NVIDIA Tesla K20Xm

The NVIDIA Tesla K20Xm is an end-of-life compute accelerator from the Kepler generation, built on the GK110 chip using TSMC's 28 nm process. It packs 7,080 million transistors into a 561 mm² die, resulting in a transistor density of 12.6M per mm². With a launch MSRP of 7,699 USD, this dual-slot card was positioned for high-performance computing rather than consumer graphics, a fact underscored by its complete lack of display outputs. The benchmark database shows an average score of 12,547, placing it at the 51st percentile of all GPUs, indicating a mid-pack standing in the overall performance hierarchy. The card's compute focus is evident in its Geekbench scores: 8,035 in Metal and 17,058 in OpenCL.

Benchmark Performance

The Tesla K20Xm's average benchmark score of 12,547 sits at the 51st percentile, meaning it outperforms just over half of all GPUs in the database. This is a modest standing, but the breakdown between the two benchmark tests reveals a distinct compute bias. The Geekbench OpenCL score of 17,058 is more than double the Metal score of 8,035, indicating that the card is heavily optimized for parallel compute workloads rather than graphics rendering. In terms of raw throughput, the K20Xm delivers 3.935 TFLOPS of FP32 performance, supported by 2688 shading units, 224 texture mapping units, and 48 raster operation units. The pixel rate is 40.99 GPixel/s, and the texture rate is 164.0 GTexel/s.

Comparing to its nearest rivals, the K20Xm leads the GeForce GTX 1650 SUPER by a mere 0.3%. The average scores of 12,547 and 12,504 are nearly identical, suggesting that in real-world applications, users would see no measurable difference. Against the GTX 880M, the K20Xm is 0.5% ahead, another negligible margin. The GTX 1070, a popular consumer card, is beaten by 1.8%, which is a more substantial lead. However, the AMD FirePro W5100 overtakes the K20Xm by 1.9%, with an average score of 12,789 versus 12,547.

These small deltas indicate that the K20Xm, despite being an older compute card, holds its own against modern mid-range GPUs. The 1.8% lead over the GTX 1070 is particularly noteworthy, as the GTX 1070 is a well-regarded graphics card. The data suggests that the K20Xm's compute-oriented architecture provides a competitive edge in the OpenCL-heavy workloads that these benchmarks likely represent. The 6 GB GDDR5 memory on a 384-bit bus, offering 249.6 GB/s of bandwidth, provides ample memory bandwidth for large datasets, which is crucial for compute tasks.

How It Compares

NVIDIA GeForce GTX 1650 SUPER: The K20Xm edges out the GTX 1650 SUPER by 0.3% in average benchmark score. With scores of 12,547 and 12,504, the two cards are effectively tied. This is a remarkable outcome for a compute card from 2012, as the GTX 1650 SUPER is a modern consumer GPU. The 0.3% delta is within typical measurement noise, so users should treat these as equivalent in performance.

NVIDIA GeForce GTX 880M: The K20Xm is 0.5% ahead of the GTX 880M. The GTX 880M is a high-end mobile GPU, and its average score of 12,490 is only slightly below the K20Xm's 12,547. This comparison highlights the K20Xm's efficiency in compute tasks, as it matches a mobile part that benefits from newer architecture optimizations.

NVIDIA GeForce GTX 1070: The K20Xm leads the GTX 1070 by 1.8%. The GTX 1070's average score is 12,331, while the K20Xm achieves 12,547. This is the most significant lead among the NVIDIA rivals. The GTX 1070 is a desktop gaming card, and the K20Xm's superiority in these benchmarks suggests that its Kepler compute architecture remains competitive in specific workloads.

AMD FirePro W5100: The K20Xm trails the FirePro W5100 by 1.9%. The FirePro W5100's average score of 12,789 is clearly higher than the K20Xm's 12,547. This is the only rival in the list that outperforms the K20Xm, and the margin is more than the K20Xm's leads over the other NVIDIA cards. The FirePro W5100 is also a workstation card, so this comparison is particularly relevant for professional compute users.

Who Should Consider It

The Tesla K20Xm is not a gaming card; it has no display outputs, so it cannot be connected to a monitor. Instead, it is a compute accelerator designed for servers and workstations. Users who run compute-intensive applications that leverage OpenCL will find the K20Xm's score of 17,058 in Geekbench OpenCL to be quite strong. This is more than double its Metal score of 8,035, confirming that the card is optimized for parallel processing.

For workloads that require large memory pools, the 6 GB GDDR5 memory on a 384-bit bus provides 249.6 GB/s of bandwidth. This is sufficient for many scientific and engineering applications that handle large datasets. The 2688 shading units and 224 TMUs provide high compute throughput, while the 48 ROPs are less critical for non-graphics tasks.

Given its end-of-life status, the K20Xm is best suited for legacy compute environments where Kepler architecture is still supported. The 235 W TDP and suggested PSU of 550 W mean that a robust power supply is necessary. Users considering this card for modern workloads should note that it lacks RT cores and tensor cores, so it cannot accelerate ray tracing or AI inference.

For resolution and settings-based recommendations, the K20Xm is not applicable for gaming. However, in compute benchmarks, it performs on par with modern mid-range GPUs like the GTX 1650 SUPER and GTX 1070. The 1.8% lead over the GTX 1070 suggests that it can handle compute tasks that would typically require a modern desktop GPU. The 0.3% lead over the GTX 1650 SUPER indicates that it is a viable option for compute workloads where a modern entry-level card would suffice.

FAQ

Q: What is the average benchmark score of the Tesla K20Xm?

A: The average benchmark score is 12,547, placing it at the 51st percentile of all GPUs in the database.

Q: How does the K20Xm compare to the GeForce GTX 1070?

A: The K20Xm is 1.8% ahead of the GTX 1070, with an average score of 12,547 versus 12,331.

Q: Does the K20Xm have ray tracing or tensor cores?

A: No. The FACT PACK lists RT cores and tensor cores as null, meaning the card does not include dedicated hardware for ray tracing or AI tensor operations.

Q: What APIs does the K20Xm support?

A: It supports DirectX 12 (11_0), OpenGL 4.6, and Vulkan 1.2.175.

Q: What is the memory configuration of the K20Xm?

A: It has 6 GB of GDDR5 memory on a 384-bit bus, providing 249.6 GB/s of bandwidth. The memory clock is 1300 MHz, with an effective data rate of 5.2 Gbps.

Q: Can the K20Xm be used for gaming?

A: No, because it has no display outputs. It is a compute accelerator without any video outputs, making it unsuitable for direct gaming use.

Ray Tracing and Feature Set

The Tesla K20Xm is built on the Kepler architecture, specifically the GK110 chip, and it does not include any RT cores or tensor cores. This means it lacks hardware acceleration for ray tracing and AI-based tensor operations, which are common in modern GPUs. Instead, the card relies on its raw compute capabilities: 3.935 TFLOPS of FP32 performance, 2688 shading units, 224 TMUs, and 48 ROPs.

The memory subsystem is a 6 GB GDDR5 pool on a 384-bit interface, yielding 249.6 GB/s of bandwidth. The memory clock is 1300 MHz, with an effective data rate of 5.2 Gbps. This configuration provides substantial bandwidth for compute workloads, though it is not as fast as modern GDDR6 or HBM solutions.

In terms of API support, the K20Xm supports DirectX 12 (11_0), OpenGL 4.6, and Vulkan 1.2.175. This allows it to run modern graphics APIs, but without display outputs, it cannot render to a screen. The card is a dual-slot design with a 235 W TDP and a suggested PSU of 550 W. It measures 267 mm (10.5 inches) in length.

The 28 nm process and 7,080 million transistors indicate an older manufacturing node, but the transistor density of 12.6M per mm² is a result of the large 561 mm² die. The card's production status is end-of-life, and it was released on 2012-11-11. Its predecessor is Tesla Fermi, and its successor is Tesla Maxwell.

For compute-oriented users, the OpenCL score of 17,058 is the standout feature, showing that the card excels in parallel compute tasks. The Metal score of 8,035 is lower, reflecting the card's focus on compute rather than graphics. The absence of RT and tensor cores means that users looking for modern features like ray tracing or DLSS will need to look elsewhere, but for raw compute throughput, the K20Xm remains a competitive option.