After the release of Tesla FSD V14 Lite's real-world driving videos in Korea, controversy over HW3 performance has emerged, suggesting a complex interplay of hardware limitations, model distillation, and the unique characteristics of Korean road environments.
Want the next one from this channel too?
Following the release of real-world driving videos of Tesla FSD V14 in Korea, many discussions have ensued. In particular, scenes where FSD makes unstable judgments in complex situations, such as night intersections or U-turn lanes, have amplified user questions about the performance limits of HW3 hardware.
In the US, Tesla vehicles equipped with HW3 are known to handle traffic lights relatively stably. This stability tends to be maintained even during night driving. However, in Korea, the same HW3 hardware shows different patterns, which may stem from differences in software versions and national road and signal systems.
The camera installed in Tesla HW3 is confirmed to use an AR0136 image sensor. This sensor is approximately 1.2 megapixels with a resolution of 1280x960 and is a 1/3-inch CMOS sensor. It also supports HDR (High Dynamic Range) mode and adopts a rolling shutter method, designed to collect optical data in various environments.
Based on currently available data, it is difficult to definitively confirm whether the exact CFA (Color Filter Array) arrangement of the Tesla Hardware 3 camera is RCCB, RCCC, or another method. Therefore, drawing premature conclusions about the technical limitations of the HW3 camera based solely on fragmented sensor specification information can be risky. The precise technical specifications remain unconfirmed.
Tesla HW3 is equipped with a dual FSD SOC (System-on-Chip) and a Neural Network Accelerator (NNA). According to official announcements at the time, these two NNAs combined were said to deliver performance of 72 trillion operations per second, or 72 TOPS (Tera Operations Per Second). This was considered significant computational power for its launch in 2019.
In complex deep learning models like Transformer models, simply computational power is not the only important factor. How quickly data can be retrieved from memory and brought to the processing unit, i.e., data bandwidth, critically affects overall performance. Even if the processing unit is excellent, performance can degrade due to bottlenecks if data is not supplied in a timely manner.
In a Tesla official webcast in Q1 2026, Elon Musk directly stated that HW3's memory bandwidth is about one-eighth of HW4's. This suggests that HW3 has significant limitations in data processing speed compared to HW4. This difference in memory bandwidth can be interpreted as a warning that it could cause a significant bottleneck in the performance of HW3-based FSD.
According to the FSD V14 release notes, the intelligence of the HW4 version was 'distilled' to fit HW3. This is a common technique where a smaller 'student model' learns to mimic the behavior of a larger, higher-performing 'teacher model' operating on HW4, making it runnable on HW3. This process aims to reduce model size while preserving as much important knowledge as possible, but performance degradation can occur in specific situations, even if average performance is maintained.
Model distillation techniques can be effective in improving or maintaining overall average performance, but they do not guarantee identical performance in all situations. According to distillation research by Google Research, even if overall average performance improves, performance may actually decrease in certain subgroups or classes with insufficient data. Additionally, there is a risk that errors from the teacher model could be transferred to or amplified in the student model, potentially weakening its ability to handle edge cases.
The road environment in Korea presents numerous complex and unique situations, or 'long-tail scenarios,' that differ from those in the United States. Specific examples include U-turn conditions, lane-specific signals, bus-only lanes, repeated merge and diverge points over short distances, Korean character road markings, and complex intersection geometries. These factors present greater challenges for autonomous driving systems compared to the US road environment.
According to Tesla's official Q1 2026 data, the V14.3 mainline version improved its reinforcement stage to better handle 'long-tail cases.' It also explained that perception capabilities in visibility scenarios were enhanced, demonstrating efforts to strengthen responsiveness to complex environments.
Tesla officially announced that customers drove over 1 million km in Korea within a month of the FSD supervised release. However, the distance driven by customers and the actual data used for training are not identical, and what matters is the distribution of data, rather than just the quantity. That is, whether data effectively reflecting Korea's unique and complex long-tail situations has been collected and learned is crucial to FSD performance.
Strong evidence confirms that the HW3 computer is constrained in running current large FSD models. Tesla's disclosed HW3 architecture and the actual memory configuration of the board support this, and Elon Musk explicitly stated that HW3's memory bandwidth is about one-eighth of HW4's. The hypothesis that these limited hardware resources, a scaled-down model, and Korea's unique road environment combine to cause the current unstable phenomena of FSD in Korea is gaining traction.
Answers come from the transcript, with the exact spot cited.
Want the next one from this channel too?
When ElectrochemK S publishes, we'll write it up like the one you just read and email it to you.
{channel} published 9 in the last 7 days.
We skip Shorts. You can unfollow any time.
