AMD Instinct MI350X vs NVIDIA GeForce RTX 5090 D V2 Comparison

FAQ

Q: What is the core architectural difference between the AMD Instinct MI350X and the NVIDIA GeForce RTX 5090 D V2?

A: The MI350X uses AMD's CDNA 4.0 architecture on a 3 nm TSMC process, while the RTX 5090 D V2 uses NVIDIA's Blackwell 2.0 architecture on a 5 nm TSMC process. The MI350X is built around the MI350 256CU chip, whereas the RTX 5090 D V2 uses the GB202 chip.

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

A: The MI350X has 288 GB of HBM3e memory on an 8192-bit bus with 8.19 TB/s bandwidth, while the RTX 5090 D V2 has 24 GB of GDDR7 memory on a 384-bit bus with 1.34 TB/s bandwidth. The MI350X offers 12 times the memory capacity and roughly six times the bandwidth.

Q: Which card has higher FP32 compute throughput?

A: The RTX 5090 D V2 delivers 104.8 TFLOPS of FP32 performance, which is 32.71 TFLOPS higher than the MI350X's 72.09 TFLOPS. The NVIDIA card also matches this figure for FP16 at 104.8 TFLOPS (1:1), while the MI350X delivers 72.09 TFLOPS FP16 (1:1).

Q: What are the power requirements for each card?

A: The MI350X has a 1000 W TDP and requires a 1400 W suggested PSU, while the RTX 5090 D V2 has a 575 W TDP and requires a 950 W suggested PSU. The MI350X uses no power connectors (OAM Module slot width), whereas the RTX 5090 D V2 uses a single 16-pin connector.

Q: What is the percentile ranking for each card in the database?

A: The RTX 5090 D V2 sits at the 59th percentile among all GPUs with an average benchmark score of 16504, while the MI350X is at the 50th percentile with an average benchmark score of 0. The NVIDIA card has recorded benchmark data, while the AMD card has no benchmark entries in the database.

Q: When were these cards released?

A: The MI350X launched on 2025-06-11, and the RTX 5090 D V2 launched on 2025-08-14. The RTX 5090 D V2's production status is listed as Active, while the MI350X has no production status recorded.

The Verdict

The data presents two fundamentally different products. The AMD Instinct MI350X is a compute-oriented accelerator with 288 GB of HBM3e memory, an 8192-bit bus, and 8.19 TB/s of bandwidth, built for workloads that demand massive memory capacity. It has no display outputs, no DirectX, OpenGL, or Vulkan API support, and no pixel rate. The NVIDIA GeForce RTX 5090 D V2 is a consumer graphics card with 24 GB of GDDR7, 170 RT cores, 680 tensor cores, full API support including DirectX 12 Ultimate, and a measured 3DMark Steel Nomad DX12 score of 16504.

The RTX 5090 D V2 is the only card with recorded benchmark results, placing it at the 59th percentile. Its nearest rivals in the database are the NVIDIA T400 at 16508 (0% delta), AMD Radeon PRO W7500 at 16415 (0.5% delta), NVIDIA RTX PRO 6000 Blackwell at 16408 (0.6% delta), and AMD Radeon RX 5700 XT at 16361 (0.9% delta). The MI350X has no benchmark scores, no nearest rivals, and an average benchmark score of zero.

For a builder choosing between these two, the decision rests entirely on workload. The MI350X is for memory-bound compute tasks that can use 288 GB of HBM3e. The RTX 5090 D V2 is for rendering, gaming, and general GPU compute where its 104.8 TFLOPS FP32 throughput and API support matter. The MI350X has a 1000 W TDP and requires a 1400 W PSU; the RTX 5090 D V2 draws 575 W and requires a 950 W PSU. The launch MSRP for the RTX 5090 D V2 is 2,299 USD. The MI350X has no launch MSRP recorded.

Head-to-Head Benchmarks

No head-to-head benchmark entries exist in the database for these two cards. The only recorded benchmark belongs to the RTX 5090 D V2: 3DMark Steel Nomad DX12 with a score of 16504. This places the NVIDIA card at the 59th percentile among all GPUs. The MI350X has no benchmark scores at all, so direct performance comparisons cannot be made from measured data.

The RTX 5090 D V2's nearest rival comparisons show a tight cluster of scores. The NVIDIA T400 leads with 16508, a 0% delta. The AMD Radeon PRO W7500 scores 16415, which is 0.5% behind. The NVIDIA RTX PRO 6000 Blackwell scores 16408, 0.6% behind. The AMD Radeon RX 5700 XT scores 16361, 0.9% behind. These deltas indicate the RTX 5090 D V2 sits within a narrow performance band relative to these cards in the Steel Nomad workload.

The MI350X's 72.09 TFLOPS FP32 is 32.71 TFLOPS lower than the RTX 5090 D V2's 104.8 TFLOPS. However, the MI350X counters with 8.19 TB/s memory bandwidth versus 1.34 TB/s, a difference of 6.85 TB/s in favor of the AMD card. The MI350X also has 288 GB of memory versus 24 GB, a 264 GB advantage. These figures indicate the MI350X is designed for memory capacity and bandwidth, while the RTX 5090 D V2 prioritizes raw compute throughput.

Specification Differences

The two cards differ across nearly every major specification category.

Process and die: The MI350X uses a 3 nm process with 185,000 million transistors on a 2380 mm² die, giving a transistor density of 77.7M per mm². The RTX 5090 D V2 uses a 5 nm process with 92,200 million transistors on a 750 mm² die, giving a transistor density of 122.9M per mm². The AMD chip has 92,800 million more transistors but a lower density due to its much larger die.

Clocks: The MI350X has a 1000 MHz base clock and 2200 MHz boost clock. The RTX 5090 D V2 has a 2017 MHz base clock and 2407 MHz boost clock, with a 1017 MHz higher base and 207 MHz higher boost.

Memory: The MI350X has 288 GB HBM3e on an 8192-bit bus with 8.19 TB/s bandwidth and 2000 MHz memory clock (8 Gbps effective). The RTX 5090 D V2 has 24 GB GDDR7 on a 384-bit bus with 1.34 TB/s bandwidth and 1750 MHz memory clock (28 Gbps effective).

Compute units: The MI350X has 16384 shading units, 1024 TMUs, and 0 ROPs. The RTX 5090 D V2 has 21760 shading units, 680 TMUs, and 176 ROPs. The NVIDIA card has 5376 more shading units but 344 fewer TMUs.

Ray tracing and tensor cores: The MI350X has no RT cores and no tensor cores listed. The RTX 5090 D V2 has 170 RT cores and 680 tensor cores.

Rates: The MI350X has a pixel rate of 0 MPixel/s and a texture rate of 2,252.8 GTexel/s. The RTX 5090 D V2 has a pixel rate of 423.6 GPixel/s and a texture rate of 1,636.8 GTexel/s.

Power and cooling: The MI350X has a 1000 W TDP, OAM Module slot width, no power connectors, and a 1400 W suggested PSU. The RTX 5090 D V2 has a 575 W TDP, dual-slot width, one 16-pin connector, and a 950 W suggested PSU.

Physical dimensions: The MI350X is 102 mm long and 165 mm wide. The RTX 5090 D V2 is 304 mm long, 137 mm high, and 48 mm wide.

Display and API support: The MI350X has no display outputs and no DirectX, OpenGL, or Vulkan support. The RTX 5090 D V2 has 1x HDMI 2.1b and 3x DisplayPort 2.1b outputs, with DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4 support.

Bus interface: Both use PCIe 5.0 x16.

Architecture Differences

The MI350X is built on CDNA 4.0, AMD's compute-focused architecture designed for data center and AI workloads. The RTX 5090 D V2 uses Blackwell 2.0, NVIDIA's architecture for consumer graphics. The MI350X's MI350 256CU chip is a 2380 mm² die fabricated on TSMC's 3 nm process. The RTX 5090 D V2's GB202 chip is a 750 mm² die on TSMC's 5 nm process.

The memory architectures reflect their different purposes. The MI350X uses HBM3e with an 8192-bit bus, which explains its 8.19 TB/s bandwidth and 288 GB capacity. This configuration suits workloads that hold very large datasets in memory. The RTX 5090 D V2 uses GDDR7 on a 384-bit bus, prioritizing lower latency and consumer-friendly packaging at 24 GB.

Compute resources differ sharply. The MI350X has 16384 shading units and 1024 TMUs but zero ROPs, no RT cores, and no tensor cores listed. The RTX 5090 D V2 has 21760 shading units, 680 TMUs, 176 ROPs, 170 RT cores, and 680 tensor cores. The absence of ROPs and display outputs on the MI350X confirms it is not intended for graphical output. The RTX 5090 D V2's RT and tensor cores support ray tracing and AI-accelerated features.

The MI350X has no API support for DirectX, OpenGL, or Vulkan, which reinforces its compute-only design. The RTX 5090 D V2 supports DirectX 12 Ultimate, OpenGL 4.6, and Vulkan 1.4, making it compatible with consumer software stacks.

Power delivery also reflects their roles. The MI350X draws 1000 W and uses an OAM module format with no external power connectors, indicating it is designed for server racks with dedicated power delivery. The RTX 5090 D V2 draws 575 W and uses a standard 16-pin connector, suitable for desktop builds.

Where Each One Wins

The RTX 5090 D V2 wins in raw compute throughput. Its 104.8 TFLOPS FP32 and FP16 figures exceed the MI350X's 72.09 TFLOPS by 32.71 TFLOPS. It also has higher base and boost clocks, more shading units, and the only recorded benchmark score in this comparison. For any workload that relies on shader compute, ray tracing, or tensor operations, the NVIDIA card has the hardware resources.

The RTX 5090 D V2 also wins in graphics capability. It has 176 ROPs, a 423.6 GPixel/s pixel rate, RT cores, tensor cores, and full DirectX 12 Ultimate support. The MI350X has zero ROPs, zero pixel rate, no RT cores, and no graphics API support. Any task that produces a visual output requires the NVIDIA card.

The MI350X wins decisively in memory capacity and bandwidth. Its 288 GB of HBM3e is 264 GB more than the RTX 5090 D V2's 24 GB of GDDR7. Its 8.19 TB/s bandwidth is 6.85 TB/s higher. For workloads that need to keep massive datasets resident in GPU memory, such as large model inference or scientific computing, the MI350X is the only viable option between these two.

The MI350X also wins in texture rate at 2,252.8 GTexel/s versus 1,636.8 GTexel/s for the RTX 5090 D V2. This is a 616 GTexel/s advantage, though the MI350X has no display outputs to use that texture throughput for graphics.

The RTX 5090 D V2 wins on physical integration. It is a dual-slot card with a 16-pin connector, a 575 W TDP, and a 950 W suggested PSU, making it installable in a desktop system. The MI350X is an OAM module with a 1000 W TDP, a 1400 W suggested PSU, and no power connectors, requiring specialized server infrastructure.

The RTX 5090 D V2 also has the advantage of being an active production product with benchmark data. Its 16504 Steel Nomad score and 59th percentile ranking provide a measurable baseline. The MI350X has no recorded performance data and no production status, leaving its real-world performance unverified in the database.