Is Huawei Catching Nvidia in Chip Performance?

Some claim the Ascend 910B & 910C outperform the Nvidia A100 & H100

Huawei, one of China’s tech giants, is making headline waves with its latest AI chips.

Huawei’s “Ascend series” of chips (910B & 910C) are being touted as a formidable challenger to NVIDIA’s dominance in the AI hardware market.

While Huawei has made significant progress in GPU performance, they will likely never catch NVIDIA… the ascent is too steep for the “Ascend.”

---

Huawei’s Latest AI Chips (September 2024)

1. Huawei Ascend 910B (vs. NVIDIA A100)

The Ascend 910B chip has demonstrated promising performance metrics:

• Efficiency: Benchmarks show it delivers 80% of the efficiency of NVIDIA's A100 for large language model (LLM) training. In some cases, it reportedly outperforms the A100 by 20%.

• Executive Commentary: Wang Tao, COO of Huawei’s Ascend and Kunpeng ecosystem, stated that there is “not much difference” between the computing power of the Ascend 910B and NVIDIA’s A100.

Context for Performance Claims:

Huawei’s performance claims are based on controlled benchmarks, often optimized for its proprietary ecosystem.

While impressive, these results may not reflect real-world performance across diverse AI workloads.

Additionally, factors such as energy efficiency and cost-effectiveness at scale remain unverified.

---

2. Huawei Ascend 910C (vs. NVIDIA H100)

Huawei is also developing the Ascend 910C, which is positioned as a direct competitor to NVIDIA’s H100 GPU:

• Competitive Edge: Reports suggest that the 910C is designed to match the H100’s performance.

• Launch Timeline: Expected to ship as early as October 2024.

• Orders and Valuation: Huawei is reportedly in discussions for orders exceeding 70,000 chips, valued at around $2 billion.

Challenges with 910C:

While Huawei aims to rival NVIDIA’s H100, this claim remains aspirational.

The company has yet to release comprehensive benchmarks or demonstrate the chip’s performance at scale in real-world applications, such as training large AI models or deploying generative AI solutions.