AMD Radeon Instinct MI25 vs NVIDIA GB10 Comparison

The NVIDIA GB10 and AMD Radeon Instinct MI25 represent two different eras of accelerator design. The benchmark data shows a clear overall performance leader, but the architectural gap tells a more nuanced story about intended workloads, system integration, and longevity.

Where Each One Wins

The recorded benchmark results show a single head-to-head comparison, and the NVIDIA GB10 wins it decisively. In the Geekbench OpenCL test, the GB10 scores 120,137, while the MI25 scores 68,562. This is a 75.2% advantage for the GB10, a massive margin that places these two products in entirely different performance classes.

The NVIDIA GB10 sits in the 95th percentile of all GPUs tracked in the database. Its nearest rivals include the NVIDIA RTX 4000 SFF Ada Generation (average score 117,088, only 0.3% behind the GB10) and the AMD Radeon PRO W7700 (average score 118,976, which is 1.3% ahead of the GB10). This places the GB10 in a competitive segment alongside modern workstation-class accelerators. The GB10 also leads the NVIDIA Tesla V100 SXM2 16 GB by 2.6% and the NVIDIA RTX A5500 Mobile by 3%, showing it holds its own against established data center parts.

The AMD Radeon Instinct MI25, by contrast, sits in the 90th percentile. Its nearest rivals are the Intel Arc A770 (average score 68,809, just 0.4% ahead of the MI25) and the NVIDIA CMP 90HX (average score 69,000, 0.6% ahead). The MI25 also trails the AMD Radeon Pro WX 8200 by 1.9% and the NVIDIA Quadro P6000 by 2%. These are older or mid-range parts, and the MI25 is competitive within that group, but the gap to the GB10 is enormous.

The GB10 wins the only recorded comparison, and it wins across the board in raw compute. The MI25 does not win any benchmark in the database. For workloads that rely on OpenCL compute, the GB10 is the clear choice. The MI25 retains relevance only in legacy deployments or scenarios where its specific memory characteristics matter, and those are not reflected in the recorded scores.

Architecture Differences

The NVIDIA GB10 is built on the Blackwell 2.0 architecture, using the GB20B chip, and fabricated on a 5 nm process at TSMC. The die size is 382 mm². The AMD Radeon Instinct MI25 uses the GCN 5.0 architecture with the Vega 10 chip, fabricated on a 14 nm process at GlobalFoundries, with a die size of 495 mm² and 12,500 million transistors. The MI25’s larger die is a product of its older, less dense manufacturing process, while the GB10 achieves more performance from a smaller chip.

The GB10 has 6,144 shading units, 384 texture mapping units, and 48 raster output units. It also includes 48 ray tracing cores and 384 tensor cores, which the MI25 lacks entirely. The MI25 has 4,096 shading units, 256 TMUs, and 64 ROPs. The GB10’s compute advantage is substantial: it delivers 29.71 TFLOPS of FP32 performance and 29.71 TFLOPS of FP16 (at a 1:1 ratio). The MI25 delivers 12.29 TFLOPS of FP32 but 24.58 TFLOPS of FP16, at a 2:1 ratio. So while the MI25 has some FP16 capability, the GB10 matches its FP32 and FP16 throughput, which is unusual and reflects a design that does not sacrifice FP32 for tensor work.

Memory configurations could not be more different. The GB10 uses 128 GB of LPDDR5X on a 256 bit bus, delivering 273.2 GB/s of bandwidth. The MI25 uses 16 GB of HBM2 on a 2048 bit bus, delivering 436.2 GB/s of bandwidth. The MI25 has a much narrower capacity but significantly higher bandwidth per byte, which is typical of HBM-based accelerators. The GB10’s 128 GB capacity is eight times larger, making it suitable for large models or datasets that cannot fit in the MI25’s memory. The MI25’s bandwidth advantage, however, means it may handle streaming workloads more efficiently per gigabyte.

The GB10 is an integrated graphics processor (IGP) with a 140 W TDP, no power connectors, and a suggested power supply of 300 W. It is a 150 mm by 51 mm card with a single HDMI output. The MI25 is a dual-slot card with a 300 W TDP, two 8-pin power connectors, and a suggested power supply of 700 W. It measures 267 mm by 111 mm and has no display outputs. The GB10 is clearly designed for compact, power-constrained systems, while the MI25 is a traditional high-power accelerator card.

The GB10 uses PCIe 5.0 x16, while the MI25 uses PCIe 3.0 x16. The GB10 also supports newer APIs, though its DirectX, OpenGL, and Vulkan support is listed as N/A, while the MI25 supports DirectX 12 (12_1), OpenGL 4.6, and Vulkan 1.3. The GB10 is an active production part, released in October 2025, with a predecessor in Server Hopper and a successor in Server Rubin. The MI25 is end-of-life, released in June 2017, with a predecessor in FirePro Data Center and no successor listed.

Head-to-Head Benchmarks

The only recorded head-to-head benchmark is Geekbench OpenCL, and the result is decisive. The NVIDIA GB10 scores 120,137, while the AMD Radeon Instinct MI25 scores 68,562. That is a 75.2% difference, meaning the GB10 completes OpenCL workloads nearly twice as fast as the MI25. This is not a marginal improvement; it is a generational leap.

To put the GB10’s score in context, its average benchmark score across all recorded tests is 117,393. Its nearest rival, the NVIDIA RTX 4000 SFF Ada Generation, scores 117,088, which is only 0.3% lower. The AMD Radeon PRO W7700 scores 118,976, which is 1.3% higher. The GB10 is essentially at parity with these modern workstation cards, despite being an integrated part with a 140 W TDP. It also beats the Tesla V100 SXM2 16 GB by 2.6% and the RTX A5500 Mobile by 3%, both of which are established accelerators with dedicated cooling and power.

The MI25’s average benchmark score is 68,562, and its nearest rivals are all within 2% of that figure. The Intel Arc A770 scores 68,809 (0.4% higher), the NVIDIA CMP 90HX scores 69,000 (0.6% higher), the AMD Radeon Pro WX 8200 scores 69,870 (1.9% higher), and the NVIDIA Quadro P6000 scores 69,986 (2% higher). The MI25 is firmly in that performance band, but that band is roughly half the performance of the GB10’s band.

The delta of 75.2% in OpenCL is the only recorded comparison, and it aligns with the raw specification differences. The GB10 has 2.4 times the FP32 throughput of the MI25, and its shading unit count is 50% higher. The MI25’s FP16 advantage is real but narrower, and in OpenCL workloads that typically stress FP32 or memory, the GB10 dominates. The MI25’s higher memory bandwidth does not translate into a benchmark win in the recorded data.

FAQ

Q: Which GPU has higher memory capacity?

A: The NVIDIA GB10 has 128 GB of LPDDR5X memory, while the AMD Radeon Instinct MI25 has 16 GB of HBM2. The GB10 offers eight times the capacity.

Q: Does the AMD MI25 support ray tracing or tensor operations?

A: No, the MI25 has no ray tracing cores and no tensor cores. The NVIDIA GB10 includes 48 ray tracing cores and 384 tensor cores.

Q: How do their power requirements compare?

A: The GB10 has a 140 W TDP and a suggested power supply of 300 W, and it requires no power connectors. The MI25 has a 300 W TDP and a suggested power supply of 700 W, requiring two 8-pin connectors.

Q: Which GPU is faster in the recorded OpenCL benchmark?

A: The NVIDIA GB10 scores 120,137, which is 75.2% higher than the MI25’s score of 68,562.

Q: What are the production statuses of these two GPUs?

A: The GB10 is listed as active production, released in October 2025. The MI25 is end-of-life, released in June 2017.

Q: Do both GPUs support modern graphics APIs?

A: The MI25 supports DirectX 12 (12_1), OpenGL 4.6, and Vulkan 1.3. The GB10 lists N/A for DirectX, OpenGL, and Vulkan, indicating it is not designed for traditional graphics API workloads.

Specification Differences

| Specification | NVIDIA GB10 | AMD Radeon Instinct MI25 |

|---|---|---|

| Architecture | Blackwell 2.0 | GCN 5.0 |

| Process Node | 5 nm | 14 nm |

| Foundry | TSMC | GlobalFoundries |

| Die Size | 382 mm² | 495 mm² |

| Transistors | Unknown | 12,500 million |

| Shading Units | 6,144 | 4,096 |

| TMUs | 384 | 256 |

| ROPs | 48 | 64 |

| Ray Tracing Cores | 48 | None |

| Tensor Cores | 384 | None |

| FP32 Performance | 29.71 TFLOPS | 12.29 TFLOPS |

| FP16 Performance | 29.71 TFLOPS (1:1) | 24.58 TFLOPS (2:1) |

| Memory Size | 128 GB | 16 GB |

| Memory Type | LPDDR5X | HBM2 |

| Memory Bus Width | 256 bit | 2048 bit |

| Memory Bandwidth | 273.2 GB/s | 436.2 GB/s |

| TDP | 140 W | 300 W |

| Slot Width | IGP | Dual-slot |

| Power Connectors | None | 2x 8-pin |

| Suggested PSU | 300 W | 700 W |

| Bus Interface | PCIe 5.0 x16 | PCIe 3.0 x16 |

| Display Outputs | 1x HDMI | No outputs |

| Release Date | October 2025 | June 2017 |

| Production Status | Active | End-of-life |

| Launch MSRP | 3,999 USD | Not listed |

The GB10 is smaller, newer, more power-efficient, and vastly more capable in raw compute. The MI25 has a wider memory bus and higher bandwidth, but it is older, larger, and consumes more than twice the power. The GB10 also includes hardware for ray tracing and tensor operations, which the MI25 lacks entirely.

The Verdict

The data is unambiguous: the NVIDIA GB10 is the superior accelerator in every recorded metric. It wins the only head-to-head benchmark by 75.2%, delivers 29.71 TFLOPS of FP32 against the MI25’s 12.29 TFLOPS, and does so at 140 W versus 300 W. The GB10 also offers 128 GB of memory, which is critical for modern workloads that require large model residency, while the MI25 is limited to 16 GB.

The MI25’s only advantages are memory bandwidth (436.2 GB/s versus 273.2 GB/s) and a wider bus (2048 bit versus 256 bit). In streaming workloads that are bandwidth-bound, the MI25 could theoretically perform better per byte of data moved, but the recorded OpenCL benchmark does not reflect any such advantage. The MI25 also supports traditional graphics APIs, while the GB10 lists N/A for DirectX, OpenGL, and Vulkan, so for display or graphics rendering tasks, the MI25 has a functional edge.

For buyers choosing between these two, the decision depends on the workload. If the task is compute-heavy, involves large datasets, or requires modern features like ray tracing or tensor cores, the GB10 is the only rational choice. Its performance percentile (95th) and active production status indicate it will remain relevant for years. The MI25, at the 90th percentile and end-of-life, is a legacy part. It may still serve in existing deployments or bandwidth-sensitive applications, but new purchases should favor the GB10. The launch MSRP of the GB10 is 3,999 USD, and it is available as an active product, while the MI25 has no listed launch MSRP and is no longer in production. The verdict is straightforward: the GB10 wins on performance, efficiency, capacity, and longevity.