AMD Instinct MI300 vs NVIDIA GeForce RTX 5090 SE Comparison

FAQ

Q: What are the core architectural differences between the AMD Instinct MI300 and the NVIDIA GeForce RTX 5090 SE?

A: The MI300 uses AMD's CDNA 3.0 architecture on the Aqua Vanjaram chip, while the RTX 5090 SE uses NVIDIA's Blackwell 2.0 architecture on the GB202 chip. Both are built on TSMC's 5 nm process, but the MI300 has a larger die at 1017 mm² compared to 750 mm² for the RTX 5090 SE.

Q: How do memory configurations compare between these two cards?

A: The MI300 ships with 128 GB of HBM3 memory across an 8192-bit bus, yielding 5.32 TB/s of bandwidth. The RTX 5090 SE has 24 GB of GDDR7 memory on a 384-bit bus, delivering 1.34 TB/s. The MI300 has over 5x the capacity and nearly 4x the bandwidth.

Q: What are the transistor counts and densities for each chip?

A: The MI300 contains 153,000 million transistors on a 1017 mm² die, giving a density of 150.4M per mm². The RTX 5090 SE has 92,200 million transistors on a 750 mm² die, for a density of 122.9M per mm². The MI300 has both more total transistors and higher density.

Q: How do the shading unit counts compare?

A: Both cards have exactly 14,080 shading units. However, the MI300 has 880 texture mapping units and no ROPs, while the RTX 5090 SE has 440 texture mapping units and 160 ROPs.

Q: What are the power requirements for each card?

A: The MI300 has a TDP of 600 W with 2x 8-pin power connectors and a suggested PSU of 1000 W. The RTX 5090 SE has a TDP of 500 W with a single 16-pin connector and a suggested PSU of 900 W.

Q: What display outputs does each card offer?

A: The MI300 has no display outputs, making it a compute-only accelerator. The RTX 5090 SE provides 1x HDMI 2.1b and 3x DisplayPort 2.1b outputs, making it suitable for direct display connection.

Where Each One Wins

The AMD Instinct MI300 is built for compute-heavy workloads where memory capacity and bandwidth dominate. Its 128 GB HBM3 pool and 5.32 TB/s bandwidth provide an enormous advantage for large datasets that must reside on the GPU. The 8192-bit memory bus is unmatched in this comparison. The MI300's 880 texture mapping units also give it a lead in texture-heavy compute tasks, and its 1,496.0 GTexel/s texture rate far exceeds the competition. The card is designed for server and datacenter deployments, with its lack of display outputs confirming a pure compute focus.

The NVIDIA GeForce RTX 5090 SE wins in scenarios requiring rasterization, ray tracing, and graphics output. Its 160 ROPs and 380.3 GPixel/s pixel rate make it the clear choice for rendering frames. The 110 ray tracing cores and 440 tensor cores provide dedicated hardware for ray-traced workloads and AI acceleration. The card's 66.94 TFLOPS FP32 performance is significantly higher than the MI300's 47.87 TFLOPS, giving it a raw compute advantage in workloads that scale with FP32 throughput. The RTX 5090 SE also supports DirectX 12 Ultimate, OpenGL 4.6, and Vulkan 1.4, whereas the MI300 reports no API support for these graphics standards.

Architecture Differences

The MI300 uses AMD's CDNA 3.0 architecture, which is optimized for compute acceleration rather than graphics rendering. The architecture prioritizes memory bandwidth and FP32 throughput for scientific and AI workloads. The chip, codenamed Aqua Vanjaram, features a massive 1017 mm² die with 153,000 million transistors. The architecture includes 14,080 shading units and 880 TMUs but omits ROPs entirely, with a pixel rate of 0 MPixel/s. This confirms the design goal is data processing, not image output.

The RTX 5090 SE uses NVIDIA's Blackwell 2.0 architecture on the GB202 chip. This architecture balances compute and graphics capabilities. It includes 14,080 shading units, 440 TMUs, and 160 ROPs, providing full rasterization capability. The architecture also integrates 110 ray tracing cores and 440 tensor cores, which are absent from the MI300's specification sheet. The Blackwell 2.0 design supports the full modern graphics API stack, including DirectX 12 Ultimate, OpenGL 4.6, and Vulkan 1.4. The 750 mm² die with 92,200 million transistors is smaller than the MI300 but still substantial.

Both chips are manufactured on TSMC's 5 nm process, but the MI300 achieves a higher transistor density at 150.4M per mm² versus 122.9M per mm² for the RTX 5090 SE. The MI300's larger die and higher density result in a significantly higher total transistor count. The clock strategies differ as well: the MI300 runs at a 1000 MHz base and 1700 MHz boost, while the RTX 5090 SE operates at 1740 MHz base and 2377 MHz boost, giving the NVIDIA card a substantial clock advantage.

Specification Differences

The two cards differ across nearly every major specification category. The MI300 has 153,000 million transistors versus 92,200 million for the RTX 5090 SE. Die size is 1017 mm² versus 750 mm². Transistor density is 150.4M per mm² versus 122.9M per mm².

Clock speeds show the RTX 5090 SE at a significant advantage: 1740 MHz base versus 1000 MHz, and 2377 MHz boost versus 1700 MHz. Memory clocks also differ, with the MI300 at 1300 MHz (5.2 Gbps effective) and the RTX 5090 SE at 1750 MHz (28 Gbps effective).

Memory configuration is a major differentiator. The MI300 has 128 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.

The shading unit count is identical at 14,080, but TMU counts differ: 880 for the MI300 versus 440 for the RTX 5090 SE. The MI300 has 0 ROPs while the RTX 5090 SE has 160. The RTX 5090 SE adds 110 ray tracing cores and 440 tensor cores, which have no equivalent on the MI300.

Pixel rate is 0 MPixel/s for the MI300 versus 380.3 GPixel/s for the RTX 5090 SE. Texture rate is 1,496.0 GTexel/s versus 1,045.9 GTexel/s. FP32 performance is 47.87 TFLOPS versus 66.94 TFLOPS. FP16 performance mirrors FP32 for both cards.

Power specifications differ: TDP is 600 W for the MI300 and 500 W for the RTX 5090 SE. Power connectors are 2x 8-pin versus 1x 16-pin. Suggested PSU is 1000 W versus 900 W. The RTX 5090 SE is dual-slot and has display outputs, while the MI300 has none. Both cards share the same length and height at 267 mm and 111 mm, but only the RTX 5090 SE has a listed width of 40 mm.

The API support differs completely: the MI300 lists N/A for DirectX, OpenGL, and Vulkan, while the RTX 5090 SE supports DirectX 12 Ultimate, OpenGL 4.6, and Vulkan 1.4. The RTX 5090 SE has an active production status and a launch MSRP of 1,499 USD, while the MI300 has no listed production status or launch MSRP.

Head-to-Head Benchmarks

The recorded data shows no direct head-to-head benchmark results between these two cards. Both have zero benchmark entries and zero wins in the comparison. The percentile versus all GPUs is identical at 50 for both, and the average benchmark score is 0 for both. The nearest rivals lists are empty for both cards.

Without direct benchmark data, the specification differences provide the clearest basis for comparison. The MI300 leads in memory capacity by a factor of 5.33, offering 128 GB versus 24 GB. Its memory bandwidth of 5.32 TB/s is 3.97x higher than the RTX 5090 SE's 1.34 TB/s. The memory bus width of 8192 bits versus 384 bits gives the MI300 a 21.3x advantage in bus width. The MI300 also has 2x the TMUs at 880 versus 440, and its texture rate of 1,496.0 GTexel/s is 43% higher than the 1,045.9 GTexel/s of the RTX 5090 SE.

The RTX 5090 SE leads in several compute metrics. Its FP32 throughput of 66.94 TFLOPS is 40% higher than the MI300's 47.87 TFLOPS. The clock advantage is substantial: the boost clock of 2377 MHz is 40% higher than the MI300's 1700 MHz. The RTX 5090 SE delivers 380.3 GPixel/s of pixel throughput while the MI300 delivers none. The RTX 5090 SE also has 160 ROPs, 110 ray tracing cores, and 440 tensor cores, all of which are absent from the MI300's specifications.

The power efficiency favors the RTX 5090 SE based on the recorded data. The RTX 5090 SE delivers 66.94 TFLOPS at 500 W TDP, while the MI300 delivers 47.87 TFLOPS at 600 W TDP. The RTX 5090 SE achieves 0.13 TFLOPS per watt, while the MI300 achieves 0.08 TFLOPS per watt. The RTX 5090 SE also has a lower suggested PSU at 900 W versus 1000 W.

The Verdict

The AMD Instinct MI300 is the choice for memory-bound compute workloads. Its 128 GB HBM3 pool and 5.32 TB/s bandwidth are decisive advantages for large-scale AI training, scientific simulation, and data processing tasks where dataset size exceeds the 24 GB capacity of the RTX 5090 SE. The 8192-bit memory bus and 880 TMUs support high-throughput data movement and texture-heavy compute. The MI300's higher transistor count and density indicate a design focused on brute-force parallel compute, with the larger die and 153,000 million transistors providing substantial raw resources for compute-heavy applications. The absence of display outputs and graphics API support confirms this is a datacenter accelerator, not a workstation graphics card.

The NVIDIA GeForce RTX 5090 SE is the choice for graphics, ray tracing, and general-purpose compute where FP32 throughput matters. Its 66.94 TFLOPS FP32 performance is 40% higher than the MI300, providing a clear advantage in compute workloads that scale with FP32 throughput. The 160 ROPs, 110 ray tracing cores, and 440 tensor cores enable features the MI300 cannot match, including DirectX 12 Ultimate, OpenGL 4.6, and Vulkan 1.4 support. The 380.3 GPixel/s pixel rate enables real-time rendering, while the display outputs allow direct connection to monitors. The RTX 5090 SE also operates more efficiently, delivering higher performance at a lower 500 W TDP with a 900 W suggested PSU. Its 1740 MHz base and 2377 MHz boost clocks are substantially higher than the MI300's 1000 MHz and 1700 MHz, contributing to its compute advantage.

The data indicates these cards target different use cases. The MI300 prioritizes memory capacity and bandwidth for workloads that cannot fit in smaller memory pools. The RTX 5090 SE prioritizes raw FP32 compute, graphics features, and efficiency. Users with massive datasets that require on-GPU storage should select the MI300. Users needing graphics output, ray tracing, or maximum FP32 throughput should select the RTX 5090 SE. Both cards share the same shading unit count at 14,080 and identical physical length and height, but their architectural priorities diverge completely from that point.