” stressed the company’s CTO, Sachin Katti. Translated into the language of the market: Intel will not fight Nvidia where Nvidia is strongest, namely in training giant models. Instead, it is targeting the stage where it makes money from enterprise-scale AI implementations.

    Technically, Crescent Island is betting on 160GB of vintage-type memory – slower than the HBM used by competitors. It’s a trade-off: lower peak performance, but potentially better availability and cost per watt. The chip is based on Intel’s consumer GPU architecture, which suggests a shorter deployment cycle and lower manufacturing risk. However, key details are still missing: what technology process (TSMC? in-house factories?) and what real TCO against AMD Instinct or Nvidia Hopper/Blackwell chips.

    In a market that has suffered from a chronic GPU shortage since the release of ChatGPT, room for a third player obviously exists – but patience is running out. Hyperscaler customers want annual launches, interoperability and an open ecosystem. Intel promises exactly that: modularity and the ability to mix chips from different vendors. It’s a defensive/offensive move – if it can’t win solo, it wants to be indispensable as the CPU in every AI system. This is borne out by the recent deal with Nvidia, which invested $5bn and took an approximate 4% stake in Intel.

    Will it be enough? Intel is playing for time and a second chance. AI is no longer a power race – it’s starting to be an economics race. If Crescent Island proves ‘performance per dollar’, Intel can get back to the table. If not – it will remain a factory for the other winners of this revolution.

  • Blackwell successor coming. Nvidia announces Ruby CPX chips

    Blackwell successor coming. Nvidia announces Ruby CPX chips

    Nvidia has announced a new generation of AI chips, signalling its intention to maintain its annual innovation cycle and further strengthen its dominant position in the market.

    The platform, dubbed ‘Rubin’, is intended to replace the recently announced ‘Blackwell’ architecture. The first chip based on it, Rubin CPX, is expected to hit the market by the end of 2026.

    The decision to unveil Blackwell’s successor so soon highlights the intensity of technological and market pressures in the AI sector. As AI models become more complex, their computational requirements also increase.

    The new chips are designed for the most advanced tasks, such as real-time video generation or complex software development with AI assistance.

    The problem Nvidia is trying to solve is the processing bottleneck. The company points out that processing an hour of video through an advanced AI model can require the analysis of up to a million tokens – units of data – which is a huge challenge for traditional GPUs.

    To tackle this task, the Rubin CPX chip is expected to integrate several key processes on a single platform: video decoding and encoding and so-called inference, i.e. model-generated results. This approach aims to dramatically increase performance and energy efficiency.

    The announcement also has a strong business dimension. Against the backdrop of increasing pressure from Wall Street to demonstrate a return on its giant investment in AI infrastructure, Nvidia is presenting concrete calculations.

    The company suggests that an investment of $100 million in new systems could translate into $5 billion in so-called ‘token’ revenue.

    This sends a clear signal to customers and investors that the next generation of equipment is not only to push the boundaries of technological possibilities, but above all to generate tangible financial benefits.

  • A hidden feature in Nvidia chips? China speaks of ‘back doors’ and demands explanations

    A hidden feature in Nvidia chips? China speaks of ‘back doors’ and demands explanations

    China’s cybersecurity regulator (CAC) has summoned Nvidia for talks on the H20 chip, raising concerns about its security. It is a new chapter in the technology rivalry that puts the US giant in a complicated position. Beijing is demanding explanations and documentation from the company about potential security vulnerabilities and so-called backdoors.

    The situation is complex in that the H20 model is not a standard product. It was developed by Nvidia specifically for the Chinese market as a response to restrictive US export sanctions. Washington, concerned about China’s military development, blocked the sale of the most powerful AI chips such as the H100. As a result, the H20 is a deliberately downgraded unit, designed to meet US export standards while satisfying demand from Chinese technology companies.

    The Chinese regulator’s decision can be seen in two ways. On the one hand, it may stem from genuine concerns about data security and user privacy. On the other, it is likely to be part of a broader strategy by Beijing to increase pressure on Western technology companies and promote domestic alternatives as part of a drive towards technological self-sufficiency.

  • Microsoft’s AI chip delay. Braga won’t threaten Nvidia before 2026

    Microsoft’s AI chip delay. Braga won’t threaten Nvidia before 2026

    Microsoft, like Google and Amazon, has been developing its own computing chips for artificial intelligence for several years. However, according to The Information, the Redmond-based conglomerate’s plans have encountered major obstacles. A new AI chip called Braga – which is the successor to the Maia chip unveiled in 2023 – will not go into production before 2026. This is at least a six-month delay from earlier plans.

    The reason? A combination of factors: team turnover, staff shortages and changes to the chip architecture. To make matters worse, once Braga hits the market, its performance is expected to be lower than the latest chips from Nvidia – the current AI market leader.

    This is bad news for Microsoft, which – by investing billions in artificial intelligence infrastructure (including OpenAI) – is trying to reduce its reliance on Nvidia’s expensive GPUs. GPUs, especially the H100 series and its successor Blackwell (B200), are a key component of AI’s generative infrastructure, but also a bottleneck – their availability is limited and their prices high.

    Microsoft made its debut in the AI chip market in late 2023 with the Maia 100 chip, designed to train and run large language models. At the same time, Amazon was developing its Trainium line and Google had already been investing in proprietary TPU processors for years. And while all of these companies are aiming to become independent of third-party suppliers, only Google can today claim to be a fully-fledged alternative to Nvidia – its TPU v5p chip debuted in April and went straight into commercial cloud offerings.

    Amazon is not slowing down either – its Trainium3 chip is still expected to be ready in 2025. Microsoft, meanwhile, could be left without its own GPU alternative for the next several months.

    The delay of the Braga chip is not just a technology problem – it is also a blow to Microsoft’s cloud and AI strategy. The company recently announced a partnership with OpenAI to build dedicated AI data centres, and its own chips were to be the foundation of this infrastructure. In the absence of a finished product, Microsoft will most likely be forced to continue using Nvidia chips – and on terms dictated by the supplier.

    From a market perspective, this means maintaining the status quo – with Nvidia as the dominant hardware provider for generative AI, and a growing gap between Big Tech in their hardware ambitions. Google and Amazon are consolidating as pioneers of their own chips, while Microsoft – despite its aggressive investment – has to resign itself to lagging. In the AI race, each quarter can cost billions.