EN / 中文

Dense, Sparse or Equivalent TOPS: Why Autonomous Driving Chip Computing Power Numbers Are Hard to Compare

by zhijiazuiqianyan·November 19, 2026

On October 30, Yu Kai, Founder and CEO of Horizon Robotics, addressed a significant issue in the intelligent driving chip industry: how should the computing power metrics labeled by different vendors be compared?

He cited the NVIDIA Orin X as an example, noting that its dense compute is 127 TOPS. If calculated using a 2x metric convention, the sparse compute can reach 254 TOPS. He also questioned the practice where some products claim their equivalent compute reaches 6 to 7 times the dense compute.

From dense compute to sparse compute, and various forms of equivalent compute, manufacturers use many metrics to describe chip performance. However, the calculation conditions behind these numbers are not identical. Consumers and automakers often mistakenly compare results from different standards if they only look at TOPS.

So, what exactly are the differences between these types of computing power? And which metrics should be used to evaluate the actual capabilities of an autonomous driving chip?

01. What Are the Differences Between Dense, Sparse, and Equivalent Compute?

First, let's look at dense compute.

Dense compute refers to the theoretical peak computing capability a chip can achieve under specific data precision and calculation methods without utilizing sparse computation. The unit is TOPS (Tera Operations Per Second), meaning trillions of operations per second.

For example, the dense compute of the NVIDIA Orin X is 127 TOPS. This figure describes the chip's theoretical peak under specific conditions and does not imply that it will sustainably reach 127 TOPS when running a specific autonomous driving model.

It is worth noting that the statistical standard for TOPS itself needs clarification.

Take MAC (Multiply-Accumulate) operations in matrix calculations as an example. One MAC operation includes one multiplication and one addition. In TOPS statistics, this is typically counted as two operations, provided both parties adopt the same counting convention.

If a manufacturer does not specify how operation counts are aggregated, even if both are labeled as TOPS, they may not be directly comparable.

Furthermore, peak computing power under different data precisions such as FP16, INT8, and INT4 cannot be conflated. Calculation precision, operation type, and whether the statistical scope covers NPU, GPU, or multiple computing units must all be clearly defined.

Next, let's look at sparse compute.

The core of sparse computation lies in leveraging the characteristic that some weights in neural networks are zero, thereby reducing the number of operations actually executed.

Using common 2:4 structured sparsity as an example, within specific weight grouping dimensions, at least two out of every four weights are zero. Hardware supporting this method can skip certain zero-value operations, thereby improving effective computing throughput.

In compliant computing scenarios, 2:4 sparse computation can theoretically provide up to approximately twice the computing throughput improvement. This is why some chips report higher figures under sparse compute metrics than under dense compute.

However, there are two prerequisites here.

First, the hardware must genuinely support the corresponding sparse computation method, and the software compiler must be able to identify and invoke the relevant computing units. Just because zeros exist in a model does not guarantee performance gains from sparse computation.

Second, the model itself must meet the corresponding sparsity conditions. Artificially increasing zero values through pruning may affect model accuracy and usually requires training or fine-tuning to restore model performance.

NVIDIA TensorRT documentation provides an example where 53 layers in a network met the sparsity conditions, but ultimately only 22 layers adopted the sparse computation strategy.

The reason for this discrepancy is that for certain computation scales, dense computation might actually be faster.

In other words, even if the hardware supports sparse computation, it does not mean all operators are suitable for this approach.

Therefore, higher sparse compute does not equate to the chip achieving proportional performance improvements across all models and scenarios.

Finally, let's look at equivalent compute.

Compared to the previous two metrics, equivalent compute requires more explicit justification of its calculation basis.

It is not a uniformly defined hardware peak metric, but rather a metric derived by manufacturers after converting the chip's actual runtime capabilities based on specific models, calculation precision, software optimizations, and test conditions.

For instance, the same chip, after operator optimization, model quantization, and computation graph optimization, may achieve higher inference efficiency on a specific model.

Manufacturers can provide an equivalent compute value based on this, but details regarding which model, precision, test conditions, and conversion formula were used must be disclosed.

Without this information, equivalent compute is difficult to serve as a basis for horizontal comparison.

It is particularly important to note that improved model inference efficiency does not mean the physical computing units' peak capability has changed. Directly converting performance gains on a specific model into a larger TOPS figure can easily mislead users into thinking the chip possesses higher general-purpose computing power.

02. Why Does Higher TOPS Not Necessarily Mean Stronger Actual Autonomous Driving Capability?

Even if the statistical standard for dense compute is unified, one cannot judge the actual capability of an autonomous driving chip solely by TOPS.

This is because running models on a chip is not accomplished by computing units alone.

Memory bandwidth is a critical factor.

During inference, autonomous driving models require continuous reading of weights, feature data, and intermediate results. If the computing units process data quickly but memory bandwidth cannot supply data in time, the computing units may remain idle.

In such cases, no matter how high the chip's theoretical peak computing power is, it may not translate into actual inference performance.

Computing unit utilization is equally important.

Different models vary in network structure, operator types, and computation scale. Some models fully utilize the chip's parallel computing capabilities, while others may be limited by operator scale, data access patterns, and task scheduling.

This explains why the same chip can exhibit significantly different actual performance when running different models.

The software stack also impacts final results.

Whether the compiler can effectively map model operators to hardware, and whether it supports efficient operator fusion, memory scheduling, and parallel execution, affects computing unit utilization.

Even if two chips have similar theoretical computing power, the one with a more mature software toolchain and better operator support for common autonomous driving models may perform better in actual inference.

Model quantization follows similar principles.

Converting higher-precision data like FP16 to lower-precision formats like INT8 or INT4 reduces data storage and transmission overhead and improves computing efficiency on some hardware.

However, quantization is not cost-free. Changes in data representation range and precision can affect model output.

Factors such as whether quantization-aware training is used, how calibration is performed, and whether post-quantization model accuracy meets requirements all influence the final result.

Therefore, simply comparing TOPS across different precisions does not indicate which chip is more suitable for autonomous driving.

Additionally, in automotive scenarios, power consumption and heat dissipation are particularly crucial factors.

Peak performance achieved in laboratories or short-term tests may not be sustainable during long-term vehicle operation. Automotive power supply, cooling space, and ambient temperature impose constraints, and the chip may need to simultaneously handle tasks like perception, localization, prediction, and planning.

If a chip only achieves advertised performance at high power consumption, its sustained computing capability in actual vehicles must be evaluated considering power and thermal conditions.

For intelligent driving, computing power must not only be sufficiently high but also stable under whole-vehicle constraints.

03. What Should Be Tested to Truly Compare Chip Capabilities?

If TOPS cannot fully reflect the actual capabilities of an autonomous driving chip, how should the industry establish more valuable evaluation methods?

According to leading perspectives in autonomous driving, the key is not requiring all manufacturers to adopt the same chip architecture or technology route, but rather disclosing metrics at different levels separately.

First, hardware theoretical peak computing power should have clear standards.

Manufacturers must at least specify the calculation precision used, MAC counting rules, the scope of computing units, and whether sparse computation is enabled. Dense and sparse compute should be labeled separately, and results from different conditions should not be mixed.

Moreover, model inference performance should be tested independently.

Representative autonomous driving models should be selected, and tests conducted under the same model version, input size, data precision, and quality requirements, with necessary software environments and optimization conditions disclosed.

Test results should not only provide a peak TOPS figure or average inference speed for a single model. For intelligent driving systems, stability of inference latency is equally important.

Beyond average latency, attention should be paid to high-percentile latency metrics like P95 and P99, observing whether the vast majority of inference tasks complete within specified times and whether slower tasks impact system real-time performance.

Actual throughput, power consumption, and sustained operation performance also need to be disclosed.

Especially under multi-task concurrency, the chip's ability to stably handle different tasks like perception, prediction, and planning is more valuable than short-term peaks from a single model.

It is also necessary to distinguish between chip testing and whole-vehicle system testing.

Chip benchmarks can compare the computing efficiency of hardware and software stacks, but final autonomous driving performance is influenced by sensor inputs, model design, system architecture, and vehicle engineering factors.

A chip's test score cannot be directly equated with the capability of the entire intelligent driving system.

In reality, the industry has begun establishing relevant evaluation criteria. The "Technical Requirements and Test Methods for Computing Chips for Automated Vehicle Driving" (QC/T 1278-2026) was implemented on November 1, 2026, providing industry standards for technical requirements and test evaluations of automated driving computing chips.

However, chip technical testing standards and the computing power conversion metrics used in corporate marketing are not entirely the same issue.

To make computing power data from different manufacturers truly comparable, further clarification of metric definitions, test conditions, and result disclosure methods is still needed.

04. Competition in Autonomous Driving Chips Should Not Become a Race of Computing Power Numbers

From an industry development perspective, dense compute, sparse compute, and equivalent compute each have their own value.

Dense compute reflects the chip's theoretical peak capability under specific calculation precision; sparse compute demonstrates the hardware's potential in utilizing sparse calculations; and equivalent compute, when test conditions are clearly defined, helps understand the chip's actual efficiency when running specific models.

The problem lies not in which metric manufacturers use, but in whether they clearly explain the conditions behind the metrics.

If a chip is labeled with hundreds of TOPS but does not specify the data precision used, whether sparse computation is employed, or which model and test environment corresponds to the figure, then this number fails to illustrate its true capability.

Conversely, if a chip's dense compute is not outstanding, but it achieves better inference performance under the same precision and power constraints through higher computing utilization, a more mature software toolchain, and reasonable model deployment, this is equally valuable technological progress.

Intelligent driving chips ultimately serve vehicle operation, not parameter sheets.

What automakers truly need is not a seemingly larger TOPS figure, but the amount of effective computing that can be stably completed under actual models, real-time requirements, and whole-vehicle power constraints.

Therefore, what the industry needs is not forcing all chips onto the same technology route, but establishing a clear, transparent, and reproducible evaluation method.

Computing power can be calculated in different ways, but comparison standards must not be ambiguous.