AMD Instinct MI350P vs AMD Instinct MI355X Comparison

Head-to-Head Benchmarks

The recorded data for the AMD Instinct MI350P and AMD Instinct MI355X contains no direct benchmark scores, so the comparison relies entirely on their computed specification-derived throughput values. The MI355X dominates the MI350P in raw computational output. In FP32 floating-point work, the MI355X delivers 78.64 TFLOPS against the MI350P's 36.04 TFLOPS, a lead of more than 2.18x. The FP16 figures mirror this exactly, with the MI355X again at 78.64 TFLOPS and the MI350P at 36.04 TFLOPS, both listed as 1:1 ratios, meaning the accelerators do not gain any throughput advantage when switching to reduced precision.

Texture throughput follows the same pattern. The MI355X reaches 2,457.6 GTexel/s, while the MI350P produces 1,126.4 GTexel/s. That is roughly 2.18x higher texture fill performance for the MI355X, consistent with the doubling of its shading units and texture mapping units. The MI350P fields 8,192 shading units and 512 TMUs, while the MI355X doubles both to 16,384 shading units and 1,024 TMUs. This direct scaling explains the near-perfect ratio in both FP32 and texture rates.

Memory bandwidth is one area where the two are identical. Both cards use HBM3e memory with an 8,192-bit bus and 8.19 TB/s of bandwidth. The memory clock is also the same at 2000 MHz, with 8 Gbps effective data rate. The difference lies in capacity. The MI350P carries 144 GB, while the MI355X carries exactly double that at 288 GB. Since bandwidth matches, neither accelerator has a transfer-rate advantage, but the MI355X can keep twice as much data resident on-device.

Pixel rate is a non-factor for both. Each part lists 0 MPixel/s, and both have no display outputs and no DirectX, OpenGL, or Vulkan API support. These are compute-only accelerators, and their render-output pipeline is not populated. The texture rate figures are the only pixel-pipeline-adjacent metrics that matter here, and the MI355X wins decisively.

Neither accelerator has any recorded wins in head-to-head benchmark comparisons, and both sit at the 50th percentile among all GPUs in the database, with an average benchmark score of 0. The lack of measured application results means the specification-derived numbers above are the only quantitative basis for separating the two.

The Verdict

The data clearly separates these two accelerators by capability tier. The MI355X is the stronger compute part in every throughput metric that differs. Its FP32 output of 78.64 TFLOPS is more than double the MI350P's 36.04 TFLOPS, and its texture rate of 2,457.6 GTexel/s likewise doubles the MI350P's 1,126.4 GTexel/s. Its 288 GB memory capacity doubles the MI350P's 144 GB, while bandwidth stays constant at 8.19 TB/s.

The MI350P is the lower-power, lower-throughput option. Its 600 W TDP is less than half of the MI355X's 1400 W, and its suggested PSU of 1000 W is far below the MI355X's 1800 W. The MI350P also uses a conventional dual-slot form factor with a single 16-pin power connector, which makes it adaptable to standard server chassis. The MI355X is an OAM module with no power connectors of its own, meaning it requires a baseboard designed for OAM carrier modules.

For workloads constrained by power delivery or chassis compatibility, the MI350P is the more practical fit. Its power envelope and dual-slot layout allow deployment in systems that cannot support a 1400 W module. For workloads where raw throughput and memory capacity dominate, the MI355X is the obvious choice. The data shows no scenario where the MI350P outperforms the MI355X on compute or memory capacity, but it does show a clear trade-off in power and physical integration.

The release dates also separate the two. The MI355X appeared first, with a release date of June 11, 2025, while the MI350P arrived later on May 6, 2026. Both list the Radeon Instinct family as their predecessor, and neither has a successor listed in the database.

Architecture Differences

Both accelerators share the CDNA 4.0 architecture, the Instinct (MIx) generation, and a 3 nm process node from TSMC. The foundry is identical, and both use the MI350 chip family. The MI350P uses the MI350 128CU chip, while the MI355X uses the MI350 256CU chip, which doubles the compute unit count and therefore doubles the shading units and TMUs.

Transistor counts differ substantially. The MI350P has 73,000 million transistors on a 1190 mm² die, giving a transistor density of 61.3M per mm². The MI355X has 185,000 million transistors on a 2380 mm² die, a density of 77.7M per mm². The MI355X die is exactly double the area of the MI350P, but it packs more than double the transistors, which explains the higher density. The MI355X not only adds more compute units but does so with greater transistor efficiency per square millimeter.

Base clocks are identical at 1000 MHz. The boost clock differs, with the MI350P at 2200 MHz and the MI355X at 2400 MHz. That 200 MHz boost advantage adds to the MI355X's throughput lead beyond the pure core-count scaling. The combination of double the shading units and a higher boost clock pushes the MI355X FP32 figure to 78.64 TFLOPS versus the MI350P's 36.04 TFLOPS.

Neither part has ray tracing cores, tensor cores, or render output units listed. Both are compute accelerators without graphics-oriented hardware. The API support is N/A across DirectX, OpenGL, and Vulkan for both, reinforcing that these are not consumer graphics cards. The memory subsystem is identical in bus width, type, and bandwidth, but the MI355X doubles capacity to 288 GB. Both run their memory at 2000 MHz with 8 Gbps effective transfer.

Specification Differences

The MI350P and MI355X differ in several specification fields. The chip designation changes from MI350 128CU to MI350 256CU. Transistors go from 73,000 million to 185,000 million. Die size goes from 1190 mm² to 2380 mm². Transistor density goes from 61.3M per mm² to 77.7M per mm². Boost clock goes from 2200 MHz to 2400 MHz. Shading units double from 8,192 to 16,384. TMUs double from 512 to 1,024. Texture rate goes from 1,126.4 GTexel/s to 2,457.6 GTexel/s. FP32 and FP16 both go from 36.04 TFLOPS to 78.64 TFLOPS.

Memory capacity doubles from 144 GB to 288 GB. TDP goes from 600 W to 1400 W. Slot width changes from dual-slot to OAM module. Power connectors change from 1x 16-pin to none. Suggested PSU goes from 1000 W to 1800 W. Dimensions change: the MI350P is 267 mm long, 111 mm high, and 40 mm wide, while the MI355X is 102 mm long and 165 mm wide with no height listed. Release dates differ, with the MI355X on June 11, 2025 and the MI350P on May 6, 2026.

Fields that remain identical include the architecture, process node, foundry, base clock, memory clock, memory type, memory bus width, memory bandwidth, pixel rate, RT cores, tensor cores, API support, display outputs, bus interface, production status, predecessor, successor, and launch MSRP (none listed for either).

FAQ

Q: Which accelerator has higher FP32 throughput?

A: The AMD Instinct MI355X delivers 78.64 TFLOPS in FP32, while the AMD Instinct MI350P delivers 36.04 TFLOPS. The MI355X is more than double the FP32 output of the MI350P.

Q: Do the two cards have the same memory bandwidth?

A: Yes. Both use HBM3e memory with an 8,192-bit bus and 8.19 TB/s of bandwidth. The memory clock is identical at 2000 MHz with 8 Gbps effective transfer.

Q: How much memory does each accelerator carry?

A: The MI350P has 144 GB of HBM3e memory. The MI355X has 288 GB, exactly double the MI350P's capacity.

Q: What are the power requirements for each?

A: The MI350P has a 600 W TDP and a suggested PSU of 1000 W. The MI355X has a 1400 W TDP and a suggested PSU of 1800 W. The MI350P uses a 1x 16-pin power connector, while the MI355X is an OAM module with no power connectors.

Q: Which accelerator uses a larger die?

A: The MI355X has a die size of 2380 mm² with 185,000 million transistors. The MI350P has a die size of 1190 mm² with 73,000 million transistors. Both are built on a 3 nm TSMC process.

Q: When did each accelerator release?

A: The MI355X released on June 11, 2025. The MI350P released later, on May 6, 2026. Both are in the Instinct (MIx) generation with the CDNA 4.0 architecture.