Recent buzz surrounding the launch of the domestically developed Kimi K3 large language model has sparked widespread discussion both within China and internationally. With an impressive 28 trillion parameters, the K3 model has reached a top-tier level of performance globally. Remarkably, it ranks first in front-end programming, demonstrating the significant advancements made in AI development. However, running such a massive model demands robust computing platforms, making the performance of AI hardware an essential factor.
In terms of parameters, current market offerings favor NVIDIA’s Blackwell series, specifically the B200 and B300 models, which are considered the most powerful. Nonetheless, Kimi K3 is undergoing optimization for domestic AI platforms, with some foreign analysts comparing Huawei’s Ascend 950DT against NVIDIA’s B300 in terms of efficiency when running the K3 model. It’s worth noting that these comparisons are based on estimated parameters rather than direct machine tests, as Huawei’s Ascend 950DT is still in the early deployment stages domestically, and overseas availability remains limited.
Visual data shared by industry insiders indicates that Huawei’s Ascend 950DT outperforms Nvidia’s B300 in certain benchmarks, especially in terms of token generation efficiency per watt—a critical factor for large-scale AI tasks. The data also compares the 950DT to Huawei’s previous-generation Ascend 910C platform, highlighting that the newer chip achieves superior energy efficiency, an impressive feat considering it is built on two generations behind the B300’s manufacturing process.
Huawei’s self-developed AI platform continues to demonstrate resilience despite its older manufacturing process, underscoring the effectiveness of their strategic approach—often referred to as “Tao’s Law,” which emphasizes system-level optimization over raw hardware advancements alone.
Further insights reveal the Ascend 950 series comprises two main models: the 950PR and 950DT. While both share the same core processor, they differ in memory configurations, catering to distinct AI tasks. Huawei states that the 950PR employs the Ascend 950 core paired with HiBL 1.0 memory, optimized for inference prefill and recommendation scenarios. This setup significantly reduces costs compared to high-performance HBM3e/4e memory options.
In contrast, the 950DT focuses on inference decode stages and training tasks. It features Huawei’s HiZQ 2.0 technology, which boosts memory capacity to 144GB and delivers a memory access bandwidth of 4TB/s. In addition, the interconnect bandwidth has been increased to 2TB/s, enabling more efficient data transfer and processing.
Huawei’s Atlas 950 supernode cluster—a flagship AI supercomputing system—can support up to 8,192 Ascend 950DT-based cards interconnected at high speeds. This scale is over 20 times larger than the Atlas 900 supernodes (which contain 384 cards) and nearly 57 times larger than NVIDIA’s NVLink V144 system. The aggregate computing power of these 8,192 cards reaches a staggering 8 exaFLOPS in FP8 precision and 16 exaFLOPS in FP4, with a total interconnect bandwidth of 16.3 petabytes per second and a total memory capacity hitting 1,152 terabytes.
Compared to the Atlas 900, this configuration offers a 17-fold performance increase, reaching approximately 4.91 million transactions per second (TPS). Such advancements highlight Huawei’s significant strides in developing competitive AI hardware to support the demands of next-generation large language models like Kimi K3.