ascend

The semiconductor research organization SemiAnalysis has released the actual test results of Huawei's Ascend 950PR. The organization selected 1,055 sets of commonly used matrix calculations for large models. Huawei's own underlying program has allowed the 950PR to achieve 86% of its theoretical computing power, higher than NVIDIA's H100 at 75.5%; however, the actual computation speed is still only 43% of the H100, and only 21% compared to the relatively newer B300. This set of results has not yet utilized the newly ported operators from DeepSeek.DeepSeek is working on reducing data movement and waiting from the software side. In attention computation, after adjusting storage with FlashMLA, the two sets of data that originally needed to be read separately can now be retrieved at once; TileLang allows the chip to compute while preloading the next batch of data. SemiAnalysis measured that automatic scheduling has reduced the time for some small computing tasks by up to about 40%. DeepSeek has also ported DeepEP, which is responsible for MoE expert communication, but communication and computation will still compete for memory bandwidth, and Huawei plans to continue improving the firmware.Last year, DeepSeek attempted to train R2 using Huawei's Ascend chips, but even with Huawei engineers on-site for assistance, it could not be completed due to issues related to chip stability, inter-card communication, and software support. This year, both parties continue to collaborate, and by the end of September, DeepSeek focused on publicly releasing six sets of Ascend adaptation code. In October, it was reported that Huawei is adjusting the next generation of chip designs according to DeepSeek's model requirements. The 950PR tested this time is mainly aimed at inference input processing, while DeepSeek is preparing to use the 950DT for large-scale pre-training as early as the end of the year. Previously, both parties had conducted tests at the thousand-card level, but to enable tens of thousands of chips to complete a round of large model training continuously, more complex coordination and stability issues still need to be resolved.

MiniMax announced the open-source release of the MiniMax M3.1 Flash model weights on October 23, with a total of 428 billion parameters and 23 billion active parameters. Compared to the previous generation M3 model, the M3.1 Flash has achieved a comprehensive improvement in model capabilities. According to official data, within 72 hours of the model going live, the number of tokens processed by real users exceeded one trillion.On the technical side, M3.1 Flash improves inference efficiency through optimizations such as full model sparse attention, quantization-aware training (QAT), and power-aware scheduling. In terms of ecosystem, MiniMax collaborates with chip manufacturers such as Huawei Ascend, T-head, Kunlun Core, Moore Threads, Biren, Haiguang, Muxi, Tiandao, NVIDIA, AMD, and cloud platforms like Together AI, Fireworks, Nebius, Baseten, Modal, DigitalOcean to promote access on the first day of open source. vLLM, SGLang, Zhiyuan FlagOS, as well as Hugging Face, Modao, and other communities simultaneously support the deployment of the model across different chips and development environments.

DeepSeek announced the full open-source of its internal inference framework DeepInfer, and simultaneously open-sourced the production-level token encoding and decoding component deepseek-recipe. DeepInfer integrates high-performance operators and a deeply optimized inference engine, supporting distributed inference, speculative decoding, SWA Bounded Replay, and structured output, compatible with NVIDIA and Huawei Ascend platforms.DeepSeek stated that it has implemented high throughput and low latency large-scale inference services for DeepSeek V4.1 Flash based on DeepInfer, and future inference code and optimizations for new models will continue to be updated here. deepseek-recipe is fully aligned with the processing logic of the DeepSeek official API, aiming to ensure that the community receives an experience consistent with the official deployment. The open-source repository and technical report have been released simultaneously.

According to The Information, DeepSeek is collaborating closely with Huawei to jointly design the next generation of AI chips, aiming to build models that do not rely on NVIDIA's computing power. Insiders revealed that Huawei engineers have entered DeepSeek's office to understand its model architecture and algorithm requirements, and to test how the chips handle data transmission and compression, with the first compatible chip targeting Ascend 950DT.The report pointed out that this move comes as China pushes to reduce its dependence on American chips. Limited by TSMC's foundry services and U.S. export controls, Huawei's Ascend 950DT is planned for release in 2027, but mass production may be delayed until 2028. To address the chip shortage, DeepSeek is using the TileLang programming language to bypass CUDA restrictions, while still attempting to acquire the banned NVIDIA Blackwell chips through third-party data centers. In addition, DeepSeek is forming a chip team and has begun testing clusters with over 1,000 processors.

The artificial intelligence company DeepSeek has officially open-sourced infrastructure components for the Huawei Ascend computing platform, covering the TileLang high-level language compilation tool, computing libraries, and distributed communication libraries, corresponding to the previously open-sourced components for the NVIDIA platform. TileLang aims to provide a general-purpose, simpler programming language that can achieve the hardware performance limits, improving development efficiency and simplifying logic compared to CUDA, while its programming model can also leverage chip features.The TileLang route was first validated on the NVIDIA platform and has already supported the implementation of most operators in the training of the DeepSeek V4 series models. The open-sourced Ascend version encapsulates the underlying instructions of the Ascend C, providing a high-level programming approach without sacrificing hardware performance. Currently, every TileLang operator used in DeepSeek's training has a corresponding high-performance implementation on Ascend.The components open-sourced at the same time also include DeepGEMM, DeepEP, TileKernels, FlashMLA, and DeepSelect, which are used for general matrix operations, large-scale cross-device communication, conventional vector calculations and memory access, long-context sparse attention, and data filtering, respectively. DeepSeek claims that in multiple key tests, the related computing and communication performance has approached hardware limits; during the R&D process, the Huawei team provided support, and both parties collaborated to advance the 128-card supernode solution based on Ascend 950, with deep optimizations made for computing and communication.

0G launched Compute Finance (ComFi) on September 21, and introduced the liquid staking product Ascend, while also announcing that the computing power-based digital asset Infinite AI (iAI) will be launched on September 29. Users can stake 0G to receive a0G, use a0G to mint iAI, and then stake eligible iAI to earn computing power points, which can be used for supported 0G AI products such as 0G Private Computer.According to the initial parameters, eligible staked iAI can earn computing power points daily, with a calibrated use value exceeding 1 dollar in supported 0G AI services. When 0G Private Computer goes live, it will connect to over 130 AI models. Computing power points are only used for eligible AI services and do not constitute cash, cash equivalents, or any guaranteed financial returns. iAI is also expected to be launched on 0G's agent and token issuance platform Comfy.fun.

Huawei's rotating chairman Wang Tao stated at the Huawei Connect Conference 2026 that the Ascend 960 chip is ready ahead of schedule, achieving a doubling of performance. Among them, the Ascend 960DT will be released in Q1 2027, three quarters earlier than originally planned; the Ascend 960PR will be released in Q3 2027, one quarter earlier than originally planned.Wang Tao mentioned that they will adhere to an annual evolution pace, with the Ascend 970 planned for release in 2028 and the Ascend 980 planned for release in 2029.

According to a Reuters exclusive report on September 10, Chinese AI chip manufacturers have significantly raised prices due to a shortage of high-bandwidth memory. Huawei has raised the indicated price of the Ascend 950DT accelerator card to over 250,000 yuan, an increase of 20% to 50% compared to two months ago. Cambricon has raised the indicated price of its planned 690 chip by 20% to 30%, and both Muxi and Tian Shu Zhi Xin have also implemented similar price increases.The report states that the reasons include U.S. export controls and the high costs of high-bandwidth memory supplies in the gray market. Prices for earlier Huawei products such as the Ascend 950PR and 910C have also increased by about 30%. These price increases have raised the construction costs of AI computing power in China. Huawei, Cambricon, and other companies did not comment.

According to a report by Bloomberg on September 4, the artificial intelligence company DeepSeek plans to deploy at least 160,000 top accelerators from Huawei Technologies Co. in a large data center being built in Inner Mongolia, potentially forming one of the largest known clusters of Huawei AI chips and advancing China's efforts to replace Nvidia.DeepSeek intends to use Huawei's next-generation Ascend 950DT chips to run its models, with the installation timeline depending on Huawei's production capacity. Sources say that the company currently does not plan to use the 950DT chips for training, although Huawei has designed and marketed these processors for more demanding training functions. DeepSeek had previously attempted to train models using Huawei chips but has so far relied on Nvidia accelerators to complete this critical step.

Research institution Delphi Digital stated that hardware limitations are driving Chinese laboratories to shift towards cheaper models and more efficient technology stacks. Export control restrictions limit their access to advanced chips, while domestic accelerators can handle inference, but cutting-edge training remains difficult to achieve. Chinese laboratories are innovating in system efficiency, model architecture, training data, and reinforcement learning. DeepSeek has found a way to avoid GPU idle data movement, and ByteDance and Moonshot are set to release similar versions within months.In terms of architecture, training is shifting to cheaper digital formats that are compatible with domestic chips. Sparse architectures reduce the runtime per token, and the attention mechanism has been redesigned three times to control the memory costs of long context. In terms of data, denser prediction targets and better optimizers enhance the learning signal per token. In reinforcement learning, DeepSeek's GRPO has become the default recipe, with ByteDance, Alibaba, and MiniMax each launching successor plans within a year. Chinese models are now among the lowest-cost cutting-edge adjacent systems.On OpenRouter, the proportion of token consumption by Chinese developed models is expected to rise from less than 1.2% at the end of 2024 to the majority by 2026. The Chinese AI market is experiencing a price war from 2024 to early 2026, with several laboratories implying revenue multiples of about 47-117 times, while Anthropic and OpenAI are around 15-21 times. Huawei's Ascend chips are currently available for inference, and DeepSeek reportedly attempted to train R2 on Ascend but returned to NVIDIA after encountering technical issues. Domestic accelerator supply remains below estimated demand.

According to official news, the cryptocurrency trading platform AscendEX has announced that it will cease operations, attributing the reasons to the current market environment and the impact of the EU's Markets in Crypto-Assets Regulation (MiCA).In response, on-chain detective ZachXBT pointed out that AscendEX's public hot wallet currently lacks sufficient liquid assets to cover the verified withdrawal demands of users amounting to millions of dollars. Affected users are advised to report to law enforcement and regulatory authorities in their respective countries or regions to hold AscendEX co-founder George (Jing) Cao accountable.

According to on-chain detective ZachXBT, multiple reports indicate that the centralized cryptocurrency exchange AscendEX (formerly Bitmax) has recently delayed or failed to process users' withdrawal requests for several days or even weeks, yet the platform continues to accept user deposits.As early as the initial warning on June 26, ZachXBT pointed out that, after reviewing on-chain data from Arkham and TRM, several known hot wallets of AscendEX (covering EVM, Tron, and Solana networks) severely lack major market cap tokens such as ETH, USDT, and SOL, indicating that the platform is likely facing a serious liquidity crisis.The latest situation shows that, in the 9 days since the first warning, the official X account of AscendEX has remained inactive, and the platform's co-founder George (Jing) Cao has not responded to inquiries from users with large amounts of trapped funds. Currently, ZachXBT strongly advises users unable to withdraw their funds to report to law enforcement and regulatory agencies in their respective countries or regions as soon as possible.It is reported that AscendEX was founded by George (Jing) Cao and Ariel Ling in 2018. In December 2021, the platform was attacked by the hacker group Lazarus Group, resulting in approximately $78 million in asset losses.

ZachXBT released a community alert on platform X, reporting that the centralized exchange AscendEX (formerly Bitmax) has been reported by multiple users for withdrawal delays ranging from days to weeks, and in some cases, withdrawals cannot be processed at all. ZachXBT checked the known hot wallets of the exchange through Arkham and TRM and found that its reserves lack mainstream large-cap tokens such as ETH, USDT, and SOL, which may indicate liquidity issues.AscendEX was founded by George Cao and Ariel Ling in 2018 and was hacked by the Lazarus Group in December 2021, resulting in a loss of approximately $78 million.

ChainCatcher news, according to Cointelegraph, the SOL/ETH trading pair has broken below a rising wedge pattern that has persisted for several months. Based on this pattern, the potential downside target is 0.038 ETH, which could represent a 40% pullback from current levels. Data from May 29 shows that the short-term support for this trading pair is around 0.0628 ETH.On-chain data indicates that memecoin activity in the Solana ecosystem has significantly cooled down. The daily fee revenue of the main issuance platform Pump.fun has dropped to a nearly one-year low, a substantial decline from the peak in the first quarter of 2025. A recent report from Standard Chartered Bank pointed out that the development of Ethereum Layer 2 solutions could pose competitive pressure on Solana.

ChainCatcher news, on-chain analyst Ali Martinez stated on social media that despite the market's volatility, Ethereum remains strong within an ascending parallel channel, pointing towards $6,000.

ChainCatcher news, the former executive director of the Bank of Japan responsible for monetary policy stated that Trump's victory in the U.S. presidential election has increased uncertainty for the Bank of Japan, and if the trend of a weaker yen further expands, it could become a potential catalyst for a rate hike in the near future.Former Bank of Japan official Kazuo Momma said in an interview on Thursday: "The U.S. election has led to increased uncertainty, not just for the Bank of Japan, but for everyone globally. The only reason the Bank of Japan might consider raising rates early could be a rapid depreciation of the yen."Momma believes that a yen-to-dollar exchange rate falling to 160 could be a key threshold for policymakers, but the timing of any exchange rate changes will also depend on the level of public discontent.

ChainCatcher news, according to Decrypt, the dynamic Bitcoin inscription artwork "Ascend" will participate in the live auction at Christie's London on October 10. This piece is the only digital artwork in the auction and is the first dynamic Bitcoin inscription artwork to be featured at Christie's.It is reported that the artwork depicts the beauty of decay of the Tskaltubo sanatorium in Georgia, which was a famous health resort during the Soviet era, built between 1952 and 1962, and has now become a ruin. The piece is inscribed on the Bitcoin blockchain using the Ordinals protocol, switching between day and night modes, reflecting the real-time of Georgia through a 30-minute transition process.

ChainCatcher news, AscendEX (formerly BitMax) has launched a new DEX contract trading product, AscendEX DEX, in collaboration with Orderly Network. AscendEX DEX will initially support Arbitrum, Optimism, and Base chains, allowing the use of USDC as margin.In addition, users trading contracts on AscendEX DEX can earn Orderly Network Merits points. Holding Merits points will qualify users for Orderly Network token airdrops.According to the Web3 asset data platform RootData, AscendEX (Peak) was launched in 2018 and is a full-stack cryptocurrency trading platform co-built by Wall Street investors, providing simple solutions for investment, trading, and profit, helping users maximize the returns on their portfolios.