NVIDIA CMP 90HX

Benchmark Performance

The NVIDIA CMP 90HX presents a peculiar benchmark profile: its raw compute specifications suggest substantial capability, yet the available data shows an average benchmark score of zero and a percentile rank of 50 among all GPUs. This combination indicates that the card's mining-oriented design, with no display outputs and a PCIe 1.0 x4 interface, effectively nullifies its utility in conventional gaming or workstation benchmarks. The FP32 throughput of 21.89 TFLOPS, paired with an equal FP16 rate at a 1:1 ratio, places it theoretically in the upper mid-range of Ampere-based accelerators, but the absence of any recorded benchmark scores in the dataset means no direct performance deltas can be calculated against rivals.

Interpreting the 50th percentile is instructive: it suggests that when positioned against the entire GPU landscape, the CMP 90HX sits at the median, not because of mediocre silicon but because its restricted ecosystem (no outputs, limited PCIe bandwidth) prevents it from competing in standard test suites. The GA102 chip itself is a proven design, and the 21.89 TFLOPS figure is identical for both FP32 and FP16, indicating a 1:1 ratio that is unusual for consumer cards but consistent with compute-oriented variants. However, without benchmark entries, the data cannot confirm whether this theoretical throughput translates into real-world mining or rendering performance. The pixel rate of 136.8 GPixel/s and texture rate of 342.0 GTexel/s are derived from the 80 ROPs and 200 TMUs at the 1710 MHz boost clock, but these metrics are meaningless for a card that cannot drive a display.

The 50th percentile rank, while seemingly neutral, masks a critical reality: the CMP 90HX is an end-of-life product released on July 27, 2021, and the lack of benchmark data likely reflects its niche status. In any comparative analysis, the card's scores—or lack thereof—must be weighted against the fact that it targets a market segment (cryptocurrency mining) that does not rely on traditional GPU benchmarks. The data shows zero average benchmark score, which is not an anomaly but a consequence of the card's design philosophy: no display outputs mean no frame rates, no DirectX tests, and no OpenGL workloads. Thus, any performance interpretation must rely solely on the compute specifications listed, which indicate a capable but unproven accelerator.

How It Compares

The nearestRivals array in the fact pack is empty, which means there are no direct competitor scores, deltaPct values, or names to reference. This absence is itself a finding: the CMP 90HX does not have a measurable peer group in the benchmark database, reinforcing its status as a specialized mining product rather than a general-purpose GPU. Without rival data, the analysis must pivot to the card's internal specifications to infer positioning. The GA102 chip, built on Samsung's 8 nm process with 28,300 million transistors on a 628 mm² die, gives it a transistor density of 45.1M per mm²—this is the same silicon found in several high-end Ampere cards, but the CMP 90HX's 6400 shading units and 200 tensor cores are configured for compute throughput rather than graphics output.

Comparisons to hypothetical rivals would be speculative, so the data-driven statement is that the CMP 90HX lacks any recorded nearest rivals, meaning its benchmark percentile of 50 is an isolated data point. This isolation is typical for mining-specific SKUs that are excluded from standard review suites. The lack of display outputs (listed as "No outputs") and the PCIe 1.0 x4 interface (a severe bottleneck compared to typical PCIe 4.0 x16 slots) further distance it from conventional GPUs. In practical terms, the card cannot be compared to gaming cards because it cannot render frames, and it cannot be compared to compute cards because its benchmark scores are unregistered. The data tells a clear story: this is a product that exists outside the normal comparative framework.

Ray Tracing and Feature Set

The CMP 90HX includes 50 RT cores and 200 tensor cores, which are the hardware blocks required for ray tracing and AI-accelerated workloads. However, the card's API support is robust on paper: DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4. These APIs would enable ray-traced effects in supported games or applications, but the absence of display outputs renders this capability moot for end-user graphics. The RT cores are present, but there is no way to output a rendered image to a monitor. Tensor cores, similarly, could accelerate DLSS or other neural network tasks, but again, the card's mining focus and lack of outputs preclude such use cases.

The feature set is a study in contradiction: it has the hardware for advanced graphics (50 RT cores, 200 tensor cores, DirectX 12 Ultimate) but no way to present the results. The Vulkan 1.4 support is particularly notable, as it is a recent API version, but it cannot be exercised without a display path. In a mining context, these features are irrelevant—mining workloads do not use RT or tensor cores in the same way—so their inclusion is likely a byproduct of using the GA102 die rather than a deliberate design choice. The data shows no benchmark scores to verify ray tracing performance, so any claims about RT capability would be unsubstantiated. For a benchmark database, the CMP 90HX's feature set is a list of unused capabilities.

Power and Cooling

The CMP 90HX has a thermal design power (TDP) of 320 W, which classifies it as a high-power card requiring serious cooling. The fact pack lists it as a dual-slot design, which is standard for this power level, and it requires two 8-pin power connectors. The suggested power supply unit (PSU) rating is 700 W, which accounts for the card's draw plus system overhead. These figures are consistent with other GA102-based cards, but the mining use case typically involves sustained 24/7 operation, which places additional stress on the power delivery system. The 320 W TDP is a fixed number from the fact pack, and the dual-slot cooler is designed to dissipate that heat, but the data does not specify the cooler type (e.g., blower vs. open-air), so no further thermal analysis is possible.

The power requirements are straightforward: 2x 8-pin connectors and a 700 W PSU recommendation. For a mining rig, this means the card needs a robust power supply and adequate airflow. The dual-slot form factor (285 mm in length, 112 mm in height) is typical for high-TDP cards, but the lack of display outputs means the card's I/O bracket is likely just a blank plate, which simplifies installation. The 320 W TDP is a fixed number from the fact pack, and the dual-slot cooler is designed to dissipate that heat, but the data does not specify the cooler type (e.g., blower vs. open-air), so no further thermal analysis is possible. The end-of-life production status suggests that NVIDIA has ceased manufacturing this card, which may affect availability of replacement coolers or support.

FAQ

Q: What is the CMP 90HX's chip architecture and manufacturing process?

A: It uses the GA102 chip on NVIDIA's Ampere architecture, manufactured on Samsung's 8 nm process with 28,300 million transistors on a 628 mm² die.

Q: Does the CMP 90HX have any display outputs?

A: No. The fact pack explicitly lists "No outputs" for display outputs, making it unusable for connecting a monitor.

Q: What is the memory configuration of the CMP 90HX?

A: It has 10 GB of GDDR6X memory on a 320-bit bus, delivering 760.3 GB/s of bandwidth with an effective 19 Gbps memory clock.

Q: What are the power connector and PSU requirements?

A: The card requires 2x 8-pin power connectors and a suggested PSU of 700 W. Its TDP is 320 W.

Q: Which APIs does the CMP 90HX support?

A: It supports DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4, though these are unused due to the lack of display outputs.

Q: When was the CMP 90HX released, and what is its current production status?

A: It was released on July 27, 2021, and its production status is listed as end-of-life.

Who Should Consider It

The CMP 90HX is a card that should only be considered by miners operating in a niche where display output is irrelevant and PCIe bandwidth is not a bottleneck. Its 21.89 TFLOPS FP32 performance and 10 GB of GDDR6X memory with 760.3 GB/s bandwidth are theoretically sufficient for compute-heavy mining algorithms, but the PCIe 1.0 x4 interface severely limits data transfer rates, which could hamper workloads that require frequent host communication. The 50th percentile ranking, combined with zero benchmark scores, means there is no verified performance data to recommend it for any gaming or professional rendering scenario. For high-resolution gaming or content creation, the card is unsuitable because it cannot output video at any resolution or settings.

The data suggests that the intended user is someone running a mining rig where the card is installed alongside other GPUs and never needs to drive a display. The dual-slot design and 320 W TDP require adequate cooling and power, but the card's end-of-life status means buyers are limited to the used market. Given the lack of benchmark scores, potential buyers cannot verify performance against alternatives, making this a high-risk purchase. The 1:1 FP16/FP32 ratio might appeal to certain compute workloads, but again, the PCIe bottleneck and no outputs are disqualifying for most use cases. In summary, only a miner with a specific need for a high-bandwidth, no-output card should consider it, and even then, the absence of data is a red flag.

Memory Subsystem

The CMP 90HX is equipped with 10 GB of GDDR6X memory, a 320-bit bus, and a bandwidth of 760.3 GB/s. The memory clock runs at 1188 MHz base, translating to 19 Gbps effective. These figures are substantial: 760.3 GB/s is a high-bandwidth configuration that would typically support 4K textures and large datasets in gaming or rendering. However, for a card with no display outputs, the memory bandwidth is solely for compute or mining workloads, where it can be fully utilized without the overhead of frame buffer writes. The 10 GB capacity is moderate by modern standards, but for mining algorithms that fit within that limit, the card can operate efficiently.

The 320-bit bus width and GDDR6X type are consistent with a high-end memory subsystem, and the 19 Gbps effective speed is a key factor in achieving the 760.3 GB/s figure. For high-resolution scenarios, this bandwidth would be more than adequate—if the card could output video. In a mining context, the memory subsystem is likely the card's strongest asset, as memory-bound algorithms can leverage the full bandwidth. The absence of benchmark scores, however, means there is no empirical evidence of how this memory configuration performs under load. The data shows the specifications but not the results, leaving the memory subsystem's real-world efficacy unproven. For a card that cannot display images, the memory serves a purely computational role, and 10 GB may limit the size of datasets or DAG files in certain mining applications.