AMD Radeon Instinct MI25 vs NVIDIA CMP 90HX Comparison

The NVIDIA CMP 90HX and AMD Radeon Instinct MI25 are both end-of-life, compute-focused cards with no display outputs, aimed at workloads where rendering to a screen is irrelevant. Benchmark data shows they are extremely close in overall OpenCL performance, with the CMP 90HX scoring 69,000 and the MI25 scoring 68,562, a delta of just 0.6%. Both sit at the 90th percentile of all GPUs, placing them in the upper tier of installed hardware, though their architectural approaches to achieving that performance are radically different.

Where Each One Wins

The NVIDIA CMP 90HX takes the single head-to-head benchmark victory, winning the Geekbench OpenCL test with a score of 69,000 against the MI25's 68,562. This represents a 0.6% lead, which is within the margin of what could be considered run-to-run variance, but the data consistently shows the CMP 90HX edging ahead. In terms of its nearest rival comparisons, the CMP 90HX is 0.3% ahead of the Intel Arc A770 and 1.2% behind the AMD Radeon Pro WX 8200, indicating it sits in a tightly contested performance band.

The AMD Radeon Instinct MI25, while losing the direct comparison by a hair, is not without its own advantages. It is 0.4% ahead of the Intel Arc A770 and 1.9% behind the Radeon Pro WX 8200, showing it trades blows with the same competitive set. The MI25's key winning territory is not in raw OpenCL throughput but in its memory configuration, offering 16 GB of HBM2 on a 2048-bit bus. This is a 60% larger memory capacity than the CMP 90HX's 10 GB, which can be decisive for workloads that need to hold large datasets in VRAM. The data suggests the CMP 90HX wins on raw compute speed, while the MI25 wins on memory capacity and bandwidth characteristics.

Architecture Differences

The two cards come from completely different design philosophies and manufacturing processes. The NVIDIA CMP 90HX is built on the GA102 chip using the Ampere architecture, fabricated on Samsung's 8 nm process. It packs 28,300 million transistors onto a 628 mm² die, yielding a transistor density of 45.1 million per square millimeter. This is a modern, dense design that prioritizes raw shader throughput.

The AMD Radeon Instinct MI25 uses the Vega 10 chip with the older GCN 5.0 architecture, built on GlobalFoundries' 14 nm process. It contains 12,500 million transistors on a 495 mm² die, with a density of 25.3 million per square millimeter. This older process and architecture mean the MI25 achieves its performance with fewer, larger transistors. The MI25 does have a unique advantage: its FP16 throughput is 24.58 TFLOPS, which is double its FP32 rate of 12.29 TFLOPS due to a 2:1 ratio, whereas the CMP 90HX has a 1:1 FP16 to FP32 ratio at 21.89 TFLOPS. This makes the MI25 potentially better suited for workloads that can utilize half-precision math.

Feature-wise, the CMP 90HX includes 50 RT cores and 200 tensor cores, which are absent from the MI25's specification sheet. This means the NVIDIA card can accelerate ray tracing and AI tensor operations, while the AMD card relies purely on its shader array. The CMP 90HX also supports DirectX 12 Ultimate (12_2) and Vulkan 1.4, while the MI25 tops out at DirectX 12 (12_1) and Vulkan 1.3, indicating a newer feature set on the NVIDIA side.

Head-to-Head Benchmarks

The only direct benchmark available is Geekbench OpenCL, and the results are remarkably close. The NVIDIA CMP 90HX scores 69,000, while the AMD Radeon Instinct MI25 scores 68,562. The CMP 90HX wins by 438 points, or 0.6%. To contextualize this, the CMP 90HX is 0.3% ahead of the Intel Arc A770 and 1.4% behind the NVIDIA Quadro P6000. The MI25 is 0.4% behind the Arc A770 and 2.0% behind the Quadro P6000.

This narrow margin suggests that in pure compute throughput, neither card has a decisive edge over the other. The CMP 90HX's higher FP32 performance of 21.89 TFLOPS versus the MI25's 12.29 TFLOPS does not translate into a proportional OpenCL lead, likely because the benchmark's workload is not purely shader-bound. The MI25's texture rate is actually higher at 384.0 GTexel/s compared to the CMP 90HX's 342.0 GTexel/s, which could offset some of the NVIDIA card's raw FP32 advantage in certain compute tasks. The pixel rate also favors the CMP 90HX at 136.8 GPixel/s versus 96.00 GPixel/s, but these metrics do not dominate the OpenCL result.

Specification Differences

The two cards differ across nearly every core specification. The NVIDIA CMP 90HX has 6,400 shading units, 200 TMUs, and 80 ROPs, while the AMD MI25 has 4,096 shading units, 256 TMUs, and 64 ROPs. This gives the CMP 90HX a 56% advantage in shader count, but the MI25 has 28% more TMUs. Clock speeds are similar, with the CMP 90HX boosting to 1710 MHz from a 1500 MHz base, and the MI25 boosting to 1500 MHz from a 1400 MHz base, so the NVIDIA card operates at a higher frequency.

Memory is the most significant differentiator. The CMP 90HX uses 10 GB of GDDR6X on a 320-bit bus, delivering 760.3 GB/s of bandwidth. The MI25 uses 16 GB of HBM2 on a 2048-bit bus, delivering 436.2 GB/s. While the NVIDIA card has 74% more bandwidth, the AMD card has 60% more capacity. The MI25's memory runs at 852 MHz (1704 Mbps effective) compared to the CMP 90HX's 1188 MHz (19 Gbps effective), but the HBM2's wider bus compensates for the lower speed.

Power and physical specs are also distinct. The CMP 90HX has a TDP of 320 W, while the MI25 is rated at 300 W. Both are dual-slot cards with 2x 8-pin power connectors and a suggested 700 W PSU. The CMP 90HX is longer at 285 mm (11.2 inches) versus 267 mm (10.5 inches) for the MI25, but their heights are nearly identical at 112 mm and 111 mm. The bus interface differs notably: the CMP 90HX uses PCIe 1.0 x4, which is a strange and limiting choice, while the MI25 uses PCIe 3.0 x16, providing significantly more host bandwidth.

FAQ

Q: Which card has more raw compute power?

A: The NVIDIA CMP 90HX has a higher FP32 rating at 21.89 TFLOPS compared to the AMD MI25's 12.29 TFLOPS. However, in the Geekbench OpenCL test, the CMP 90HX only leads by 0.6%, scoring 69,000 versus 68,562.

Q: Can either card be used for gaming or display output?

A: No. Both cards have "No outputs" listed for their display outputs, meaning they cannot connect to a monitor. They are designed exclusively for compute workloads.

Q: Which card is better for large datasets?

A: The AMD Radeon Instinct MI25 is better suited for large datasets due to its 16 GB of HBM2 memory, which is 60% more than the NVIDIA CMP 90HX's 10 GB of GDDR6X.

Q: Does the NVIDIA card support ray tracing?

A: Yes, the NVIDIA CMP 90HX includes 50 RT cores and 200 tensor cores, while the AMD MI25 has no RT or tensor cores listed. The CMP 90HX also supports DirectX 12 Ultimate (12_2), whereas the MI25 only supports DirectX 12 (12_1).

Q: What is the difference in memory bandwidth?

A: The NVIDIA CMP 90HX has significantly higher memory bandwidth at 760.3 GB/s compared to the AMD MI25's 436.2 GB/s, a 74% advantage. However, the MI25 uses a much wider 2048-bit bus compared to the CMP 90HX's 320-bit bus.

Q: Are these cards currently in production?

A: No, both the NVIDIA CMP 90HX and the AMD Radeon Instinct MI25 are listed as "End-of-life" production status. The CMP 90HX was released in 2021, while the MI25 dates back to 2017.

Q: How do these cards compare to the Intel Arc A770?

A: In the Geekbench OpenCL test, the NVIDIA CMP 90HX is 0.3% ahead of the Arc A770, while the AMD MI25 is 0.4% behind it. All three cards perform within a 0.7% performance band.

The Verdict

The data presents a clear picture of two cards that are nearly tied in overall benchmark performance but serve different practical needs. The NVIDIA CMP 90HX is the choice for raw compute speed and modern feature support. Its 21.89 TFLOPS FP32 performance, 50 RT cores, and 200 tensor cores provide a more versatile compute platform, even if the mining-focused design limits its host interface to PCIe 1.0 x4. The 0.6% lead in OpenCL and 74% higher memory bandwidth make it the stronger option for bandwidth-sensitive workloads.

The AMD Radeon Instinct MI25 is the pick for memory capacity and half-precision compute. Its 16 GB of HBM2 is 60% larger than the CMP 90HX's 10 GB, which can be a deciding factor for inference workloads or large matrix operations that need to fit entirely in VRAM. The 2:1 FP16 ratio, delivering 24.58 TFLOPS, gives it a distinct edge in mixed-precision tasks despite its older GCN 5.0 architecture.

The real-world data shows a negligible 0.6% performance gap, so the deciding factor should be workload requirements. If the task is FP32-heavy and benefits from high bandwidth, the CMP 90HX is the logical choice. If the task is memory-capacity-bound or can exploit FP16 throughput, the MI25's larger frame buffer and doubled half-precision rate make it the more practical option. Both cards are end-of-life, so availability and system compatibility—particularly the CMP 90HX's unusual PCIe 1.0 x4 interface—should be carefully checked before purchase.