AMD Radeon Pro VII vs NVIDIA CMP 90HX Comparison

The Verdict

The AMD Radeon Pro VII and NVIDIA CMP 90HX are fundamentally different tools aimed at different jobs. The Radeon Pro VII is a professional workstation card with display outputs and high memory bandwidth, while the CMP 90HX is a mining-specific GPU with no video outputs at all. The data shows the Radeon Pro VII leads in compute performance: it scores 90,148 in Geekbench OpenCL against the CMP 90HX's 69,000, a 30.6% advantage. For anyone needing a GPU for professional rendering, compute workloads, or any task that requires a monitor connection, the Radeon Pro VII is the clear choice. The CMP 90HX, with its higher FP32 throughput on paper, only makes sense for scenarios where raw shader math matters more than memory bandwidth or display support, but benchmark results indicate the Radeon still wins in the recorded test. If you need a GPU with display outputs and proven driver support, pick the Radeon Pro VII. If you are building a compute-only rig and can accept no display outputs, the CMP 90HX offers higher theoretical FP32 but loses the actual benchmark test.

FAQ

Q: Which GPU is faster in Geekbench OpenCL?

A: The AMD Radeon Pro VII scores 90,148 versus 69,000 for the NVIDIA CMP 90HX, a 30.6% performance advantage for the AMD card.

Q: Can the NVIDIA CMP 90HX be used for normal desktop tasks?

A: No. The CMP 90HX has no display outputs, making it unsuitable for any use requiring a monitor. The Radeon Pro VII has 6x mini-DisplayPort 1.4a outputs.

Q: How do their memory systems differ?

A: The Radeon Pro VII has 16 GB of HBM2 with a 4096-bit bus and 1.02 TB/s bandwidth. The CMP 90HX has 10 GB of GDDR6X with a 320-bit bus and 760.3 GB/s bandwidth.

Q: Which card has more raw FP32 computing power?

A: The CMP 90HX has a higher theoretical FP32 rate at 21.89 TFLOPS, compared to 13.06 TFLOPS for the Radeon Pro VII, though this does not translate to a win in the recorded benchmark.

Q: Are both cards still in production?

A: No. Both the Radeon Pro VII and the CMP 90HX are listed as end-of-life products in the database.

Q: Which GPU supports newer graphics APIs?

A: The NVIDIA CMP 90HX supports DirectX 12 Ultimate (12_2), Vulkan 1.4, and OpenGL 4.6. The Radeon Pro VII supports DirectX 12 (12_1), Vulkan 1.3, and OpenGL 4.6.

Architecture Differences

The architecture gap between these two is stark, representing two very different design philosophies. The AMD Radeon Pro VII uses the Vega 20 chip built on GCN 5.1 architecture, fabricated on a 7 nm process at TSMC. This is a dedicated workstation part from the Radeon Pro Vega generation. It packs 13,230 million transistors on a 331 mm² die, with a transistor density of 40.0M per mm². The GCN design prioritizes memory bandwidth and compute consistency, which is why it uses HBM2 memory with a massive 4096-bit bus.

The NVIDIA CMP 90HX uses the GA102 chip with Ampere architecture, built on Samsung's 8 nm process. This is a much larger chip: 28,300 million transistors on a 628 mm² die, with a transistor density of 45.1M per mm². Ampere introduces dedicated hardware for ray tracing (50 RT cores) and tensor operations (200 tensor cores). The Radeon Pro VII has no RT cores and no tensor cores, relying on pure shader compute through its 3840 shading units. The CMP 90HX has 6400 shading units, a significant count advantage.

The memory architecture also tells a story. The Radeon uses HBM2, which provides enormous bandwidth via a wide 4096-bit interface. The CMP uses GDDR6X on a 320-bit bus, which is narrower but runs at higher effective clock. The Radeon's FP16 performance is listed as 26.11 TFLOPS with a 2:1 ratio, while the CMP does FP16 at 1:1 (21.89 TFLOPS), meaning it does not accelerate half-precision as aggressively.

The CMP also has no display engine, which is a major architectural difference. The Radeon Pro VII has 6x mini-DisplayPort 1.4a outputs, making it a full workstation card. The CMP 90HX is designed solely for mining, which is why NVIDIA gave it no display outputs and a low-speed PCIe 1.0 x4 interface.

Specification Differences

The specification sheets reveal two cards with very different priorities. Clock speeds: the Radeon Pro VII has a base clock of 1400 MHz and boost to 1700 MHz, while the CMP 90HX has a base of 1500 MHz and boosts to 1710 MHz. The CMP is slightly faster at the top end, but not by a wide margin.

Memory is a major divide. The Radeon Pro VII has 16 GB of HBM2 with a 4096-bit bus and 1.02 TB/s bandwidth. The CMP 90HX has 10 GB of GDDR6X on a 320-bit bus with 760.3 GB/s bandwidth. The Radeon has 34% more memory capacity and 34% more bandwidth, while the CMP has GDDR6X at 19 Gbps effective.

Compute units: the Radeon has 3840 shading units, 240 TMUs, and 64 ROPs. The CMP has 6400 shading units, 200 TMUs, and 80 ROPs. The CMP has 67% more shading units but 17% fewer TMUs. Pixel rate favors the CMP at 136.8 GPixel/s versus 108.8 GPixel/s. Texture rate favors the Radeon at 408.0 GTexel/s versus 342.0 GTexel/s.

Power is also different, with the Radeon at 250 W and the CMP at 320 W. The CMP requires a 700 W PSU versus 600 W for the Radeon. Physical dimensions are close: the Radeon is 305 mm long, the CMP is 285 mm long; both are 111-112 mm tall and dual-slot.

The CMP uses PCIe 1.0 x4, a severely limited interface, while the Radeon uses PCIe 4.0 x16. The CMP has no display outputs, while the Radeon has 6x mini-DisplayPort. The CMP has 50 RT cores and 200 tensor cores, which the Radeon lacks entirely.

Head-to-Head Benchmarks

The database contains one direct head-to-head benchmark between these two cards: Geekbench OpenCL. The results are decisive. The AMD Radeon Pro VII scores 90,148 points, while the NVIDIA CMP 90HX scores 69,000 points. That is a 30.6% advantage for the AMD card.

This is significant because the CMP 90HX has a higher theoretical FP32 throughput (21.89 TFLOPS versus 13.06 TFLOPS). Yet in the actual OpenCL benchmark, it loses by a large margin. The data suggests that the Radeon's massive memory bandwidth (1.02 TB/s) and compute-oriented GCN architecture deliver better real-world compute results than the CMP's raw shader count. The CMP also has 50 RT cores and 200 tensor cores, but those do not help in OpenCL compute workloads.

The Radeon also posts a Geekbench Metal score of 108,383 and a Vulkan score of 92,862, though the CMP has no recorded scores for those tests.

Looking at the average benchmark scores, the Radeon Pro VII averages 97,131 across its benchmark set, while the CMP 90HX averages 69,000. The Radeon sits at the 93rd percentile of all GPUs, while the CMP is at the 90th percentile. The Radeon also compares favorably to its nearest rivals, with a 6% edge over the NVIDIA RTX A4500 and a 5% edge over the AMD Radeon Instinct MI60. The CMP's nearest rival is the Intel Arc A770, where it leads by only 0.3%.

Where Each One Wins

AMD Radeon Pro VII wins in:

• Compute workloads: 30.6% ahead of the CMP in Geekbench OpenCL, despite lower theoretical FP32. The memory bandwidth and GCN compute design make it a stronger performer.
• Workstation use: With 6x mini-DisplayPort outputs, it can drive multiple displays and handle professional visual workloads.
• Memory capacity and bandwidth: 16 GB HBM2 at 1.02 TB/s versus 10 GB at 760.3 GB/s. This helps with large datasets and compute tasks.
• API support for OpenCL: The recorded OpenCL score of 90,148 is the only direct benchmark, and it wins.
• Texture rate: 408.0 GTexel/s versus 342.0 GTexel/s for the CMP, a 19% advantage.
• Overall benchmark average: 97,131 versus 69,000, a 40.7% better average.

NVIDIA CMP 90HX wins in:

• Theoretical FP32: 21.89 TFLOPS versus 13.06 TFLOPS. If a workload can use raw FP32 shader throughput and not memory bandwidth, the CMP has a theoretical edge, but the recorded benchmark does not support that.
• Pixel rate: 136.8 GPixel/s versus 108.8 GPixel/s, a 25.7% higher pixel output.
• Ray tracing and tensor: 50 RT cores and 200 tensor cores, which the Radeon lacks. This matters for any future workloads using ray tracing or AI tensor operations, but not for the current benchmark set.
• Shading units: 6400 versus 3840, a 67% more shading units.
• Newer API support: DirectX 12 Ultimate and Vulkan 1.4 versus DirectX 12 (12_1) and Vulkan 1.3.

The verdict is clear for builders: the Radeon Pro VII is the better overall GPU for compute and professional use, with a decisive win in the only recorded head-to-head benchmark. The CMP 90HX is a specialized mining part with no display outputs, and despite its higher theoretical FP32, it fails to win the actual OpenCL test. If you need a working GPU with display support and proven performance, the Radeon Pro VII is the choice. If you have a compute-only workload that specifically uses FP32 shaders and you can work around the lack of display outputs, the CMP 90HX offers a higher FP32 spec, but the data does not show it winning any benchmark.