AMD Instinct MI300X vs NVIDIA GeForce RTX 5090 SE Comparison

Head-to-Head Benchmarks

The available benchmark data paints an incomplete but still informative picture. The AMD Instinct MI300X has a recorded Geekbench OpenCL score of 317,994, while the NVIDIA GeForce RTX 5090 SE has no recorded benchmark scores in the database. This makes a direct numerical comparison impossible on the same test, but the MI300X's position relative to other accelerators gives context for where it stands.

The MI300X sits at the 100th percentile among all GPUs in the database, meaning no other recorded GPU scores higher in aggregate. Its average benchmark score of 317,994 places it ahead of the NVIDIA L40S by 7.5%, and ahead of the NVIDIA RTX 6000 Ada Generation by 10.7%. Against higher-end data center parts, it trails the NVIDIA H200 NVL by 5% and the NVIDIA B200 by 8%. These deltas position the MI300X in the upper tier of compute accelerators, competitive with NVIDIA's flagship data center offerings despite slight deficits to the newest parts.

The RTX 5090 SE has no recorded scores, so its percentile of 50 and average score of 0 reflect missing data rather than performance. The database cannot currently rank it against the MI300X or any other accelerator. Any performance assessment for the 5090 SE must rely on architectural specifications rather than measured results.

Architecture Differences

The two GPUs target fundamentally different workloads, and the architecture reflects that split. The AMD Instinct MI300X uses the CDNA 3.0 architecture on the Aqua Vanjaram chip, built for compute-intensive data center tasks. The NVIDIA GeForce RTX 5090 SE uses the Blackwell 2.0 architecture on the GB202 chip, designed for client graphics and gaming with added compute capability.

Both chips are fabricated on a 5 nm process at TSMC, but the MI300X is substantially larger. The MI300X die measures 1017 mm² with 153,000 million transistors, yielding a density of 150.4 million transistors per mm². The RTX 5090 SE die is 750 mm² with 92,200 million transistors, for a density of 122.9 million per mm². The MI300X packs roughly 66% more transistors into a 36% larger die.

Memory architecture diverges sharply. The MI300X uses 192 GB of HBM3 on an 8192-bit bus, delivering 5.32 TB/s of bandwidth. The RTX 5090 SE uses 24 GB of GDDR7 on a 384-bit bus, delivering 1.34 TB/s. The MI300X offers 8 times the capacity and roughly 4 times the bandwidth, which directly supports large model inference and training workloads. The 5090 SE's GDDR7 memory is faster per pin but far narrower in total throughput.

Compute resources differ in configuration. The MI300X has 19,456 shading units, 1,216 texture mapping units, and no ROPs, reflecting its compute-first design with zero pixel output. The RTX 5090 SE has 14,080 shading units, 440 TMUs, and 160 ROPs, enabling full graphics rasterization. The 5090 SE also includes 110 ray tracing cores and 440 tensor cores, while the MI300X lists no RT or tensor core counts in the database. The MI300X does not expose graphics APIs at all: DirectX, OpenGL, and Vulkan are all N/A. The RTX 5090 SE supports DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4.

Clock behavior differs as well. The MI300X runs a 1000 MHz base and 2100 MHz boost, while the RTX 5090 SE runs a 1740 MHz base and 2377 MHz boost. The 5090 SE boosts higher, but the MI300X compensates with far wider execution resources. Memory clocks show the MI300X at 1300 MHz (5.2 Gbps effective) versus the 5090 SE at 1750 MHz (28 Gbps effective), confirming the GDDR7 advantage in raw signal rate but not in aggregate bandwidth.

FAQ

Q: Which GPU has higher raw FP32 compute?

A: The AMD Instinct MI300X delivers 81.72 TFLOPS of FP32, while the NVIDIA GeForce RTX 5090 SE delivers 66.94 TFLOPS. The MI300X is roughly 22% ahead in this metric.

Q: How much memory does each GPU have, and what type?

A: The MI300X has 192 GB of HBM3 on an 8192-bit bus with 5.32 TB/s bandwidth. The RTX 5090 SE has 24 GB of GDDR7 on a 384-bit bus with 1.34 TB/s bandwidth.

Q: Can the MI300X output video to a display?

A: No. The MI300X has no display outputs and reports zero pixel rate. It is a compute-only accelerator. The RTX 5090 SE includes 1x HDMI 2.1b and 3x DisplayPort 2.1b outputs.

Q: What power delivery does each card require?

A: The MI300X is rated at 750 W TDP with no power connectors, as it uses an OAM module form factor and requires a suggested 1150 W PSU. The RTX 5090 SE is rated at 500 W TDP, uses a single 16-pin connector, and requires a suggested 900 W PSU.

Q: Which GPU supports ray tracing?

A: The RTX 5090 SE includes 110 RT cores and supports DirectX 12 Ultimate. The MI300X lists no RT cores and has no DirectX support, so it cannot execute ray tracing workloads.

Q: What is the release timing for each?

A: The MI300X released on December 5, 2023. The RTX 5090 SE has a release date of December 31, 2025, and is marked as Active in production status.

Specification Differences

The specification table below isolates the fields where the two GPUs differ. Shared traits include the 5 nm TSMC process node, PCIe 5.0 x16 bus interface, and 1:1 FP16 to FP32 ratio.

| Specification | AMD Instinct MI300X | NVIDIA GeForce RTX 5090 SE |

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

| Architecture | CDNA 3.0 | Blackwell 2.0 |

| Transistors | 153,000 million | 92,200 million |

| Die Size | 1017 mm² | 750 mm² |

| Transistor Density | 150.4M / mm² | 122.9M / mm² |

| Base Clock | 1000 MHz | 1740 MHz |

| Boost Clock | 2100 MHz | 2377 MHz |

| Memory Clock | 1300 MHz (5.2 Gbps effective) | 1750 MHz (28 Gbps effective) |

| Memory Size | 192 GB | 24 GB |

| Memory Type | HBM3 | GDDR7 |

| Memory Bus Width | 8192 bit | 384 bit |

| Memory Bandwidth | 5.32 TB/s | 1.34 TB/s |

| Shading Units | 19,456 | 14,080 |

| TMUs | 1,216 | 440 |

| ROPs | 0 | 160 |

| RT Cores | N/A | 110 |

| Tensor Cores | N/A | 440 |

| Pixel Rate | 0 MPixel/s | 380.3 GPixel/s |

| Texture Rate | 2,553.6 GTexel/s | 1,045.9 GTexel/s |

| FP32 | 81.72 TFLOPS | 66.94 TFLOPS |

| FP16 | 81.72 TFLOPS | 66.94 TFLOPS |

| TDP | 750 W | 500 W |

| Slot Width | OAM Module | Dual-slot |

| Power Connectors | None | 1x 16-pin |

| Suggested PSU | 1150 W | 900 W |

| Display Outputs | No outputs | 1x HDMI 2.1b, 3x DisplayPort 2.1b |

| DirectX | N/A | 12 Ultimate (12_2) |

| OpenGL | N/A | 4.6 |

| Vulkan | N/A | 1.4 |

| Dimensions | Not recorded | 267 mm x 111 mm x 40 mm |

| Release Date | 2023-12-05 | 2025-12-31 |

| Predecessor | Radeon Instinct | GeForce 40 |

| Successor | Not recorded | GeForce 60 |

| Launch MSRP | None | 1,499 USD |

Where Each One Wins

The AMD Instinct MI300X wins decisively in memory capacity, memory bandwidth, and raw compute throughput. Its 192 GB HBM3 pool with 5.32 TB/s bandwidth suits large-scale AI model training, inference for very large models, and scientific computing workloads that exceed the 24 GB ceiling of the RTX 5090 SE. The MI300X's 81.72 TFLOPS FP32 and FP16 output beats the 5090 SE's 66.94 TFLOPS in both precision levels. Its higher texture rate of 2,553.6 GTexel/s versus 1,045.9 GTexel/s reflects the broader execution pipeline, even though the MI300X lacks ROPs and pixel output entirely. The MI300X also leads in transistor count, die size, and shading unit count, confirming its role as a massive compute accelerator. Its benchmark percentile of 100 and average score of 317,994 demonstrate measured dominance in compute-oriented tests, with leads of 7.5% over the L40S and 10.7% over the RTX 6000 Ada Generation.

The NVIDIA GeForce RTX 5090 SE wins in everything related to graphics and client-side workloads. It has 160 ROPs and a pixel rate of 380.3 GPixel/s, enabling full rasterization output. It supports DirectX 12 Ultimate, OpenGL 4.6, and Vulkan 1.4, making it usable for gaming and professional graphics applications. The 110 RT cores enable hardware ray tracing, while the 440 tensor cores provide AI acceleration for features like DLSS and other client-side inference tasks. The 5090 SE runs at higher clocks: 1740 MHz base and 2377 MHz boost versus 1000 MHz and 2100 MHz for the MI300X. It uses GDDR7 memory running at 28 Gbps effective, a much faster signal rate than the MI300X's 5.2 Gbps HBM3, though the narrower 384-bit bus limits total bandwidth. The 5090 SE also draws less power at 500 W versus 750 W, fits in a standard dual-slot form factor, and includes display outputs. Its 24 GB VRAM is sufficient for many gaming and workstation tasks, and its launch MSRP is 1,499 USD, though pricing details are not the focus of this analysis.

The Verdict

The data indicates two accelerators built for separate markets with no overlap in intended use. The AMD Instinct MI300X is a data center compute accelerator with no display output, no graphics API support, and no ROPs. Its 192 GB HBM3 memory and 5.32 TB/s bandwidth make it a strong candidate for large-scale AI training and inference where memory capacity dominates. The recorded Geekbench OpenCL score of 317,994 places it at the 100th percentile among all GPUs, ahead of the L40S by 7.5% and the RTX 6000 Ada Generation by 10.7%, while trailing the H200 NVL by 5% and the B200 by 8%. These results confirm it competes at the top tier of compute accelerators.

The NVIDIA GeForce RTX 5090 SE is a client graphics card with full rasterization, ray tracing, and tensor core support. It cannot match the MI300X in memory capacity, bandwidth, or raw FP32 throughput, but it offers capabilities the MI300X lacks entirely: display outputs, graphics APIs, and a standard dual-slot form factor with a 16-pin power connector. Its 500 W TDP and 900 W suggested PSU make it easier to integrate into a conventional workstation than the MI300X's OAM module with its 1150 W PSU requirement.

For users running large-scale compute workloads that fit within the MI300X's 192 GB memory envelope, the MI300X is the clear choice based on measured benchmarks and memory bandwidth. For users needing a graphics-capable GPU with ray tracing, display output, and a lower power footprint, the RTX 5090 SE is the only viable option of the two. The absence of benchmark scores for the 5090 SE means its compute performance remains unquantified in the database, so any comparison of raw compute speed between the two relies on the MI300X's measured results and the 5090 SE's architectural specifications. The MI300X wins on raw compute and memory, the 5090 SE wins on graphics features and practical client integration. Choose based on workload, not on any single specification.