AMD Radeon Instinct MI60 vs NVIDIA GeForce RTX 5090 Comparison

FAQ

Q: How does the AMD Radeon Instinct MI60 compare to the NVIDIA GeForce RTX 5090 in raw compute benchmarks?

A: The RTX 5090 wins both head-to-head compute tests decisively. It scores 334,370 in Geekbench OpenCL versus 92,488 for the MI60 (a -72.3% delta for AMD), and 376,728 in Geekbench Vulkan versus 92,444 (-75.5% delta). The NVIDIA card is roughly 3.6x faster in OpenCL and over 4x faster in Vulkan.

Q: What is the performance percentile ranking for each card?

A: The MI60 sits at the 93rd percentile among all GPUs, while the RTX 5090 ranks at the 92nd. Despite the RTX 5090’s massive benchmark lead, its average benchmark score of 79,842 is dragged down by its Passmark results, whereas the MI60’s average of 92,466 is based solely on its two high Geekbench scores.

Q: Which card has more memory bandwidth?

A: The RTX 5090 offers 1.79 TB/s of bandwidth via its 512-bit GDDR7 interface, compared to the MI60’s 1.02 TB/s over a 4096-bit HBM2 bus. NVIDIA’s advantage here is 75% higher bandwidth despite the MI60’s wider memory bus.

Q: What are the transistor and die size differences?

A: The RTX 5090 packs 92,200 million transistors on a 750 mm² die (122.9M transistors per mm²), while the MI60 has 13,230 million transistors on a 331 mm² die (40.0M per mm²). The RTX 5090 uses a 5 nm process versus 7 nm for the MI60.

Q: What is the power draw comparison?

A: The RTX 5090 has a TDP of 575 W with a suggested 950 W PSU, while the MI60 draws 300 W with a suggested 700 W PSU. NVIDIA consumes nearly double the power but delivers over 7x the FP32 throughput (104.8 TFLOPS vs 14.75 TFLOPS).

Q: Which card supports newer APIs?

A: The RTX 5090 supports DirectX 12 Ultimate (12_2) and Vulkan 1.4, while the MI60 supports DirectX 12 (12_1) and Vulkan 1.3. Both cards support OpenGL 4.6.

Where Each One Wins

The RTX 5090 wins every head-to-head benchmark in this comparison, and the wins are not marginal. In Geekbench OpenCL, the NVIDIA card scores 334,370 versus 92,488 for the MI60, a 241,882-point gap. In Geekbench Vulkan, the RTX 5090 scores 376,728 versus 92,444, a 284,284-point gap. There are zero benchmark wins for the MI60 in direct comparison.

However, the MI60 has its own territory where it excels relative to its peers. Its average benchmark score of 92,466 places it 0.9% ahead of the NVIDIA RTX A4500 (91,671) and 1.5% ahead of the RTX A4500 Mobile (91,134). The MI60 trails the AMD Radeon Pro VII (97,131) by 4.8% and the AMD Radeon RX 7900M (97,487) by 5.2%. In contrast, the RTX 5090’s average score of 79,842 is only 0.3% ahead of the NVIDIA Tesla P100 PCIe 16 GB (79,605) and 0.6% ahead of the Tesla P100 PCIe 12 GB (79,396), while trailing the AMD Radeon Pro Vega 64X (80,959) by 1.4%.

For the MI60, the win condition is efficiency per watt in a datacenter context. It delivers 14.75 TFLOPS FP32 at 300 W, which is 49.2 GFLOPS per watt, versus the RTX 5090’s 104.8 TFLOPS at 575 W, or 182.3 GFLOPS per watt. The MI60’s compute density per watt is actually lower, but its absolute power envelope is far smaller, making it suitable for power-constrained rack deployments.

Architecture Differences

The MI60 is built on AMD’s GCN 5.1 architecture with the Vega 20 chip, fabricated on a 7 nm process at TSMC. It uses 13,230 million transistors across a 331 mm² die, yielding a transistor density of 40.0M per mm². The RTX 5090 uses NVIDIA’s Blackwell 2.0 architecture with the GB202 chip, also from TSMC but on a 5 nm node. It packs 92,200 million transistors into a 750 mm² die, achieving 122.9M transistors per mm² — roughly three times the density.

The MI60 has 4,096 shading units, 256 TMUs, and 64 ROPs. It has no dedicated ray tracing cores or tensor cores. The RTX 5090 features 21,760 shading units, 680 TMUs, and 176 ROPs, plus 170 RT cores and 680 tensor cores. The NVIDIA card’s FP16 throughput is 104.8 TFLOPS at a 1:1 ratio with FP32, whereas the MI60’s FP16 is 29.49 TFLOPS at a 2:1 ratio, meaning it achieves half-rate FP16.

Memory architecture differs fundamentally. The MI60 uses 32 GB of HBM2 on a 4096-bit bus, delivering 1.02 TB/s. The RTX 5090 uses 32 GB of GDDR7 on a 512-bit bus, delivering 1.79 TB/s. The MI60’s HBM2 offers lower per-pin bandwidth but a much wider interface, while GDDR7 achieves higher total bandwidth with fewer pins.

The MI60 supports PCIe 4.0 x16, while the RTX 5090 uses PCIe 5.0 x16. Display outputs differ: the MI60 has a single mini-DisplayPort 1.4a, while the RTX 5090 offers 1x HDMI 2.1b and 3x DisplayPort 2.1b. The MI60 uses a 1x 6-pin + 1x 8-pin power configuration, whereas the RTX 5090 uses a single 16-pin connector.

Specification Differences

| Specification | AMD Radeon Instinct MI60 | NVIDIA GeForce RTX 5090 |

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

| Architecture | GCN 5.1 | Blackwell 2.0 |

| Process Node | 7 nm | 5 nm |

| Transistors | 13,230 million | 92,200 million |

| Die Size | 331 mm² | 750 mm² |

| Transistor Density | 40.0M / mm² | 122.9M / mm² |

| Base Clock | 1200 MHz | 2017 MHz |

| Boost Clock | 1800 MHz | 2407 MHz |

| Memory Type | HBM2 | GDDR7 |

| Memory Bus | 4096 bit | 512 bit |

| Memory Bandwidth | 1.02 TB/s | 1.79 TB/s |

| Shading Units | 4096 | 21760 |

| TMUs | 256 | 680 |

| ROPs | 64 | 176 |

| RT Cores | — | 170 |

| Tensor Cores | — | 680 |

| Pixel Rate | 115.2 GPixel/s | 423.6 GPixel/s |

| Texture Rate | 460.8 GTexel/s | 1,636.8 GTexel/s |

| FP32 | 14.75 TFLOPS | 104.8 TFLOPS |

| FP16 | 29.49 TFLOPS (2:1) | 104.8 TFLOPS (1:1) |

| TDP | 300 W | 575 W |

| Power Connectors | 1x 6-pin + 1x 8-pin | 1x 16-pin |

| Suggested PSU | 700 W | 950 W |

| Bus Interface | PCIe 4.0 x16 | PCIe 5.0 x16 |

| DirectX | 12 (12_1) | 12 Ultimate (12_2) |

| Vulkan | 1.3 | 1.4 |

| Release Date | 2018-11-17 | 2025-01-29 |

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

Head-to-Head Benchmarks

The two available head-to-head benchmarks both favor the RTX 5090 overwhelmingly. In Geekbench OpenCL, the RTX 5090 scores 334,370 against the MI60’s 92,488. The delta of -72.3% for the MI60 means the NVIDIA card is 3.61x faster. In Geekbench Vulkan, the RTX 5090 scores 376,728 versus 92,444, a -75.5% delta, meaning it is 4.08x faster.

These results align with the raw compute specifications. The RTX 5090’s FP32 throughput of 104.8 TFLOPS is 7.1x the MI60’s 14.75 TFLOPS. Its texture rate of 1,636.8 GTexel/s is 3.55x the MI60’s 460.8 GTexel/s, and its pixel rate of 423.6 GPixel/s is 3.68x the MI60’s 115.2 GPixel/s.

The benchmark gap is larger than the FP32 ratio would suggest, particularly in Vulkan where the RTX 5090’s lead extends to 4.08x. This likely reflects architectural advantages beyond raw shader count, such as the RTX 5090’s dedicated tensor cores and RT cores, which can accelerate certain compute workloads. The MI60 has no equivalent hardware.

The Verdict

Choose the RTX 5090 for any workload where raw compute performance is the primary criterion. It wins both head-to-head benchmarks by margins of 72-76%, delivers 104.8 TFLOPS FP32 versus 14.75 TFLOPS, and offers 1.79 TB/s of memory bandwidth versus 1.02 TB/s. It also brings hardware ray tracing and tensor cores, supports DirectX 12 Ultimate and Vulkan 1.4, and is an active production product with a successor already announced.

Choose the MI60 if you are constrained by power or form factor. At 300 W TDP with a 700 W suggested PSU, it draws nearly half the power of the RTX 5090’s 575 W TDP and 950 W suggested PSU. Its 267 mm length makes it shorter than the RTX 5090’s 304 mm. It also uses a standard 6-pin + 8-pin power configuration rather than the newer 16-pin connector, which may be easier to accommodate in existing infrastructure.

The MI60’s end-of-life status and 2018 release date mean it is not a forward-looking investment. Its GCN 5.1 architecture lacks the dedicated RT and tensor hardware of the Blackwell 2.0 chip. However, its 93rd percentile ranking versus the RTX 5090’s 92nd percentile shows that its compute efficiency per benchmark point remains competitive, even if the absolute performance is far lower.

For datacenter deployments requiring massive parallel throughput, the RTX 5090 is the clear choice. For legacy installations with power budgets and older power connectors, the MI60 remains a functional option, but the data shows it is outclassed by a factor of 3.6x to 4.1x in compute benchmarks.