First, let’s supplement whether Xiaomi’s Xuanjie 3nm chip production is compliant. This article provides a comprehensive analysis: Research on the compliance of Xiaomi’s Xuanjie 3nm chip.
In summary, the new regulations for January’s bis production have ultimately landed at 30 billion transistors, mainly restricting cloud high-performance computing chips. The Xuanjie has 19 billion transistors, and as long as it is packaged by companies on the overseas whitelist, it complies with the latest production due diligence rules. There are no issues here.
However, the biggest problem is that the rules will change dynamically, which mainly depends on the geopolitical power struggles, making it quite difficult to predict.
From Fin’s expert:
The Xuanjie SoC uses the same X925 big core as MediaTek, and both are based on the N3E process. However, in single-core benchmark scores, it surprisingly surpasses MediaTek’s years of accumulation (3100 vs. 2900), with a gap of over 200 points, which is close to generational differences.
From the parameters currently available, this benchmark score may have a traceable basis.
MediaTek is only responsible for the external 5G modem chip. The SoC/CPU/GPU is developed by Xiaomi using ARM’s public version.
Compared to the Dimensity 9400, the Xuanjie’s X925 super big core actually stacks two cores, and the frequency is also quite impressive. This may be one of the main reasons why the Xuanjie has significantly higher single-core/multi-core benchmark scores than the Dimensity 9400: the benchmark score is 8% higher, and the frequency is 7.4% higher.
The CPU frequency is higher, and the downstream configuration is also quite substantial. Rumor has it that the X925 has a 1MB L2 cache and a 16MB L3 cache, which is much larger than MediaTek’s public version X925 with a 512KB L2 cache and an 8MB L3 cache.
Another possible reason is that the Dimensity 9400 has an integrated 5G modem chip, while the Xuanjie SoC’s modem chip is external, which significantly reduces the AP power density, potentially allowing for more headroom in heat dissipation.
Currently, the first impression is that the Xuanjie chip may have made many cost-ignoring trade-offs to achieve impressive benchmark scores.
The Dimensity 9400’s benchmark power consumption is already astonishing, starting at 7W for single-core and approaching 20W for multi-core. The Xuanjie chip’s CPU frequency is 7.4% higher, and the voltage is likely higher as well, meaning the single-core benchmark power consumption is at least 20% higher than that of the Dimensity.
If it were a real load in mass production, the single-core difference would likely not be this significant under similar power consumption walls.
With two super big core X925 CPUs, a 10-core CPU configuration, and 16MB L3 cache, the chip area is certainly not small, which will put pressure on costs. The Dimensity 9400 and Qualcomm 8E are priced at $200 each, while Xiaomi’s chip will not be low-cost.
After reviewing the information, it seems that the impressive benchmark scores of the Xuanjie SoC do not involve any black technology; the parameters are traceable and align with the configuration trade-offs.
However, this does not diminish the amazement brought by this SoC: in just four years since its establishment, it has managed to catch up with MediaTek’s 15 years of accumulation in SoC performance/power tuning, which is truly impressive for the team.
Lei Jun has quietly created a big news story this time, with excellent confidentiality. From Lei Jun’s Weibo, it can be seen that since the restart of the Xuanjie project over four years ago, a total of 13.5 billion RMB has been invested in R&D, with a team of 2,500 people, and this year alone, 6 billion has been invested in R&D, with plans to invest 50 billion over the next ten years.
Achieving this level in the first chip production is truly not easy. Back in the day, Zeku also invested heavily over 10 billion for several years, starting in August 2019 with a team size of about 3,000, and only in 2023 did they begin to show their potential with the second generation.
Huawei HiSilicon started making mobile phone chips in 2012, and the first generation K3V2 was like a joke compared to Google’s self-developed Tensor chip today. It took nearly 8 years to reach the top level in the industry with the Kirin 9000 in 2020, standing shoulder to shoulder with Qualcomm, which was already quite impressive. Xiaomi has achieved remarkable results in just four years.
Building a team of two to three thousand people for Zeku was not easy. Back then, they made a big splash in the global market to recruit talent, and semiconductor workers in North America were basically approached multiple times. I checked various discussion threads, and it seems that Xiaomi may have absorbed talent from Pinecone, HiSilicon, Unisoc, and the talent lost from Zeku. It feels like a significant number of elite team members from Zeku’s 2023 dissolution may have been recruited by Xiaomi.
Compared to Xiaomi’s Xuanjie, Google’s Pixel phone’s Tensor chip has far fewer resources, probably only a few hundred people. Although Pixel’s KPI only cares about daily scenarios and not benchmark scores, the scores shouldn’t fall two generations behind. When I saw Pixel moving all Qualcomm’s architecture over, I thought Tensor G4/G5 would rise, but it still looks like this.The emergence of Xiaomi’s Xuanjie SoC is comparable to achieving what Zeku could not complete back in the day. Although the performance in actual mass production remains to be seen and may not be as stunning, for a first-generation chip, such performance is already pleasantly surprising.
Lei Jun claims a 50 billion investment over ten years, and from the current results, it seems more sincere than OPPO’s Zeku, which claimed a 50 billion investment but actually invested just over 10 billion. Whether this will trigger a new round of sanctions from the U.S. on the process will depend on Lei Jun’s wisdom and skills.
