NVIDIA CMP 170HX 10 GB

The NVIDIA CMP 170HX 10 GB is a Mining GPUs-generation product released on August 31, 2021. It uses the GA100 chip fabricated at TSMC's 7 nm process, with a die size of 826 mm² and 54,200 million transistors, giving a transistor density of 65.6M per mm². The base clock is 1140 MHz and the boost clock is 1410 MHz. The card contains 4480 shading units, 280 texture mapping units, and 128 raster output units, with a memory clock of 1215 MHz corresponding to 2.4 Gbps effective. The bus interface is PCIe 1.0 x4, there is no series or codename recorded, and production status is end-of-life with no predecessor or successor entry.

Benchmark Performance

The benchmark data for the CMP 170HX 10 GB is sparse. The benchmarks array is empty, and the nearest-rivals array is likewise empty, so there are no sampled scores, no rival names, and no deltaPct values to draw on. The single placement metric in the record is the percentile vs all GPUs, which the database sets at 50. That value places the card at the median of the GPU population: half of all recorded GPUs score below it and half score above it. The average benchmark score field is 0, indicating that no aggregate performance sample has been stored.

In the absence of benchmark runs, the specification sheet provides the only quantitative performance targets. FP32 throughput is rated at 12.63 TFLOPS. FP16 throughput is 50.53 TFLOPS when computed at the 4:1 ratio. The pixel fill rate is 180.5 GPixel/s and the texture fill rate is 394.8 GTexel/s. These are raw compute rates rather than application-level scores, so they describe the card's arithmetic capacity, not its frame rates.

Because the DirectX, OpenGL, and Vulkan API fields are all null, standard graphics benchmark software would have no API entry point to target. The performance story of this card is therefore a compute story. The 50th percentile offers a positional anchor: within the database's all-GPU ranking framework, this part sits in the median band. The recorded compute rates are substantial on paper, but they cannot be translated into game-scene performance figures without API support or benchmark entries. The pixel rate of 180.5 GPixel/s and texture rate of 394.8 GTexel/s reinforce the same conclusion: this is a throughput-oriented device whose measurable output is arithmetic and texturing speed.

Ray Tracing and Feature Set

Ray tracing hardware is absent from the data. The rtCores field is null, meaning no dedicated RT core count is recorded for this GA100-based product. Without an RT-core entry, there is no evidence of hardware acceleration for ray intersection or traversal in the specification. The card does include 280 tensor cores, which are the Ampere-architecture units for matrix arithmetic; their presence supports AI-style compute but does not substitute for RT cores.

The feature set is further constrained by the API table. DirectX, OpenGL, and Vulkan are all reported as null values. The displayOutputs field is "No outputs." Together, these fields indicate a device that cannot drive a conventional graphics workload to a monitor. The architecture label is Ampere, and the chip is GA100, but the generation label is "Mining GPUs," a designation aligned with the absence of render-focused components. The 280 tensor cores remain the only specialized compute blocks listed beyond the standard shading units, texture mapping units, and raster output units. The absence of an RT core count, in particular, separates this entry from any expectation of hardware-accelerated ray tracing; the data records no such capability.

Power and Cooling

The power requirements are stated plainly. TDP is 250 W. The suggested PSU rating is 600 W. The board uses two 8-pin power connectors. Slot width is listed as "IGP," a classification that does not match standard multi-slot descriptions, and the data does not include length, height, or width dimensions. The missing dimensions leave the physical cooling solution unspecified; the 250 W TDP is the only thermal design point in the record.

The 600 W suggested PSU is the recommended system-level power budget, not merely a board-level number. The two 8-pin connectors are the physical delivery path for the 250 W TDP. Because the product is end-of-life, the power and cooling data describes a legacy installation rather than a new purchase. The combination of 250 W TDP, 600 W PSU guidance, and two 8-pin connectors is internally consistent: the connectors provide ample headroom for the TDP, and the PSU recommendation accounts for other system components.

FAQ

Q: Does the CMP 170HX support hardware ray tracing?

A: No. The rtCores field is null, so no dedicated ray tracing core count is present in the data.

Q: What memory configuration does the card use?

A: The card has 10 GB of HBM2e memory on a 5120-bit bus, with a bandwidth of 1.56 TB/s. The memory clock is 1215 MHz, listed as 2.4 Gbps effective.

Q: Does the card have display outputs?

A: No. The displayOutputs field states "No outputs," and the DirectX, OpenGL, and Vulkan API fields are all null.

Q: What is the launch MSRP?

A: The launch MSRP is 4,299 USD.

Q: What power connectors and PSU rating does it require?

A: The card requires two 8-pin power connectors. The suggested PSU rating is 600 W and the TDP is 250 W.

Q: What is the production status and release date?

A: The production status is end-of-life. The release date is August 31, 2021, and no successor is recorded.

How It Compares

The nearest-rivals array for the CMP 170HX 10 GB is empty. There are no rival names, no rival scores, and no deltaPct values to cite, so a comparison between this card and any specific named GPU cannot be assembled from the database. The only comparative figure is the percentile vs all GPUs at 50. That midpoint rank is the sole positional statement the data supports.

The generation field "Mining GPUs" places the card in a category separate from general-purpose graphics parts. The PCIe 1.0 x4 bus interface, the absence of display outputs, and the null graphics API entries all distinguish it from a typical desktop accelerator in the database's structure. The product record has no predecessor or successor, leaving the CMP 170HX 10 GB isolated in the dataset. A reader looking for a rival comparison will find none in the data; the 50th percentile is the only quantitative reference point, and it does not carry the weight of a head-to-head delta.

Memory Subsystem

The memory subsystem is defined by three quantitative values: 10 GB of HBM2e, a 5120-bit bus width, and 1.56 TB/s of bandwidth. The memory clock is 1215 MHz, or 2.4 Gbps effective. The 5120-bit interface is the widest single entry in this data record, and the 1.56 TB/s bandwidth is correspondingly large.

For high-resolution workloads, the bandwidth figure is the metric that determines how much data can be moved for textures, geometry, and computing buffers. A 1.56 TB/s rate is sufficient to keep the 4480 shading units and 280 texture mapping units fed during compute-heavy operations. The 10 GB capacity acts as the working-set ceiling; once a workload exceeds that limit, data must be re-fetched over the PCIe 1.0 x4 interface, which is narrow relative to the memory pipe. However, the card has no display outputs and no graphics API entries, so "high resolution" in the conventional rendering sense does not apply.

The HBM2e type and 5120-bit bus align with the GA100 die's design; the large 826 mm² package accommodates the memory stacks. The 1.56 TB/s bandwidth is the standout figure in the memory profile, and it is paired with a modest 10 GB capacity. That pairing suggests an emphasis on sustained data movement for compute kernels rather than on storing large rendering scenes.