AMD Radeon PRO W6800 vs NVIDIA CMP 90HX Comparison

The AMD Radeon PRO W6800 and NVIDIA CMP 90HX occupy very different corners of the GPU market, yet comparing them reveals how architectural priorities shape real-world performance. The W6800 is a workstation card built for compute and rendering, while the CMP 90HX is a mining-focused part with no display outputs. The recorded data shows only one shared benchmark, but that single result and the surrounding specifications tell a clear story about what each card was designed to do.

Head-to-Head Benchmarks

The only directly comparable measurement in the database is the Geekbench OpenCL score. Here, the AMD Radeon PRO W6800 posts 121,808 points, while the NVIDIA CMP 90HX manages 69,000. That puts the AMD card 76.5% ahead of the NVIDIA part in raw compute throughput. This is not a narrow margin; it is a dominant lead. The W6800 outperforms the CMP 90HX by more than three-quarters of the latter's entire score, which suggests a fundamental difference in how the two cards execute general-purpose workloads.

Looking at the broader context, the W6800’s average benchmark score across all recorded tests is 135,396, placing it in the 96th percentile of all GPUs in the database. Its nearest rivals include the NVIDIA A10M at 135,230 (0.1% behind), the NVIDIA RTX 4000 Ada Generation at 135,218 (0.1% behind), the AMD Radeon Pro W6800X Duo at 135,774 (0.3% ahead), and the AMD Radeon PRO V620 at 136,472 (0.8% ahead). The W6800 sits essentially at parity with these cards, all within a 1.1% band. That means the W6800 is not an outlier in its class; it is squarely among the top-tier workstation GPUs.

The CMP 90HX, by contrast, has an average benchmark score of 69,000 and sits in the 90th percentile. Its nearest rivals are the Intel Arc A770 at 68,809 (0.3% ahead), the AMD Radeon Instinct MI25 at 68,562 (0.6% ahead), the AMD Radeon Pro WX 8200 at 69,870 (1.2% behind), and the NVIDIA Quadro P6000 at 69,986 (1.4% behind). The CMP 90HX is not far from these cards, but it is clearly in a lower performance tier than the W6800. The delta between the two cards in the head-to-head test is 76.5%, which is far larger than any delta within their respective rival groups.

What makes this comparison striking is that the CMP 90HX has a higher raw FP32 throughput on paper. The NVIDIA card delivers 21.89 TFLOPS, while the AMD card delivers 17.83 TFLOPS. Yet in the OpenCL benchmark, the AMD card wins decisively. This indicates that peak theoretical throughput does not translate directly into benchmark performance, especially when memory capacity, bandwidth, and driver optimization come into play. The W6800’s 512.0 GB/s of memory bandwidth is lower than the CMP 90HX’s 760.3 GB/s, but the AMD card has 32 GB of GDDR6 versus 10 GB of GDDR6X. For compute workloads that exceed 10 GB, the CMP 90HX would either fail or spill over, while the W6800 can hold far larger datasets. The benchmark result likely reflects this advantage.

The W6800 also wins the only head-to-head metric, giving it 1 win against 0 for the CMP 90HX. There are no other shared tests in the database, so this comparison is limited in scope. However, the magnitude of the win, combined with the percentile and rival data, makes it reasonable to infer that the W6800 is the stronger compute card across the board.

FAQ

Q: Which card has the higher average benchmark score?

A: The AMD Radeon PRO W6800 has an average benchmark score of 135,396, compared to 69,000 for the NVIDIA CMP 90HX. The W6800 also ranks in the 96th percentile of all GPUs, while the CMP 90HX ranks in the 90th percentile.

Q: How does the Geekbench OpenCL score differ between the two cards?

A: The W6800 scores 121,808 in Geekbench OpenCL, while the CMP 90HX scores 69,000. This gives the AMD card a 76.5% lead in that specific test.

Q: What memory configurations do the two cards use?

A: The W6800 has 32 GB of GDDR6 memory on a 256-bit bus, providing 512.0 GB/s of bandwidth. The CMP 90HX has 10 GB of GDDR6X memory on a 320-bit bus, providing 760.3 GB/s of bandwidth.

Q: Are these cards still in production?

A: No, both are listed as end-of-life products. The W6800 was released on June 7, 2021, and the CMP 90HX was released on July 27, 2021.

Q: Which card has more shading units and tensor cores?

A: The CMP 90HX has 6,400 shading units and 200 tensor cores, while the W6800 has 3,840 shading units and no tensor cores listed in the database. The CMP 90HX also has 50 ray tracing cores, compared to 60 for the W6800.

Q: What is the bus interface difference between the two?

A: The W6800 uses PCIe 4.0 x16, a standard full-bandwidth interface. The CMP 90HX uses PCIe 1.0 x4, which is a very old and narrow connection, likely because mining workloads do not require high host-to-device throughput.

The Verdict

The data points to a straightforward conclusion: the AMD Radeon PRO W6800 is the better compute card for general workloads. Its 76.5% lead in the only shared benchmark is decisive, and its average score of 135,396 versus 69,000 places it in a completely different performance class. The W6800 also has 32 GB of memory, which is triple the CMP 90HX’s 10 GB, and it sits at the 96th percentile of all GPUs, while the CMP 90HX sits at the 90th percentile.

For anyone who needs a GPU for compute, rendering, or any task that involves running OpenCL kernels, the W6800 is the clear choice from this data. The CMP 90HX, despite having higher FP32 TFLOPS and more shading units, cannot translate that into benchmark wins. Its lack of display outputs and its PCIe 1.0 x4 interface further limit its usability outside of dedicated mining rigs.

The CMP 90HX does have one advantage in the raw specifications: memory bandwidth. At 760.3 GB/s, it outpaces the W6800’s 512.0 GB/s by a significant margin. For workloads that are bandwidth-bound and fit within 10 GB, the CMP 90HX might perform well, but the database does not contain a benchmark that isolates that scenario. Based on recorded measurements, the W6800 wins.

Specification Differences

The two cards differ in nearly every major specification. The W6800 uses a 7 nm process from TSMC, while the CMP 90HX uses an 8 nm process from Samsung. The W6800 has 26,800 million transistors on a 520 mm² die, while the CMP 90HX has 28,300 million transistors on a 628 mm² die. This gives the W6800 a higher transistor density of 51.5M per mm², compared to 45.1M per mm² for the CMP 90HX.

Clock speeds also differ. The W6800 has a base clock of 1575 MHz and a boost clock of 2322 MHz, while the CMP 90HX has a base clock of 1500 MHz and a boost clock of 1710 MHz. The memory clocks are 2000 MHz with 16 Gbps effective for the W6800, and 1188 MHz with 19 Gbps effective for the CMP 90HX.

Memory capacity and type are major differences: 32 GB of GDDR6 versus 10 GB of GDDR6X. Bus width differs as well, 256-bit versus 320-bit, and bandwidth favors the CMP 90HX at 760.3 GB/s versus 512.0 GB/s.

Compute unit counts vary. The W6800 has 3,840 shading units, 240 texture mapping units, and 96 raster operations units. The CMP 90HX has 6,400 shading units, 200 texture mapping units, and 80 raster operations units. The W6800 has 60 ray tracing cores, while the CMP 90HX has 50. The CMP 90HX has 200 tensor cores, while the W6800 has none listed.

Power and connectivity differ substantially. The W6800 has a TDP of 250 W with a suggested PSU of 600 W, while the CMP 90HX has a TDP of 320 W with a suggested PSU of 700 W. The W6800 uses 1x 6-pin and 1x 8-pin power connectors, while the CMP 90HX uses 2x 8-pin. The bus interface is PCIe 4.0 x16 for the W6800 and PCIe 1.0 x4 for the CMP 90HX.

Display outputs are a stark contrast: the W6800 has 6x mini-DisplayPort 1.4a outputs, while the CMP 90HX has no outputs at all. Physical dimensions also differ, with the W6800 at 267 mm length, 120 mm height, and 50 mm width, while the CMP 90HX is 285 mm long and 112 mm high with no width recorded. Both are dual-slot cards.

The W6800 has a launch MSRP of 2,249 USD, while the CMP 90HX has no recorded launch MSRP.

Architecture Differences

The W6800 is built on AMD’s RDNA 2.0 architecture, using the Navi 21 chip. It belongs to the Radeon Pro Navi generation, which is part of the Navi II series. The CMP 90HX is built on NVIDIA’s Ampere architecture, using the GA102 chip, and belongs to the Mining GPUs generation.

The process nodes reflect different foundry choices: TSMC at 7 nm for AMD, Samsung at 8 nm for NVIDIA. This gives the W6800 a denser transistor packing, 51.5M per mm² versus 45.1M per mm², despite the CMP 90HX having more total transistors.

Ray tracing and tensor capabilities diverge. The W6800 has 60 ray tracing cores and no tensor cores, indicating a focus on traditional compute and rendering. The CMP 90HX has 50 ray tracing cores and 200 tensor cores, suggesting it was designed with deep learning and matrix operations in mind, though its mining purpose may not leverage those features.

FP16 performance also differs. The W6800 delivers 35.67 TFLOPS at 2:1 ratio, meaning it can double its FP32 rate when using FP16. The CMP 90HX delivers 21.89 TFLOPS at 1:1 ratio, meaning it sees no benefit from FP16. This is a notable architectural difference, as the W6800 can accelerate half-precision workloads significantly.

Both cards support DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4. The W6800’s memory is 32 GB of GDDR6, while the CMP 90HX uses 10 GB of GDDR6X, with the latter offering higher bandwidth but lower capacity.

The CMP 90HX’s PCIe 1.0 x4 interface is a major architectural limitation. It severely restricts host communication, which is fine for mining where data is loaded once, but problematic for interactive or data-intensive workloads. The W6800’s PCIe 4.0 x16 is standard for modern GPUs.

Where Each One Wins

The AMD Radeon PRO W6800 wins in the recorded benchmark, which is Geekbench OpenCL. Its 121,808 score versus 69,000 shows a 76.5% advantage. It also wins on memory capacity, with 32 GB versus 10 GB, which matters for large datasets, machine learning models, or rendering scenes that exceed the smaller card’s memory. The W6800’s higher boost clock of 2322 MHz versus 1710 MHz also suggests better sustained performance in short bursts.

The W6800 wins on display outputs, offering 6x mini-DisplayPort 1.4a, making it usable in multi-monitor workstation setups. The CMP 90HX has no display outputs, so it cannot drive a monitor. The W6800 also has a lower TDP of 250 W versus 320 W, and a lower suggested PSU requirement of 600 W versus 700 W, which could make it easier to integrate into existing systems.

The NVIDIA CMP 90HX wins on raw compute specifications. It has 6,400 shading units versus 3,840, and 21.89 FP32 TFLOPS versus 17.83. It also has higher memory bandwidth at 760.3 GB/s versus 512.0 GB/s, and more tensor cores at 200 versus none. These specs suggest the CMP 90HX could excel in bandwidth-heavy or tensor-heavy workloads, but the database does not include a benchmark that confirms this.

The CMP 90HX also has a larger die at 628 mm² versus 520 mm², and more transistors at 28,300 million versus 26,800 million. Its memory type, GDDR6X, is faster per pin than the W6800’s GDDR6, which explains the bandwidth advantage despite a narrower bus.

In terms of percentile, the W6800 sits at 96th versus 90th, meaning it ranks higher among all GPUs. The nearest rivals tell a similar story: the W6800 competes with cards like the RTX 4000 Ada Generation and A10M, while the CMP 90HX competes with the Arc A770 and Quadro P6000. Those are different performance tiers.

For a user choosing between these two, the W6800 is the obvious pick for any task involving OpenCL, rendering, or general compute. The CMP 90HX, with its mining-specific design and no display outputs, only makes sense in a niche where its high bandwidth and tensor cores are fully utilized, and even then, the lack of benchmark data leaves that unproven. The recorded measurements favor AMD without ambiguity.