AMD Radeon Pro W5700X vs NVIDIA CMP 90HX Comparison

NVIDIA CMP 90HX and AMD Radeon Pro W5700X occupy very different corners of the GPU landscape, one built for a specific compute task and the other for professional graphics within a proprietary ecosystem. The recorded data reveals a stark performance gap in the one benchmark they share, but also highlights fundamental architectural and specification divergences that define their respective use cases.

Head-to-Head Benchmarks

The only direct comparison available in the database is the Geekbench OpenCL test, and the result is decisively one-sided. The NVIDIA CMP 90HX scores 69,000, while the AMD Radeon Pro W5700X scores 43,810. This gives the NVIDIA card a 57.5% advantage in raw OpenCL throughput. That is not a marginal lead; it is a dominant margin that places the two products in entirely different performance tiers for this particular workload.

Context from the nearest rivals reinforces this gap. The CMP 90HX sits at the 90th percentile among all GPUs, with its closest competitors being the Intel Arc A770 (68,809, just 0.3% behind) and the AMD Radeon Instinct MI25 (68,562, 0.6% behind). The W5700X, by contrast, sits at the 87th percentile, but its average benchmark score across all tests is only 54,828, dragged down by its OpenCL result. Its nearest rivals include the NVIDIA GeForce RTX 4080 (54,247, 1.1% behind) and the AMD Radeon RX 6750 GRE 12 GB (55,698, 1.6% ahead). The W5700X's OpenCL score is nearly 25,000 points below the CMP 90HX's result, a chasm that no other metric in the data can bridge.

It is importantly the W5700X has additional benchmark entries in Metal (75,427) and Vulkan (45,246), which are not shared with the CMP 90HX. These scores suggest the W5700X is much stronger in Apple's Metal API than in OpenCL, but since the CMP 90HX has no corresponding results, a direct head-to-head on those APIs is impossible. The only shared test, OpenCL, is a clear win for the CMP 90HX.

Where Each One Wins

The CMP 90HX wins decisively in raw compute throughput, as measured by OpenCL. Its 69,000 score versus 43,810 is a 57.5% margin, and its 90th percentile ranking confirms it is near the top of the database for this metric. The card's massive FP32 throughput of 21.89 TFLOPS, combined with 6,400 shading units, 200 TMUs, and 80 ROPs, points to a design optimized for parallel computation. Its memory subsystem, with 10 GB of GDDR6X on a 320-bit bus delivering 760.3 GB/s of bandwidth, is built to feed that compute engine at high speed. This is a card that wins on brute-force number crunching.

The W5700X wins in areas that are not directly benchmarked in the shared test. It has 16 GB of GDDR6 memory, which is 60% more capacity than the CMP 90HX, and its 256-bit bus delivers 448.0 GB/s. It also supports display outputs, with 1x HDMI 2.0b and 4x Thunderbolt ports, while the CMP 90HX has no outputs at all. The W5700X's higher boost clock of 2,040 MHz versus 1,710 MHz suggests better per-clock efficiency, and its 7 nm process node from TSMC versus the CMP 90HX's 8 nm node from Samsung indicates a more modern manufacturing process. For users who need display connectivity, larger memory capacity, or operation within Apple's MPX ecosystem, the W5700X is the only viable choice.

FAQ

Q: How much faster is the NVIDIA CMP 90HX than the AMD Radeon Pro W5700X in OpenCL?

A: The CMP 90HX scores 69,000 versus 43,810 for the W5700X, a 57.5% advantage for the NVIDIA card.

Q: Which card has more memory and what are the memory types?

A: The W5700X has 16 GB of GDDR6, while the CMP 90HX has 10 GB of GDDR6X. The CMP 90HX has a wider 320-bit bus and higher bandwidth at 760.3 GB/s versus 448.0 GB/s.

Q: Can the NVIDIA CMP 90HX be used for display output?

A: No. The CMP 90HX has no display outputs. The W5700X has 1x HDMI 2.0b and 4x Thunderbolt ports.

Q: What is the power requirement difference?

A: The CMP 90HX has a TDP of 320 W and a suggested PSU of 700 W. The W5700X has a TDP of 205 W and a suggested PSU of 550 W.

Q: How do the cards compare in transistor count and die size?

A: The CMP 90HX has 28,300 million transistors on a 628 mm² die. The W5700X has 10,300 million transistors on a 251 mm² die.

Q: What are the release dates for these products?

A: The CMP 90HX was released on 2021-07-27, and the W5700X was released on 2019-12-10.

Specification Differences

The two cards diverge on nearly every core specification. The CMP 90HX uses an 8 nm process from Samsung, while the W5700X uses a 7 nm process from TSMC, giving the AMD card a smaller and more modern manufacturing node. Transistor counts reflect this: 28,300 million for the NVIDIA chip versus 10,300 million for the AMD chip, on die sizes of 628 mm² and 251 mm² respectively. Transistor density is comparable at 45.1M per mm² for NVIDIA and 41.0M per mm² for AMD.

Clock speeds tell a mixed story. The CMP 90HX has a higher base clock at 1,500 MHz versus 1,243 MHz, but the W5700X has a much higher boost clock at 2,040 MHz versus 1,710 MHz. Memory clocks differ as well: the CMP 90HX runs at 1,188 MHz with 19 Gbps effective, while the W5700X runs at 1,750 MHz with 14 Gbps effective. The CMP 90HX uses GDDR6X memory with a 320-bit bus for 760.3 GB/s bandwidth, while the W5700X uses GDDR6 with a 256-bit bus for 448.0 GB/s.

Compute resources are heavily skewed toward the NVIDIA card: 6,400 shading units, 200 TMUs, and 80 ROPs versus 2,560 shading units, 160 TMUs, and 64 ROPs. The CMP 90HX also has 50 RT cores and 200 tensor cores, while the W5700X has none of either. Pixel and texture rates are close, with the CMP 90HX at 136.8 GPixel/s and 342.0 GTexel/s versus 130.6 GPixel/s and 326.4 GTexel/s for the W5700X. FP32 performance is 21.89 TFLOPS for the CMP 90HX versus 10.44 TFLOPS for the W5700X, a 2.1x difference. FP16 is 21.89 TFLOPS (1:1) for the CMP 90HX versus 20.89 TFLOPS (2:1) for the W5700X.

Physical and interface differences are stark. The CMP 90HX is dual-slot with 2x 8-pin power connectors, a 700 W suggested PSU, and a PCIe 1.0 x4 bus interface. The W5700X is quad-slot with no listed power connectors, a 550 W suggested PSU, and an Apple MPX bus interface. The CMP 90HX measures 285 mm (11.2 inches) in length and 112 mm (4.4 inches) in height, while the W5700X measures 305 mm (12 inches) in length. The CMP 90HX supports DirectX 12 Ultimate (12_2) and OpenGL 4.6, while the W5700X supports DirectX 12 (12_1) and OpenGL 4.6; both support Vulkan 1.4.

Architecture Differences

The underlying architectures could not be more different. The CMP 90HX is built on NVIDIA's Ampere architecture with the GA102 chip, part of the Mining GPUs generation. It is a compute-oriented design that includes dedicated RT cores and tensor cores, reflecting an architecture that can handle ray tracing and AI workloads even if this particular product has no display outputs. The W5700X uses AMD's RDNA 1.0 architecture with the Navi 10 chip, part of the Radeon Pro Mac (Navi Series) generation. It has no RT cores and no tensor cores, indicating a focus on traditional graphics and compute without hardware acceleration for ray tracing or tensor operations.

The process technology differs significantly: Samsung's 8 nm node for NVIDIA versus TSMC's 7 nm node for AMD. The CMP 90HX packs 28,300 million transistors into 628 mm², while the W5700X fits 10,300 million into 251 mm². This gives the NVIDIA chip a higher transistor count and density, but the AMD chip achieves a higher boost clock of 2,040 MHz on a smaller, more efficient process. The CMP 90HX's memory is GDDR6X, which is a higher-bandwidth technology than the W5700X's GDDR6, and its 760.3 GB/s bandwidth is 69.7% higher than the AMD card's 448.0 GB/s.

The API support also differs: the CMP 90HX supports DirectX 12 Ultimate (12_2), which includes features like mesh shaders and variable rate shading, while the W5700X only supports DirectX 12 (12_1). Both support OpenGL 4.6 and Vulkan 1.4. The CMP 90HX has no display outputs, whereas the W5700X has 1x HDMI 2.0b and 4x Thunderbolt. The bus interfaces are also incompatible: PCIe 1.0 x4 for the CMP 90HX versus Apple MPX for the W5700X, meaning the W5700X is specifically designed for Apple's Mac Pro ecosystem while the CMP 90HX is a standalone mining card.

The Verdict

The data points to a clear split. For pure OpenCL compute performance, the NVIDIA CMP 90HX is the superior product by a 57.5% margin, and its 90th percentile ranking versus the W5700X's 87th percentile confirms this. The CMP 90HX delivers 21.89 TFLOPS of FP32 performance, has 6,400 shading units, and offers 760.3 GB/s of memory bandwidth, all of which explain its dominant benchmark result. Anyone prioritizing raw compute throughput in OpenCL should choose the CMP 90HX.

The W5700X, however, is the only option for users who need display outputs, with 1x HDMI 2.0b and 4x Thunderbolt ports. It also offers 16 GB of memory versus 10 GB, which is a substantial advantage for large datasets. Its lower TDP of 205 W versus 320 W makes it more power-efficient, and its smaller die size of 251 mm² on a 7 nm process indicates a more compact and efficient design. Its Apple MPX bus interface means it is tailored for Mac Pro systems, whereas the CMP 90HX's PCIe 1.0 x4 interface is unusual and potentially limiting for general-purpose use.

The verdict is straightforward: the CMP 90HX wins on raw compute and is the pick for OpenCL-heavy workloads. The W5700X wins on memory capacity, display connectivity, and platform compatibility, making it the pick for professional graphics in Apple environments. There is no overlap in their intended use cases, and the benchmark data reinforces that separation.