Timing and fate.Two months ago, Jensen Huang made a rare comment about ASICs, stating, “90% of the many ASIC projects worldwide will fail.” Given that NVIDIA focuses on general-purpose GPU architecture, which has a clear competitive relationship with ASICs, this prediction sounds more like a curse.The key point is that this “curse” quickly manifested in Huawei’s Ascend chips. According to foreign media The Information, Huawei is planning to shift its Ascend chip strategy from ASICs to GPGPU (a subset of GPUs). The sudden change by this domestic chip giant has surprised the industry.After seven years, Ascend urgently needs to escape the ASIC isolation.In 2018, Huawei unveiled its Ascend series IP and chip products to the public at the Global Connectivity Conference in Shanghai. This was followed by Huawei’s ambition to cover the entire stack and all scenarios of artificial intelligence.However, constrained by the ASIC-specific scenario isolation, Ascend has soared for seven years but still fails to meet Huawei’s imagination.According to Info Tech, ASICs are chips designed for specific purposes. Due to their architecture being customized for specific algorithms, they perform excellently in localized applications. However, their highly targeted professional mode also limits their general computing capabilities.For example, once the algorithm changes, the performance of ASICs generally declines significantly, and they may even need to be re-customized.It is reported that Ascend’s annual sales can reach around 700,000 units, mainly concentrated in large customer groups. These users rely on the deep customization and optimization support from the Ascend team, making the implementation costs extremely high. For smaller customers with more complex scenarios, they cannot obtain the same level of technical support, which severely restricts Ascend’s scalability.Moreover, Ascend chips are currently mainly specialized for deep learning inference and training, showing weakness in tasks such as graphics rendering, parallel computing, and scientific computing. On one hand, they struggle to meet the general computing needs of diverse scenarios; on the other hand, they are gradually deviating from the mainstream ecosystem, facing the isolation dilemma of a single technical route.Jensen Huang’s comments on ASICs are quite sharp: “While ASICs are highly efficient and effective for single purposes, they lack flexibility and scalability. This ‘single-point optimization’ strategy is difficult to cope with the rapid evolution of AI applications.” He believes that “even those ASIC projects that have not failed find it hard to maintain competitiveness in the long term.”Clearly, this strikes at the core of Ascend. In the process of AI+ being implemented across various industries, the ASIC system remains trapped in a corner, which is obviously inconsistent with its strategic demands. For the future, Huawei must reassess Ascend and seek broader iterative space.GPGPU is solid, while domestic ASICs are fleeting.Looking at the domestic AI chip market, Ascend is not the first ASIC manufacturer to “collapse halfway.” Despite the many martyrs on the road to the rise of domestic chips, the stark contrast between the ASIC industry’s recession and the vigorous vitality of GPGPU remains evident.Statistics show that the domestic ASIC sector has experienced varying degrees of decline over the past three years. In 2023, the earlier launched chip manufacturer, Chipone, collapsed due to the failure of its third-generation product iteration and shortcomings in the local ecosystem. Last year, Core Technology in the PowerPC field was criticized for lacking long-term development potential in its technical route, leading to a funding chain break, with core management going missing.Moreover, in the first half of this year, both Shendong Technology and Huaxia Chip also faced crises. The former was reported to have failed to meet expectations with its 3D spatial computing chip, leading to salary arrears and the dissolution of its Shanghai and Taiwan teams; the latter, which had previously claimed to “fill the gap in domestic heterogeneous processors,” ultimately went bankrupt due to insufficient production capacity to cover R&D costs.In contrast to the bleak ASIC market, domestic GPGPU companies are collectively pushing for IPOs.In the past month, both Moore Threads and Muxi have submitted applications for listing on the Science and Technology Innovation Board. The former excels in full-function GPU layouts, covering AI training to graphics rendering; the latter focuses on general computing chips, performing well in the AI inference market.Based on the predicted fundraising amounts and share issuance ratios, Moore Threads’ market valuation upon listing is expected to reach 80 billion yuan, while Muxi is estimated at 39 billion yuan. Additionally, Wallen Technology, which also focuses on GPGPU, is also advancing its listing process vigorously.The stark contrast between these two chip routes undoubtedly reflects the favorable prospects for GPGPU in the AI era.Industry research reports indicate that GPGPU, based on traditional GPU architecture, removes some acceleration hardware units and adds dedicated vector, tensor, and matrix operation instructions for various needs, achieving higher precision and performance, making it the preferred choice for intelligent computing chips. In the coming years, the market size of China’s GPGPU industry is expected to maintain an annual growth rate of 40%-50%.Considering market realities, Huawei’s decision to abandon ASICs and reconfigure towards GPGPU is both unexpected and reasonable. However, the only question is, with Ascend now on this path, how will it catch up with other domestic GPGPUs and even NVIDIA at the end of the race?Only by reaching the end can one know to turn back; can skill overcome fate?In the marketization process of Ascend chips, the resources and efforts invested by Huawei are undeniable.Previously, media reports indicated that Huawei specifically assigned a technical team of hundreds to adapt and optimize Ascend chips for iFlytek’s large model project to support their business needs. In market development, Ascend relies on Huawei’s closed-loop hardware and software product ecosystem from chip to cloud, naturally possessing bundling sales advantages.This relentless resource empowerment has allowed Ascend chips to dominate the domestic market, even embedding into many leading user business scenarios. Unfortunately, Huawei has nearly exhausted its capabilities in the ASIC field and ultimately failed to transform Ascend’s isolation into a new continent. As it shifts to GPGPU for external competition, can this combination still be effectively applied to Ascend?“Times change, and so do circumstances.” The mainstream view in the industry believes that the barriers Ascend currently faces cannot simply be resolved by piling up resources. “NVIDIA has laid out its CUDA platform and software development community for decades; even if Huawei adopts a dual approach of ‘compatibility + open-source,’ it will take at least 2-3 years to narrow the gap at best.” This is clearly beyond external forces’ intervention.Additionally, due to the previous limitations of ASIC application scenarios, Ascend 910 has “failed” the first batch of supporting governments. Sources reveal that “currently, apart from Huawei’s own research institute repurchasing, only some research scenarios are using (Ascend 910).”Furthermore, as the Xinchuang market is restructuring its “contract quota system” procurement method, the previously designated and bundled sales strategies are gradually losing effectiveness. This also means that Huawei may find it difficult to provide external support for Ascend.Perhaps skill cannot overcome fate. Although Ascend ASICs have realized the need to turn back at a crossroads, they have already lost too much room for error. We can only hope that Ascend, stripped of all buffs, can develop real combat capabilities in the more open GPGPU market.After all, the story of “human determination conquering nature” is more worthy of applause.