NVIDIA CMP 30HX vs NVIDIA CMP 90HX Comparison

The NVIDIA CMP 90HX and NVIDIA CMP 30HX are both end-of-life mining-specific GPUs with no display outputs, but they represent two very different tiers of compute capability within NVIDIA’s professional mining lineup. The benchmark data shows a clear hierarchy, with the 90HX delivering a decisive lead in raw compute performance while the 30HX offers a lower-power alternative with a smaller physical footprint. Below is a detailed breakdown of how these two cards compare across architecture, performance, and specifications.

Head-to-Head Benchmarks

The only directly comparable benchmark result in the data is the Geekbench OpenCL test. In this test, the NVIDIA CMP 90HX scores 69,000 points, while the NVIDIA CMP 30HX scores 65,199 points. This gives the 90HX a 5.8% advantage, making it the winner in the sole head-to-head comparison. While a 5.8% lead may seem modest, it is important to note that this is a single workload, and the underlying hardware differences suggest the gap could widen significantly in memory-bandwidth-bound or FP32-heavy tasks.

Looking at the broader performance context, the 90HX’s average benchmark score of 69,000 places it in the 90th percentile of all GPUs. Its nearest rivals are the AMD Radeon Pro WX 8200 (69,870, delta -1.2%) and the NVIDIA Quadro P6000 (69,986, delta -1.4%), meaning the 90HX sits just below these professional workstation cards. Meanwhile, the 30HX’s average score of 63,842 puts it in the 89th percentile, with its closest competitor being the AMD Radeon RX 9060 XT LP (63,830, delta 0.0%) — essentially a statistical tie. The 30HX also trails the AMD Radeon Pro WX 9100 (64,212, delta -0.6%) by a small margin.

In the Vulkan test, only the 30HX has a recorded score: 62,484. There is no comparable Vulkan result for the 90HX in the data, so no direct comparison can be made. However, the 30HX’s Vulkan score is lower than its OpenCL score by roughly 4.2%, which may indicate architectural differences in how Turing handles these APIs. The 90HX’s OpenCL performance, however, is 8.2% higher than the 30HX’s best recorded score in any test, underscoring its compute advantage.

The wins tally is straightforward: the 90HX wins 1 head-to-head test, while the 30HX wins 0. This makes the 90HX the clear performance leader in the only directly comparable metric, though the margin is not overwhelming. For a mining-focused product where compute throughput is king, even a 5.8% lead can translate into meaningful efficiency gains over extended operation.

Architecture Differences

The architectural gap between these two cards is substantial. The 90HX is built on NVIDIA’s Ampere architecture, using the GA102 chip manufactured on Samsung’s 8 nm process node. This is a massive die, measuring 628 mm² and containing 28,300 million transistors, for a transistor density of 45.1M per mm². In contrast, the 30HX uses the older Turing architecture, built on the TU116 chip at TSMC’s 12 nm node. The TU116 die is 284 mm² with 6,600 million transistors, yielding a density of 23.2M per mm². The 90HX therefore has more than four times the transistor count and over twice the die area, which explains its higher compute potential.

The 90HX also features dedicated ray tracing cores (50) and tensor cores (200), which are entirely absent from the 30HX (both fields are null). This gives the 90HX support for DirectX 12 Ultimate (12_2), while the 30HX is limited to DirectX 12 (12_1). Both cards support OpenGL 4.6 and Vulkan 1.4, but the lack of RT and tensor hardware on the 30HX means it cannot accelerate ray-traced workloads or AI inference tasks. For a mining GPU, these features are largely irrelevant, but they do affect the cards’ theoretical compute versatility.

The memory subsystems are also fundamentally different. The 90HX uses 10 GB of GDDR6X memory on a 320-bit bus, delivering 760.3 GB/s of bandwidth. The 30HX uses 6 GB of GDDR6 on a 192-bit bus, with a much lower 336.0 GB/s bandwidth. This 2.26x difference in memory bandwidth directly impacts performance in memory-intensive algorithms, which are common in cryptocurrency hashing. The 90HX’s memory clock is rated at 1188 MHz (19 Gbps effective), while the 30HX’s is 1750 MHz (14 Gbps effective) — the higher effective data rate of GDDR6X is the key differentiator.

Clock speeds are somewhat counterintuitive. The 30HX has a higher base clock (1530 MHz vs 1500 MHz) and boost clock (1785 MHz vs 1710 MHz) than the 90HX. Yet the 90HX still wins on raw throughput due to its vastly higher core counts. The 90HX packs 6,400 shading units, 200 texture mapping units, and 80 ROPs, compared to the 30HX’s 1,408 shading units, 88 TMUs, and 48 ROPs. This translates to 21.89 TFLOPS of FP32 performance for the 90HX versus 5.027 TFLOPS for the 30HX — a 4.35x difference. The 90HX also achieves 342.0 GTexel/s texture rate and 136.8 GPixel/s pixel rate, while the 30HX manages 157.1 GTexel/s and 85.68 GPixel/s.

FAQ

Q: Which card has higher raw compute performance?

A: The NVIDIA CMP 90HX delivers 21.89 TFLOPS of FP32 compute, which is 4.35 times the 30HX’s 5.027 TFLOPS. In the Geekbench OpenCL test, the 90HX scores 69,000 versus 65,199 for the 30HX, a 5.8% lead.

Q: Do these cards support ray tracing?

A: Only the CMP 90HX has dedicated ray tracing cores (50) and tensor cores (200). The CMP 30HX has no RT or tensor cores, and its DirectX support is limited to 12 (12_1) rather than the 90HX’s 12 Ultimate (12_2).

Q: What are the memory specifications for each card?

A: The 90HX has 10 GB of GDDR6X on a 320-bit bus with 760.3 GB/s bandwidth. The 30HX has 6 GB of GDDR6 on a 192-bit bus with 336.0 GB/s bandwidth. The 90HX’s memory bandwidth is 2.26 times higher.

Q: Which card has a smaller physical footprint?

A: The CMP 30HX is 229 mm long (9 inches) and 111 mm tall (4.4 inches) with a width of 35 mm (1.4 inches). The 90HX is longer at 285 mm (11.2 inches) and slightly taller at 112 mm (4.4 inches), with no width specified.

Q: What is the power requirement difference?

A: The 90HX has a TDP of 320 W and requires a 700 W suggested PSU with dual 8-pin connectors. The 30HX has a TDP of 125 W and requires a 300 W suggested PSU with a single 8-pin connector.

Q: Are there any display outputs on either card?

A: No. Both the CMP 90HX and CMP 30HX have no display outputs, as they are designed exclusively for mining use.

The Verdict

The data clearly identifies the NVIDIA CMP 90HX as the superior performer for compute-intensive workloads. Its 4.35x advantage in FP32 throughput, 2.26x memory bandwidth lead, and higher Geekbench OpenCL score (69,000 vs 65,199) make it the definitive choice for anyone prioritizing raw hashing power. The 90HX also holds a percentile rank of 90 versus the 30HX’s 89, indicating it sits higher in the overall GPU performance distribution.

However, the 30HX is not without merit. Its 125 W TDP versus the 90HX’s 320 W means it consumes 61% less power, and its 300 W suggested PSU requirement makes it far easier to integrate into multi-GPU mining rigs with existing power infrastructure. The 30HX is also significantly shorter (229 mm vs 285 mm), which can be a critical factor in dense chassis configurations. Its higher boost clock (1785 MHz vs 1710 MHz) shows that it is not a slouch in clock-for-clock terms, but it simply lacks the core count to compete.

For a mining operation where electricity costs dominate, the 30HX’s lower power draw could make it more efficient per watt, but the data does not provide a direct efficiency metric. The 90HX’s 5.8% OpenCL lead is modest, but its massive compute advantage suggests that in algorithms leveraging FP32 or memory bandwidth, the gap would be far larger. The 90HX also has the advantage of GDDR6X memory, which is faster but runs hotter — a trade-off that may matter in poorly ventilated mining rigs.

The 30HX has a single launch MSRP of 799 USD, while the 90HX has no listed MSRP. This makes direct price comparison impossible, but the 30HX’s lower power and smaller size may appeal to hobbyist miners. Ultimately, the 90HX is the performance king, while the 30HX is a more practical, lower-power alternative. Choose the 90HX for maximum throughput; choose the 30HX for easier deployment and lower energy demands.

Specification Differences

| Specification | NVIDIA CMP 90HX | NVIDIA CMP 30HX |

|----------------|-----------------|-----------------|

| Chip | GA102 | TU116 |

| Architecture | Ampere | Turing |

| Process Node | 8 nm (Samsung) | 12 nm (TSMC) |

| Transistors | 28,300 million | 6,600 million |

| Die Size | 628 mm² | 284 mm² |

| Transistor Density | 45.1M / mm² | 23.2M / mm² |

| Base Clock | 1500 MHz | 1530 MHz |

| Boost Clock | 1710 MHz | 1785 MHz |

| Memory Clock | 1188 MHz (19 Gbps effective) | 1750 MHz (14 Gbps effective) |

| Memory Size | 10 GB | 6 GB |

| Memory Type | GDDR6X | GDDR6 |

| Memory Bus Width | 320 bit | 192 bit |

| Memory Bandwidth | 760.3 GB/s | 336.0 GB/s |

| Shading Units | 6400 | 1408 |

| TMUs | 200 | 88 |

| ROPs | 80 | 48 |

| RT Cores | 50 | null |

| Tensor Cores | 200 | null |

| Pixel Rate | 136.8 GPixel/s | 85.68 GPixel/s |

| Texture Rate | 342.0 GTexel/s | 157.1 GTexel/s |

| FP32 Performance | 21.89 TFLOPS | 5.027 TFLOPS |

| FP16 Performance | 21.89 TFLOPS (1:1) | 10.05 TFLOPS (2:1) |

| TDP | 320 W | 125 W |

| Power Connectors | 2x 8-pin | 1x 8-pin |

| Suggested PSU | 700 W | 300 W |

| DirectX Support | 12 Ultimate (12_2) | 12 (12_1) |

| Length | 285 mm (11.2 inches) | 229 mm (9 inches) |

| Height | 112 mm (4.4 inches) | 111 mm (4.4 inches) |

| Width | null | 35 mm (1.4 inches) |

| Release Date | 2021-07-27 | 2021-02-24 |

| Launch MSRP | null | 799 USD |

| Avg Benchmark Score | 69000 | 63842 |

| Percentile vs All GPUs | 90 | 89 |