AMD Radeon Instinct MI25 vs NVIDIA GeForce RTX 5090 D Comparison

The NVIDIA GeForce RTX 5090 D and AMD Radeon Instinct MI25 are separated by nearly eight years of GPU architecture, and the benchmark data reflects that chasm. In the only shared benchmark, Geekbench OpenCL, the RTX 5090 D delivers a score of 310,674 against the MI25’s 68,562, a staggering 353.1% advantage. The RTX 5090 D also holds the overall average score crown at 77,712, placing it in the 92nd percentile of all GPUs, while the MI25’s 68,562 average puts it in the 90th percentile. This is not a contest; it is a generational gap made visible through raw compute numbers.

Head-to-Head Benchmarks

The single direct comparison available is Geekbench OpenCL, and it is decisive. The NVIDIA GeForce RTX 5090 D scores 310,674, while the AMD Radeon Instinct MI25 scores 68,562. The delta is 353.1% in favor of the RTX 5090 D. That means for every point of OpenCL performance the MI25 produces, the RTX 5090 D produces roughly 4.5 points. This is the kind of margin that changes which workloads are feasible on a given card.

Looking at the RTX 5090 D’s broader benchmark suite, its average score of 77,712 is only 1.1% above the AMD Radeon RX 6650M XT (76,904) and 1.6% below the AMD Radeon RX 6850M XT (78,940). Its closest rivals are mobile GPUs and datacenter Tesla parts, not the MI25. The MI25’s nearest rivals are the Intel Arc A770 (68,809, -0.4% difference) and NVIDIA CMP 90HX (69,000, -0.6% difference). These rival lists confirm that the RTX 5090 D operates in a performance tier roughly 13% higher than the MI25’s tier based on average scores alone.

In specific RTX 5090 D tests, the 3DMark Steel Nomad DX12 score is 14,326, and the Geekbench Vulkan score is 376,915. The PassMark G3D score is 44,065, while the GPU Compute score is 28,396. The MI25 has no comparable scores in these tests, so the only apples-to-apples metric remains OpenCL, where the RTX 5090 D is 353.1% ahead. The data shows one winner in every head-to-head test, with the RTX 5090 D taking all 1 win and the MI25 taking 0.

FAQ

Q: Which card has a higher average benchmark score?

A: The NVIDIA GeForce RTX 5090 D has an average benchmark score of 77,712, compared to the AMD Radeon Instinct MI25’s 68,562. This places the RTX 5090 D in the 92nd percentile of all GPUs, while the MI25 sits in the 90th percentile.

Q: How much faster is the RTX 5090 D in OpenCL compute?

A: The RTX 5090 D scores 310,674 in Geekbench OpenCL versus 68,562 for the MI25, a delta of 353.1%. This is the only benchmark where both cards have recorded scores.

Q: Does the MI25 beat the RTX 5090 D in any benchmark?

A: No. In the head-to-head benchmark results, the RTX 5090 D wins the sole test (Geekbench OpenCL), giving it 1 win versus 0 for the MI25.

Q: What are the closest rivals to each card?

A: The RTX 5090 D’s nearest rival is the AMD Radeon RX 6650M XT, which is 1.1% slower, and the AMD Radeon RX 6850M XT, which is 1.6% faster. The MI25’s nearest rival is the Intel Arc A770, which is 0.4% slower, and the NVIDIA CMP 90HX, which is 0.6% faster.

Q: Which card has a higher transistor density?

A: The RTX 5090 D has a transistor density of 122.9 million transistors per square millimeter, while the MI25 has a density of 25.3 million per square millimeter. This is due to the RTX 5090 D’s 5 nm process versus the MI25’s 14 nm process.

Q: What is the production status of each card?

A: The RTX 5090 D is listed as Active production, while the MI25 is End-of-life. The MI25 was released in 2017, and the RTX 5090 D was released in 2025.

Architecture Differences

The architectural divide is massive. The RTX 5090 D uses the GB202 chip built on TSMC’s 5 nm process, featuring the Blackwell 2.0 architecture. The MI25 uses the Vega 10 chip on GlobalFoundries’ 14 nm process, running GCN 5.0. The RTX 5090 D packs 92,200 million transistors into a 750 mm² die, yielding a density of 122.9 million transistors per square millimeter. The MI25 has 12,500 million transistors on a 495 mm² die, for a density of just 25.3 million per square millimeter. That density difference explains why the RTX 5090 D can fit 21,760 shading units, 680 TMUs, and 176 ROPs, versus the MI25’s 4,096 shading units, 256 TMUs, and 64 ROPs.

The RTX 5090 D also brings dedicated hardware that the MI25 lacks entirely: 170 ray tracing cores and 680 tensor cores. The MI25 has neither, making it a compute-only card with no ray tracing or AI acceleration hardware. The RTX 5090 D’s FP32 throughput is 104.8 TFLOPS, and its FP16 is also 104.8 TFLOPS at a 1:1 ratio. The MI25’s FP32 is 12.29 TFLOPS, and its FP16 is 24.58 TFLOPS at a 2:1 ratio. This means the RTX 5090 D is over 8.5 times faster in FP32 and over 4 times faster in FP16, even accounting for the MI25’s FP16 advantage.

Memory architecture is another generational leap. The RTX 5090 D uses 32 GB of GDDR7 on a 512-bit bus, achieving 1.79 TB/s of bandwidth. The MI25 uses 16 GB of HBM2 on a 2048-bit bus, achieving 436.2 GB/s. Despite the MI25’s wider bus, the RTX 5090 D delivers over 4 times the bandwidth thanks to faster memory technology. The RTX 5090 D also supports PCIe 5.0 x16, while the MI25 is limited to PCIe 3.0 x16.

Specification Differences

The two cards differ in nearly every measurable specification. The RTX 5090 D has a base clock of 2017 MHz and a boost clock of 2407 MHz, while the MI25 runs at 1400 MHz base and 1500 MHz boost. Memory clocks are 1750 MHz (28 Gbps effective) for the RTX 5090 D versus 852 MHz (1704 Mbps effective) for the MI25. The RTX 5090 D has 32 GB of GDDR7, while the MI25 has 16 GB of HBM2. The bus width is 512-bit versus 2048-bit, but the RTX 5090 D’s bandwidth is 1.79 TB/s versus 436.2 GB/s.

Pixel and texture rates favor the RTX 5090 D: 423.6 GPixel/s and 1,636.8 GTexel/s versus 96.00 GPixel/s and 384.0 GTexel/s. Power draw is also different: the RTX 5090 D has a TDP of 575 W with a suggested PSU of 950 W and a single 16-pin connector, while the MI25 has a TDP of 300 W with a suggested PSU of 700 W and two 8-pin connectors. Both are dual-slot cards, but the RTX 5090 D is longer at 304 mm versus the MI25’s 267 mm. The RTX 5090 D has display outputs (1x HDMI 2.1b, 3x DisplayPort 2.1b), while the MI25 has no outputs. API support differs as well: the RTX 5090 D supports DirectX 12 Ultimate (12_2) and Vulkan 1.4, while the MI25 supports DirectX 12 (12_1) and Vulkan 1.3. Both support OpenGL 4.6. The RTX 5090 D’s launch MSRP is 2,299 USD; the MI25 has no listed launch MSRP.

The Verdict

The data is unambiguous: the NVIDIA GeForce RTX 5090 D is the superior card in every recorded metric. It wins the only shared benchmark by 353.1%, has a higher average score (77,712 vs 68,562), and sits in a higher percentile (92nd vs 90th). The MI25 is end-of-life, lacks ray tracing and tensor cores, and has less than half the memory capacity. If you need raw compute, the RTX 5090 D delivers 104.8 TFLOPS FP32, while the MI25 manages 12.29 TFLOPS. There is no scenario in the data where the MI25 offers a performance advantage. The only reason to pick the MI25 is if you have a legacy system that requires its specific PCIe 3.0 interface and lower 300 W power draw, but even then, the performance gap is so large that it is hard to justify outside of a very constrained power envelope.

Where Each One Wins

The RTX 5090 D wins in every category where data exists. It wins in OpenCL compute by 353.1%, in memory bandwidth by a factor of 4.1 (1.79 TB/s vs 436.2 GB/s), in FP32 throughput by a factor of 8.5, and in FP16 throughput by a factor of 4.3. It also wins on features: it has ray tracing cores, tensor cores, modern display outputs, and a PCIe 5.0 interface. The MI25 has no wins in the head-to-head benchmarks and no notable advantages in specifications, except for a lower TDP (300 W vs 575 W) and a smaller physical footprint (267 mm vs 304 mm). For workloads like AI inference, ray-traced rendering, or high-bandwidth data processing, the RTX 5090 D is the only viable option from this pair. For a low-power compute node that does not need display output or modern APIs, the MI25 exists, but the data shows it is over 8 times slower in raw FP32 and over 4 times slower in memory bandwidth. The choice is clear for anyone prioritizing performance.