AMD Radeon Instinct MI60 vs NVIDIA CMP 90HX Comparison

The AMD Radeon Instinct MI60 and NVIDIA CMP 90HX are both end-of-life accelerators, but they target fundamentally different workloads and deliver sharply contrasting benchmark results. The data shows a decisive single-head-to-head victory for the AMD part, yet a closer examination of the architecture and specification sheets reveals why each card exists and who might still find one useful. The MI60 is a compute-oriented data center product, while the CMP 90HX is a purpose-built mining device with no display outputs.

Head-to-Head Benchmarks

The only shared benchmark result in the data is the Geekbench OpenCL test, and it produces a clear winner. The AMD Radeon Instinct MI60 scores 92,488, while the NVIDIA CMP 90HX scores 69,000. This translates to a 34% advantage for the MI60 in this specific workload. The margin is substantial, placing the AMD card in a different performance tier for OpenCL compute tasks.

Context from the nearest rivals reinforces this gap. The MI60's average benchmark score is 92,466, placing it at the 93rd percentile of all GPUs. Its closest competitor in the data is the NVIDIA RTX A4500, which averages 91,671, a mere 0.9% behind. The MI60 also edges out the RTX A4500 Mobile (91,134, 1.5% behind). However, it sits below the AMD Radeon Pro VII (97,131, 4.8% ahead) and the AMD Radeon RX 7900M (97,487, 5.2% ahead). The MI60 is, therefore, a strong performer within its peer group, but not the absolute top of the pack.

The CMP 90HX, by contrast, has an average score of 69,000, placing it at the 90th percentile. Its nearest rivals show a much tighter cluster around its score. The Intel Arc A770 (68,809) is just 0.3% behind, and the AMD Radeon Instinct MI25 (68,562) is 0.6% behind. Slightly ahead are the AMD Radeon Pro WX 8200 (69,870, 1.2% ahead) and the NVIDIA Quadro P6000 (69,986, 1.4% ahead). The CMP 90HX is essentially in the middle of a dense pack of mid-range performers, whereas the MI60 is closer to the top of its own tier.

The benchmark results indicate that for general-purpose OpenCL compute, the MI60 is the far superior option. The 34% lead is not a marginal difference; it is a significant performance chasm that would impact any workload reliant on raw compute throughput. The CMP 90HX, while scoring reasonably well in absolute terms, does not compete at the same level as the MI60 in this test.

Architecture Differences

The two cards are built on fundamentally different architectures, process nodes, and design philosophies. The MI60 uses the Vega 20 chip based on AMD’s GCN 5.1 architecture, fabricated on a 7 nm process at TSMC. It contains 13,230 million transistors on a 331 mm² die, yielding a transistor density of 40.0M / mm². In contrast, the CMP 90HX uses NVIDIA’s GA102 chip based on the Ampere architecture, built on Samsung’s 8 nm process. This larger chip packs 28,300 million transistors onto a 628 mm² die, with a density of 45.1M / mm².

Memory configurations diverge significantly. The MI60 is equipped with 32 GB of HBM2 memory on a 4096-bit bus, delivering a massive 1.02 TB/s of bandwidth. The CMP 90HX has 10 GB of GDDR6X memory on a 320-bit bus, providing 760.3 GB/s. The MI60’s memory subsystem is designed for large datasets and high-bandwidth compute, while the CMP 90HX’s smaller, faster GDDR6X pool is typical for mining workloads that are less memory-capacity sensitive.

Compute resources also differ. The MI60 has 4,096 shading units, 256 TMUs, and 64 ROPs. The CMP 90HX has 6,400 shading units, 200 TMUs, and 80 ROPs. The NVIDIA card has more shading units and ROPs, but fewer TMUs. Critically, the CMP 90HX includes 50 RT cores and 200 tensor cores, which the MI60 lacks entirely. These features are irrelevant for mining but could be useful for other tasks, though the CMP 90HX's lack of display outputs limits its utility.

Clock speeds show a different trade-off. The MI60 has a base clock of 1200 MHz and a boost of 1800 MHz, while the CMP 90HX has a higher base of 1500 MHz but a lower boost of 1710 MHz. The MI60’s higher boost clock, combined with its memory bandwidth, helps explain its OpenCL lead. Pixel and texture rates follow suit: the MI60 achieves 115.2 GPixel/s and 460.8 GTexel/s, while the CMP 90HX achieves 136.8 GPixel/s and 342.0 GTexel/s. The CMP 90HX has a higher pixel rate, but the MI60 has a much higher texture rate.

FAQ

Q: Which card has a higher OpenCL benchmark score?

A: The AMD Radeon Instinct MI60 scores 92,488 in Geekbench OpenCL, which is 34% higher than the NVIDIA CMP 90HX’s score of 69,000.

Q: What are the memory capacities and types of each card?

A: The MI60 has 32 GB of HBM2 memory, while the CMP 90HX has 10 GB of GDDR6X memory.

Q: Do these cards support ray tracing?

A: The NVIDIA CMP 90HX has 50 RT cores, while the AMD Radeon Instinct MI60 has no RT cores listed in the data.

Q: What is the production status of these cards?

A: Both the AMD Radeon Instinct MI60 and the NVIDIA CMP 90HX are listed as end-of-life products.

Q: Which card has a higher transistor density?

A: The NVIDIA CMP 90HX has a transistor density of 45.1M / mm², which is higher than the AMD Radeon Instinct MI60’s density of 40.0M / mm².

Q: What is the bus interface for each card?

A: The MI60 uses PCIe 4.0 x16, while the CMP 90HX uses a much older PCIe 1.0 x4 interface.

The Verdict

The data points to a clear choice for compute-oriented users: the AMD Radeon Instinct MI60. Its 34% lead in OpenCL performance, combined with 32 GB of HBM2 memory and a modern PCIe 4.0 interface, makes it the superior accelerator for general-purpose compute tasks. Its 93rd percentile ranking versus the CMP 90HX’s 90th percentile further underscores the gap. The MI60’s higher texture rate (460.8 GTexel/s vs. 342.0 GTexel/s) and higher boost clock (1800 MHz vs. 1710 MHz) support this conclusion.

The NVIDIA CMP 90HX, however, is not without its own rationale. It has a higher pixel rate (136.8 GPixel/s vs. 115.2 GPixel/s) and more shading units (6,400 vs. 4,096), which could theoretically benefit certain workloads. Its 200 tensor cores and 50 RT cores are features the MI60 does not have, but their utility is questionable in a card with no display outputs. The CMP 90HX’s PCIe 1.0 x4 interface is a severe bottleneck that would cripple data transfer in any modern system, making it an impractical choice for general computing. Its design is clearly optimized for mining, where the CPU-GPU data path is minimal.

For a benchmark database, the verdict is straightforward: the MI60 is the better compute card. The CMP 90HX is a specialized product for a narrow use case, and its performance in the only available benchmark reflects that. Users seeking a data center accelerator should choose the MI60. Users with a specific need for a mining card with high pixel throughput might consider the CMP 90HX, but its low bus interface and lack of outputs render it a niche item. The MI60 wins the head-to-head decisively.

Specification Differences

• Chip: AMD Vega 20 vs. NVIDIA GA102
• Architecture: GCN 5.1 vs. Ampere
• Generation: Radeon Instinct (MIx) vs. Mining GPUs
• Process Node: 7 nm (TSMC) vs. 8 nm (Samsung)
• Foundry: TSMC vs. Samsung
• Transistors: 13,230 million vs. 28,300 million
• Die Size: 331 mm² vs. 628 mm²
• Transistor Density: 40.0M / mm² vs. 45.1M / mm²
• Base Clock: 1200 MHz vs. 1500 MHz
• Boost Clock: 1800 MHz vs. 1710 MHz
• Memory Clock: 2 Gbps effective vs. 19 Gbps effective
• Memory Size: 32 GB vs. 10 GB
• Memory Type: HBM2 vs. GDDR6X
• Memory Bus Width: 4096 bit vs. 320 bit
• Memory Bandwidth: 1.02 TB/s vs. 760.3 GB/s
• Shading Units: 4096 vs. 6400
• TMUs: 256 vs. 200
• ROPs: 64 vs. 80
• RT Cores: None vs. 50
• Tensor Cores: None vs. 200
• Pixel Rate: 115.2 GPixel/s vs. 136.8 GPixel/s
• Texture Rate: 460.8 GTexel/s vs. 342.0 GTexel/s
• FP32: 14.75 TFLOPS vs. 21.89 TFLOPS
• FP16: 29.49 TFLOPS (2:1) vs. 21.89 TFLOPS (1:1)
• TDP: 300 W vs. 320 W
• Power Connectors: 1x 6-pin + 1x 8-pin vs. 2x 8-pin
• Bus Interface: PCIe 4.0 x16 vs. PCIe 1.0 x4
• Display Outputs: 1x mini-DisplayPort 1.4a vs. No outputs
• DirectX Support: 12 (12_1) vs. 12 Ultimate (12_2)
• Vulkan Support: 1.3 vs. 1.4
• Dimensions: 267 mm (10.5 inches) vs. 285 mm (11.2 inches) length
• Release Date: 2018-11-17 vs. 2021-07-27
• OpenCL Benchmark: 92,488 vs. 69,000