AMD Instinct MI100 vs NVIDIA GeForce RTX 5090 D Comparison

The Verdict

The data records only a single shared benchmark between these two accelerators, but that result is decisive. In Geekbench OpenCL, the NVIDIA GeForce RTX 5090 D scores 310674 against the AMD Instinct MI100's 139035, a 55.2% margin in NVIDIA's favor. The MI100's single recorded score of 139035 places it in the 96th percentile of all GPUs, while the RTX 5090 D's average benchmark score of 77712 sits in the 92nd percentile. However, that percentile comparison is misleading, as the RTX 5090 D's average includes many lower-scoring DirectX and PassMark tests, whereas the MI100's average is derived from only its OpenCL result.

For compute-heavy OpenCL workloads, the RTX 5090 D is the clear choice based on head-to-head data. Its 104.8 TFLOPS FP32 and 104.8 TFLOPS FP16 (1:1) throughput dwarf the MI100's 23.07 TFLOPS FP32 and 46.14 TFLOPS FP16 (2:1). The MI100 remains relevant only in legacy CDNA 1.0 deployments where its 32 GB of HBM2 with 1.23 TB/s bandwidth and 4096-bit bus are already integrated into existing infrastructure. For any new deployment, the RTX 5090 D's 32 GB of GDDR7 at 1.79 TB/s bandwidth, 512-bit bus, and 170 RT cores make it the more versatile and future-proof accelerator. The MI100 is end-of-life with no display outputs and no DirectX, OpenGL, or Vulkan support, limiting it to headless compute. The RTX 5090 D, by contrast, is active, supports DirectX 12 Ultimate, OpenGL 4.6, and Vulkan 1.4, and includes display outputs.

Where Each One Wins

The RTX 5090 D wins the only direct comparison available: Geekbench OpenCL, where it leads by 55.2%. This is consistent with its raw specifications, which show advantages across every measurable compute category. Its FP32 throughput of 104.8 TFLOPS is 4.5 times the MI100's 23.07 TFLOPS. Its FP16 throughput matches its FP32 at 104.8 TFLOPS, while the MI100's FP16 is only 2:1 ratio, delivering 46.14 TFLOPS. The RTX 5090 D also leads in texture rate (1,636.8 GTexel/s vs 721.0 GTexel/s) and pixel rate (423.6 GPixel/s vs 96.13 GPixel/s).

The MI100 wins in memory latency-sensitive workloads that favor HBM2 over GDDR7, though the data does not include a direct test. Its 4096-bit memory bus is double the RTX 5090 D's 512-bit bus, and its HBM2 memory operates at 1200 MHz with 2.4 Gbps effective speed. The RTX 5090 D's GDDR7 runs at 1750 MHz with 28 Gbps effective speed, yielding higher total bandwidth (1.79 TB/s vs 1.23 TB/s), but the MI100's wider bus can be advantageous for certain access patterns. The MI100 also has a lower power draw at 300 W versus 575 W, and its suggested PSU is 700 W versus 950 W, making it easier to integrate into power-constrained systems.

Architecture Differences

The two accelerators represent fundamentally different design philosophies separated by five years of silicon evolution. The MI100 uses AMD's CDNA 1.0 architecture on the Arcturus chip, built on TSMC's 7 nm process with 25,600 million transistors on a 750 mm² die. The RTX 5090 D uses NVIDIA's Blackwell 2.0 architecture on the GB202 chip, built on TSMC's 5 nm process with 92,200 million transistors, also on a 750 mm² die. The transistor density tells the story: the RTX 5090 D packs 122.9 million transistors per mm² versus the MI100's 34.1 million per mm², a 3.6 times density advantage from the newer node.

The MI100 has 7,680 shading units, 480 texture mapping units, and 64 raster output units. It has no dedicated RT cores and no tensor cores, reflecting its compute-focused CDNA 1.0 lineage. The RTX 5090 D has 21,760 shading units, 680 TMUs, and 176 ROPs, plus 170 RT cores and 680 tensor cores. Clock speeds also differ substantially: the MI100 runs at 1000 MHz base and 1502 MHz boost, while the RTX 5090 D runs at 2017 MHz base and 2407 MHz boost.

Memory architectures diverge sharply. The MI100 uses 32 GB of HBM2 with a 4096-bit bus and 1.23 TB/s bandwidth. The RTX 5090 D uses 32 GB of GDDR7 with a 512-bit bus and 1.79 TB/s bandwidth. Both have the same 32 GB capacity, but the newer GDDR7 standard delivers 45.5% more bandwidth despite the narrower bus.

Feature support is another major differentiator. The MI100 has no display outputs, no DirectX, no OpenGL, and no Vulkan support, making it a pure compute accelerator. The RTX 5090 D includes one HDMI 2.1b and three DisplayPort 2.1b outputs, supports DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4. The RTX 5090 D also uses PCIe 5.0 x16 versus the MI100's PCIe 4.0 x16, doubling the host interface bandwidth. Power delivery differs as well: the MI100 uses two 8-pin connectors at 300 W TDP, while the RTX 5090 D uses a single 16-pin connector at 575 W TDP.

FAQ

Q: Which GPU has higher FP32 compute performance?

A: The NVIDIA GeForce RTX 5090 D delivers 104.8 TFLOPS FP32, compared to the AMD Instinct MI100's 23.07 TFLOPS, a 4.5 times advantage.

Q: Do both cards have the same memory capacity?

A: Yes, both have 32 GB, but the RTX 5090 D uses GDDR7 with 1.79 TB/s bandwidth, while the MI100 uses HBM2 with 1.23 TB/s bandwidth.

Q: Can the MI100 be used for graphics output?

A: No, the MI100 has no display outputs and supports no DirectX, OpenGL, or Vulkan APIs, making it headless compute only.

Q: What is the power requirement difference?

A: The MI100 has a 300 W TDP with a 700 W suggested PSU and two 8-pin connectors. The RTX 5090 D has a 575 W TDP with a 950 W suggested PSU and one 16-pin connector.

Q: Which card has ray tracing capabilities?

A: Only the RTX 5090 D, which includes 170 RT cores. The MI100 has no RT cores listed.

Q: How does the RTX 5090 D compare in OpenCL performance?

A: The RTX 5090 D scores 310674 in Geekbench OpenCL, which is 55.2% higher than the MI100's 139035.

Head-to-Head Benchmarks

The database contains one direct comparison: Geekbench OpenCL. The RTX 5090 D scores 310674, while the MI100 scores 139035. The delta is 55.2% in NVIDIA's favor, meaning the MI100 achieves only 44.8% of the RTX 5090 D's OpenCL score. This result is consistent with the broader specification gap between the two architectures.

Looking at the nearest rivals for each card provides context. The MI100's 139035 OpenCL score places it just 0.7% ahead of the NVIDIA Tesla V100 PCIe 16 GB (138063), 0.9% ahead of the Tesla V100 SXM2 32 GB (137731), 1.9% ahead of the AMD Radeon PRO V620 (136472), and 2.4% ahead of the AMD Radeon Pro W6800X Duo (135774). These are all narrow margins, indicating the MI100 was competitive within its contemporary compute GPU class.

The RTX 5090 D's average benchmark score of 77712 tells a different story when compared to its nearest rivals. It sits 1.1% ahead of the AMD Radeon RX 6650M XT (76904), 1.6% behind the AMD Radeon RX 6850M XT (78940), 2.1% behind the NVIDIA Tesla P100 PCIe 12 GB (79396), and 2.4% behind the NVIDIA Tesla P100 PCIe 16 GB (79605). This average includes many low-scoring legacy DirectX tests (PassMark DirectX 9 at 434, DirectX 10 at 231, DirectX 11 at 371, DirectX 12 at 219), which drag down the average despite strong modern scores like Geekbench Vulkan at 376915 and PassMark G3D at 44065.

The individual benchmark suite for the RTX 5090 D reveals its strengths. Its Geekbench Vulkan score of 376915 is 21.3% higher than its own OpenCL score of 310674, indicating strong Vulkan performance. PassMark GPU Compute shows 28396, while PassMark G2D shows 1487. The 3DMark Steel Nomad DX12 test reports 14326. These results paint a picture of a card that excels in modern APIs and compute workloads, while its legacy DirectX scores (all below 434) suggest driver overhead in older APIs.

For the MI100, the single OpenCL score of 139035 is its only recorded benchmark. Its 96th percentile ranking among all GPUs reflects that this score is strong relative to the entire database, but the RTX 5090 D's 92nd percentile with a much lower average score demonstrates how averaging across many tests can obscure top-tier performance. The head-to-head result is unambiguous: in the one test both cards share, the RTX 5090 D wins decisively by 55.2%.