How does the performance of the latest GPU chips released by Moore Threads actually compare? This is a question many people are concerned about. Based on publicly available data, the key is to understand what the manufacturers emphasize and what they deliberately avoid mentioning.

Many people want to see direct comparison data between GPU chips and mainstream international products, but such data is indeed hard to find. Interestingly, this situation does not exist in the automotive industry—at the launch of domestically produced new energy vehicles, industry benchmarks are often compared side by side. But the GPU market is a bit different.

To be honest, there is an objective gap between domestic AI chips and top international products in terms of generational advancement. Therefore, at this launch, more focus was on comparing the chips with their previous generation—improvements in compute density, better energy efficiency ratios, and other metrics. These indicators have indeed progressed. The single-card computing power may still have gaps, but the approach of domestic chips is to compensate by stacking clusters—if funding is sufficient, stacking tens of thousands or even hundreds of thousands of units is not a problem.

It is worth mentioning that the inter-chip interconnection rate reaches 1314 GB/s, which is a relatively competitive metric. Overall, giving domestic GPU chips more development time, their future performance is worth looking forward to.