NVIDIA Jetson Orin Nano Super

NVIDIA’s Jetson Orin Nano Super is a compact embedded system-on-module built around the GA10B chip, bringing Ampere architecture features to a 25 W power envelope. This analysis examines its memory subsystem, ray tracing capabilities, competitive positioning, benchmark performance, power considerations, and architectural design, using only the provided technical specifications.

Memory Subsystem

The Jetson Orin Nano Super integrates 8 GB of LPDDR5 memory across a 128-bit bus, operating at an effective speed of 6.4 Gbps. This configuration yields a memory bandwidth of 102.4 GB/s. For an integrated graphics processor (IGP) with a 25 W TDP, this bandwidth figure is notable, as it represents the primary data throughput available to both the CPU and GPU components sharing this unified memory pool.

In practical terms, this bandwidth is sufficient for high-resolution workloads, though with caveats. At 1080p, the 102.4 GB/s figure allows for smooth texture streaming and moderate resolution scaling. At 1440p or 4K, the same bandwidth must serve larger framebuffers and higher-resolution textures, which can strain the subsystem. The 128-bit bus width is moderate; wider buses (typically 256-bit or more) are found in discrete GPUs with higher bandwidth, but this module’s embedded nature prioritizes power efficiency over raw throughput.

The LPDDR5 type is significant because it offers lower power consumption compared to GDDR6 or GDDR6X, aligning with the module’s 25 W TDP. However, this comes at the cost of peak bandwidth, a discrete GPU with similar memory bus width but faster memory would exceed this figure. The effective 6.4 Gbps speed is competitive for an embedded platform in this class, yet the 8 GB capacity means large datasets or high-resolution texture packs may require careful management.

For high-resolution gaming or rendering, the data suggests that the 102.4 GB/s bandwidth will be a limiting factor in scenarios where textures exceed the 8 GB capacity or where memory access patterns are bandwidth-intensive. Benchmark results indicate that this module sits at the 50th percentile among all GPUs, implying it is neither a bandwidth leader nor a laggard. The memory subsystem’s performance is adequate for its intended use case, edge AI inference, robotics, and portable devices, where resolution demands are often secondary to computational throughput.

Ray Tracing and Feature Set

The Jetson Orin Nano Super is built on the Ampere architecture, which is notable for its support of DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4. These API levels include features such as ray tracing, variable rate shading, and mesh shaders in the case of DirectX 12 Ultimate. However, the FACT PACK lists no dedicated ray tracing (RT) cores for this chip. The absence of RT cores is a critical detail: while the APIs support ray tracing workloads, the hardware lacks specialized units to accelerate them.

Instead, the module includes 32 tensor cores, which are designed for AI and machine learning tasks such as tensor operations, inference, and upscaling. These tensor cores can theoretically be repurposed for some ray tracing calculations (e.g., denoising), but they are not a substitute for dedicated RT cores in terms of efficiency. The presence of tensor cores aligns with the Jetson product line’s focus on AI at the edge, not on high-fidelity real-time ray tracing.

The 1024 shading units handle conventional rasterization, while 32 texture mapping units (TMUs) and 16 render output units (ROPs) process texture and pixel output. The pixel rate of 16.32 GPixel/s and texture rate of 32.64 GTexel/s are derived from the clocks and unit counts, but the base and boost clocks are not specified in the FACT PACK, so these rates are fixed values from the data. The FP32 performance is 2.089 TFLOPS, which indicates the raw compute capability for non-tensor workloads.

For ray tracing, the data implies that the Jetson Orin Nano Super is not a suitable platform for real-time ray-traced effects in games. The lack of RT cores means any ray tracing would fall back to compute shaders on the shading units, which would drastically reduce frame rates. The Vulkan 1.4 and DirectX 12 Ultimate support mean APIs are present, but hardware acceleration is absent. This is a key differentiator from higher-tier Ampere GPUs, which include RT cores. The tensor cores could enable ray-traced denoising if software optimizations are applied, but the 2.089 TFLOPS FP32 performance suggests limited headroom for such workloads.

How It Compares

The FACT PACK lists no nearest rivals for the Jetson Orin Nano Super, meaning there are no direct comparison points with specific scores or delta percentages. This absence is itself informative: the module occupies a unique niche as an embedded IGP with a 25 W TDP, rather than a discrete desktop or mobile GPU. Without rival data, comparisons must be qualitative, based on the module’s own specifications.

In the absence of nearestRivals data, the percentileVsAllGpus field provides a relative anchor: this module sits at the 50th percentile among all GPUs in the database. This suggests it outperforms half of the GPUs tracked, which is remarkable for a 25 W embedded part, but it also means it lags behind the top 50% in absolute performance. The benchmark scores are empty, so no aggregate score exists to compare against specific products.

What the data does allow is a comparison against hypothetical alternatives based on the module’s own traits. For example, a discrete GPU with similar FP32 performance (around 2 TFLOPS) would likely consume far more than 25 W and have a larger physical footprint. Conversely, lower-power embedded GPUs might offer less memory bandwidth or fewer tensor cores. The 102.4 GB/s bandwidth is higher than what many low-power IGPs offer, but the absence of RT cores puts it at a disadvantage to any Ampere-based discrete card with RT hardware.

The lack of rivals also means no deltaPct values to cite. Therefore, this section cannot provide specific percentage leads or deficits. Instead, the analysis relies on the fact that the module’s 50th percentile rank and its 2.089 TFLOPS FP32 compute place it in a mid-range position among all GPUs, but its embedded nature and 25 W TDP suggest it targets efficiency over absolute speed.

FAQ

Q: What is the memory bandwidth of the Jetson Orin Nano Super?

A: The module has a memory bandwidth of 102.4 GB/s, based on 8 GB of LPDDR5 memory on a 128-bit bus running at 6.4 Gbps effective speed.

Q: Does this GPU support ray tracing hardware?

A: No. The FACT PACK lists no RT cores. While it supports DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4, which include ray tracing APIs, the hardware lacks dedicated RT cores.

Q: What is the TDP and power connector requirement?

A: The TDP is 25 W. No power connectors are listed, and no suggested PSU is specified. The module is an IGP, so power delivery is likely through the host board.

Q: How many tensor cores does it have, and what are they used for?

A: It has 32 tensor cores. These are designed for AI and machine learning tasks, such as tensor operations and inference, not for accelerating ray tracing.

Q: What is the FP32 performance?

A: The FP32 performance is 2.089 TFLOPS. The FP16 performance is 4.178 TFLOPS, achieved via a 2:1 ratio.

Q: What is the release date and launch MSRP?

A: The release date is 2024-12-16, and the launch MSRP is 249 USD. The production status is Active.

Q: What is the process node and die size?

A: The chip is fabricated on Samsung’s 8 nm process node, with a die size of 200 mm². The transistor count is listed as unknown.

Q: What is the bus interface?

A: The bus interface is PCIe 4.0 x4, which provides a moderate data transfer path for an embedded module.

Benchmark Performance

The FACT PACK contains no benchmark scores for the Jetson Orin Nano Super, the benchmarks array is empty, and the avgBenchmarkScore is 0. This means there are no synthetic or real-world performance numbers to analyze directly. However, the percentileVsAllGpus field provides a single data point: the module sits at the 50th percentile among all GPUs in the database. This percentile is not a score but a rank, indicating that its performance is median relative to the entire population of GPUs tracked.

Without nearestRivals data, there are no deltaPct values to cite for comparisons. The absence of rival scores means any statement about being “30% ahead” or “behind” is impossible. Instead, the analysis must interpret what the 50th percentile implies. A GPU at the 50th percentile would outperform half of all GPUs in the database, which is a strong result for a 25 W embedded part. However, the top 50% of GPUs include many high-end discrete cards with far higher power envelopes, so the Jetson Orin Nano Super is not competitive in absolute terms with those.

The FP32 performance of 2.089 TFLOPS is a fixed compute figure. To contextualize this, one can infer that a typical discrete GPU with similar FP32 output would be considered entry-level. The FP16 performance of 4.178 TFLOPS (2:1 ratio) doubles the compute for workloads that can utilize FP16, which is common in AI inference, a primary use case for the Jetson line. The pixel rate of 16.32 GPixel/s and texture rate of 32.64 GTexel/s are derived from the shading units and clocks, but without clock speeds, these are the only rates available. These rates suggest that at 1080p, the module could handle basic rendering, but at higher resolutions, the limited pixel throughput would become a bottleneck.

Given the empty benchmark data, the only quantitative performance signal is the 50th percentile rank. This indicates that in a broad database of GPUs, the module’s performance is average for the entire population, but that average includes many older and lower-power parts. For AI workloads, the 32 tensor cores and FP16 compute are likely the key differentiators, though no specific tensor performance numbers are provided. The data does not support any claims about specific frame rates or rendering scores.

Who Should Consider It

Based on the provided data, the Jetson Orin Nano Super is suited for specific use cases rather than general-purpose gaming. The 25 W TDP, IGP form factor, and 70 mm length (2.8 inches) suggest it is designed for embedded systems, robotics, portable devices, and edge AI applications. The 8 GB LPDDR5 memory and 102.4 GB/s bandwidth are adequate for AI inference workloads that require moderate memory capacity, but the 2.089 TFLOPS FP32 compute limits heavy graphics rendering.

For users working with AI models (e.g., image classification, object detection), the 32 tensor cores and FP16 performance of 4.178 TFLOPS are the primary assets. The tensor cores accelerate matrix operations common in neural networks, making this module a viable option for on-device inference without needing a discrete GPU. The 50th percentile rank among all GPUs suggests that it is not a high-end compute device, but for its power class, it is competitive.

For gaming or high-resolution rendering, the module is less suitable. The lack of RT cores means no hardware ray tracing, and the 16.32 GPixel/s pixel rate is low compared to discrete GPUs. At 1080p, it might handle older or less demanding titles at low settings, but the data does not provide any game-specific benchmarks to confirm this. The 8 GB memory is sufficient for 1080p textures, but the bandwidth could throttle performance in texture-heavy scenes. For 1440p or 4K, the module would likely struggle due to bandwidth and compute limitations.

The absence of power connectors and suggested PSU suggests that power is delivered via the host board, typical for an IGP. This makes it suitable for systems where a discrete GPU is impractical. The PCIe 4.0 x4 interface is adequate for data transfer from a host CPU, but it is narrower than the x8 or x16 interfaces used by discrete GPUs, which could limit performance in certain data-intensive tasks. Overall, the target user is a developer or engineer building a compact, low-power system for AI or edge computing, not a gamer seeking high frame rates.

Power and Cooling

The Jetson Orin Nano Super has a TDP of 25 W, which is extremely low for a GPU, most discrete graphics cards exceed 100 W. This low TDP means the module can be cooled by passive heatsinks or small fans, though the FACT PACK does not specify a cooler or slot width beyond “IGP” (meaning it occupies no expansion slot). The dimensions are 70 mm in length (2.8 inches) and 45 mm in height (1.8 inches), making it a compact component suitable for small form factor systems.

No power connectors are listed, implying that power is drawn from the motherboard via the PCIe interface or a dedicated board connector. No suggested PSU is provided, but given the 25 W TDP, a standard system power supply would be more than sufficient. The absence of a PSU recommendation is consistent with an embedded module that is often sold as part of a developer kit or integrated into custom boards.

The 8 nm process node from Samsung contributes to the low power draw. With a die size of 200 mm², the transistor density is not specified, but the 8 nm node allows for a balance between performance and efficiency. The memory power is included in the 25 W TDP, as LPDDR5 is designed to be power-efficient. For cooling, the data does not specify a cooler type, but the low TDP means that even a basic passive heatsink could dissipate the heat if airflow is adequate. In enclosed systems, a small fan would be prudent, but the module’s thermal requirements are minimal compared to discrete GPUs.

The pixel rate and texture rate (16.32 GPixel/s and 32.64 GTexel/s) are achieved within this 25 W envelope, indicating that the clocks are tuned for efficiency rather than peak performance. The lack of base and boost clock speeds in the FACT PACK prevents further analysis of thermal behavior under load. However, the 25 W TDP is a hard limit, so sustained workloads would not cause power spikes typical of higher-TDP GPUs. The PCIe 4.0 x4 interface also consumes power, but this is negligible compared to the GPU core.

Architecture and Design

The Jetson Orin Nano Super is built on NVIDIA’s GA10B chip, which is part of the Ampere architecture. The architecture name “Ampere” is shared with NVIDIA’s GeForce RTX 30 series, but the GA10B is a specific variant designed for Tegra (the generation field is “Tegra (Ampere)”). This chip is fabricated on Samsung’s 8 nm process node, with a die size of 200 mm². The transistor count is listed as unknown, which is a gap in the data, but the die size is moderate for a modern GPU.

The chip contains 1024 shading units, 32 TMUs, 16 ROPs, and 32 tensor cores. The shading units handle general-purpose compute and graphics, while the TMUs and ROPs process textures and pixel output. The tensor cores are a key feature, enabling AI acceleration. The absence of RT cores is notable, as it differentiates this chip from other Ampere parts like the GA102 or GA104, which include RT cores. The FP32 performance is 2.089 TFLOPS, and FP16 is 4.178 TFLOPS, with a 2:1 ratio indicating that FP16 throughput is doubled by using tensor cores or specialized FP16 paths.

The memory controller is integrated on-chip, supporting 8 GB of LPDDR5 on a 128-bit bus. The effective memory speed is 6.4 Gbps, yielding 102.4 GB/s bandwidth. The bus interface is PCIe 4.0 x4, which is a lower lane count than discrete GPUs but appropriate for an embedded module. The display outputs are listed as “Portable Device Dependent,” meaning they vary based on the host device, this module does not have fixed display connectors.

The architecture is designed for efficiency: the 25 W TDP is remarkably low for a chip with 1024 shading units. The 8 nm process from Samsung is not as advanced as TSMC’s 5 nm or 4 nm nodes used in some competing chips, but it offers a cost-effective balance for embedded applications. The die size of 200 mm² is relatively large for a 25 W part, suggesting that the chip may have some idle or low-power states to manage thermals. The release date of 2024-12-16 places it as a recent product, and the launch MSRP of 249 USD positions it as an affordable embedded solution.

The production status is Active, meaning it is currently available. The lack of a predecessor or successor in the data suggests it is a standalone product within the Jetson lineup, or that those fields were not populated. The PCIe 4.0 x4 interface supports high-speed data transfer for AI workloads, but the 4-lane width could be a bottleneck for large datasets. Overall, the architecture balances compute capability, memory bandwidth, and power efficiency, but it omits ray tracing hardware and relies on tensor cores for AI-specific tasks.