AMD Instinct MI350X vs AMD Radeon Instinct MI308X Comparison

Head-to-Head Benchmarks

The database shows a direct comparison between two AMD data center accelerators, but the recorded data contains no actual benchmark scores for either product. Both the AMD Instinct MI350X and the AMD Radeon Instinct MI308X have empty benchmark arrays, zero average benchmark scores, and zero wins in the head-to-head comparison. The percentile ranking for both GPUs against all other graphics cards in the database is identical at the 50th percentile, indicating that without test results, both occupy the median position by default. This means no performance deltas can be calculated from measured data, and any claims of superiority would lack factual grounding in the recorded metrics.

What the data does reveal is the raw theoretical throughput figures from the specification sheets. The Radeon Instinct MI308X delivers 81.72 TFLOPS of FP32 compute, while the Instinct MI350X provides 72.09 TFLOPS. That places the MI308X approximately 13% ahead in single-precision floating-point performance based on these stated figures. In FP16 compute, the gap widens substantially: the MI308X reaches 653.7 TFLOPS with an 8:1 ratio, whereas the MI350X achieves 72.09 TFLOPS with a 1:1 ratio. The MI308X therefore holds a massive advantage in half-precision workloads, offering over nine times the FP16 throughput on paper. Texture fill rates also favor the MI308X, which records 2,553.6 GTexel/s compared to the MI350X's 2,252.8 GTexel/s.

However, the MI350X counters in memory capacity. It carries 288 GB of HBM3e memory, while the MI308X comes with 192 GB of HBM3. That is a 50% capacity advantage for the MI350X. Memory bandwidth tells a different story: the MI308X achieves 10.3 TB/s, while the MI350X provides 8.19 TB/s. The MI308X thus offers roughly 26% higher bandwidth despite having less total memory. Both cards use an 8192-bit memory bus, so the bandwidth difference stems from the memory clock: the MI308X runs at 2525 MHz with 10.1 Gbps effective, while the MI350X runs at 2000 MHz with 8 Gbps effective.

Architecture Differences

The two accelerators represent successive generations of AMD's CDNA architecture. The Instinct MI350X uses CDNA 4.0, built on a 3 nm process at TSMC. The Radeon Instinct MI308X uses CDNA 3.0, fabricated on a 5 nm process, also at TSMC. This process shrink allows the MI350X to pack more transistors into a larger die: 185,000 million transistors across a 2380 mm² die, versus 153,000 million transistors on a 1017 mm² die for the MI308X. Transistor density tells an interesting story: the MI308X achieves 150.4 million transistors per square millimeter, while the MI350X reaches only 77.7 million per square millimeter. The older 5 nm chip is denser per area, but the newer 3 nm chip has more total transistors and a much larger physical footprint.

Compute unit organization differs as well. The MI350X uses the MI350 256CU chip, while the MI308X uses the Aqua Vanjaram chip. Despite the MI350X having fewer shading units (16,384 versus 19,456), it still delivers competitive FP32 performance due to its higher boost clock. The MI350X boosts to 2200 MHz, while the MI308X boosts to 2100 MHz. Both share a 1000 MHz base clock. The MI350X also has fewer texture mapping units: 1,024 versus 1,216 for the MI308X. Neither card has ROPs, with pixel rates of 0 MPixel/s for both, and neither has dedicated ray tracing or tensor cores listed in the data.

Memory technology marks a clear generational split. The MI350X uses HBM3e, while the MI308X uses HBM3. The MI350X also has a later release date of June 11, 2025, compared to December 5, 2023 for the MI308X. Power requirements differ substantially: the MI350X has a TDP of 1000 W with a suggested PSU of 1400 W, while the MI308X has a TDP of 750 W with a suggested PSU of 1150 W. Both use an OAM Module slot width, have no power connectors, and feature PCIe 5.0 x16 bus interfaces. Neither card has display outputs, and both list no supported graphics APIs in their respective entries.

Where Each One Wins

Based purely on the specification data, the Radeon Instinct MI308X wins in raw compute throughput. Its FP32 figure of 81.72 TFLOPS exceeds the MI350X's 72.09 TFLOPS, and its FP16 figure of 653.7 TFLOPS dwarfs the MI350X's 72.09 TFLOPS. The MI308X also delivers higher memory bandwidth at 10.3 TB/s versus 8.19 TB/s, and a higher texture fill rate at 2,553.6 GTexel/s versus 2,252.8 GTexel/s. For workloads that depend on half-precision math, such as certain AI training and inference tasks, the MI308X's 8:1 FP16 ratio provides a theoretical throughput advantage that the MI350X cannot match with its 1:1 ratio.

The Instinct MI350X wins in memory capacity, offering 288 GB versus 192 GB. This 96 GB difference allows larger models and datasets to reside in on-board memory without spilling to system RAM or storage. The MI350X also has a higher boost clock at 2200 MHz versus 2100 MHz, and it uses a more advanced 3 nm process node. Its larger die size of 2380 mm², compared to 1017 mm², accommodates the additional transistor count. The MI350X's newer architecture, CDNA 4.0, and later release date reflect a more recent design generation.

For power-constrained deployments, the MI308X holds an advantage with a 750 W TDP versus 1000 W, and a suggested PSU of 1150 W versus 1400 W. This means the MI308X can operate in systems with less substantial power delivery infrastructure. The MI350X, by contrast, demands a higher power envelope but offers greater memory capacity per watt in the recorded data.

Specification Differences

The two cards differ across nearly every major specification field. Process node: 3 nm for the MI350X, 5 nm for the MI308X. Transistor count: 185,000 million versus 153,000 million. Die size: 2380 mm² versus 1017 mm². Transistor density: 77.7 million per mm² versus 150.4 million per mm². Shading units: 16,384 versus 19,456. Texture mapping units: 1,024 versus 1,216. Boost clock: 2200 MHz versus 2100 MHz. Memory size: 288 GB versus 192 GB. Memory type: HBM3e versus HBM3. Memory clock: 2000 MHz with 8 Gbps effective versus 2525 MHz with 10.1 Gbps effective. Memory bandwidth: 8.19 TB/s versus 10.3 TB/s. FP32 performance: 72.09 TFLOPS versus 81.72 TFLOPS. FP16 performance: 72.09 TFLOPS (1:1) versus 653.7 TFLOPS (8:1). Texture rate: 2,252.8 GTexel/s versus 2,553.6 GTexel/s. TDP: 1000 W versus 750 W. Suggested PSU: 1400 W versus 1150 W. Release date: June 11, 2025 versus December 5, 2023.

Shared specifications include the 8192-bit memory bus width, 1000 MHz base clock, PCIe 5.0 x16 interface, OAM Module form factor, no power connectors, no display outputs, and 0 MPixel/s pixel rate for both. Both cards have no ray tracing cores and no tensor cores listed. Both come from AMD and are fabricated by TSMC. Neither has an available launch MSRP in the database.

FAQ

Q: Which GPU has more memory?

A: The AMD Instinct MI350X has 288 GB of HBM3e memory, while the AMD Radeon Instinct MI308X has 192 GB of HBM3 memory. The MI350X offers a 96 GB capacity advantage.

Q: Which GPU provides higher FP32 compute performance?

A: The AMD Radeon Instinct MI308X delivers 81.72 TFLOPS of FP32 performance, which exceeds the AMD Instinct MI350X's 72.09 TFLOPS.

Q: How do the memory bandwidth figures compare?

A: The AMD Radeon Instinct MI308X achieves 10.3 TB/s of memory bandwidth, while the AMD Instinct MI350X provides 8.19 TB/s. Both use an 8192-bit memory bus.

Q: What is the difference in power consumption?

A: The AMD Instinct MI350X has a TDP of 1000 W with a suggested PSU of 1400 W. The AMD Radeon Instinct MI308X has a TDP of 750 W with a suggested PSU of 1150 W.

Q: Which GPU uses a more advanced manufacturing process?

A: The AMD Instinct MI350X uses a 3 nm process at TSMC, while the AMD Radeon Instinct MI308X uses a 5 nm process at TSMC.

Q: What are the release dates for these GPUs?

A: The AMD Instinct MI350X was released on June 11, 2025. The AMD Radeon Instinct MI308X was released on December 5, 2023.

The Verdict

The recorded data presents a clear split between the two accelerators. The AMD Radeon Instinct MI308X is the stronger choice for raw compute throughput, particularly in FP16 workloads where its 653.7 TFLOPS (8:1) vastly exceeds the MI350X's 72.09 TFLOPS (1:1). Its higher FP32 figure of 81.72 TFLOPS, higher memory bandwidth of 10.3 TB/s, and higher texture rate of 2,553.6 GTexel/s reinforce this position. The MI308X also requires less power, with a 750 W TDP compared to 1000 W, making it suitable for systems with a 1150 W PSU rather than 1400 W.

The AMD Instinct MI350X is the choice for memory-bound workloads. Its 288 GB capacity, compared to 192 GB, enables larger datasets to fit on the card. The MI350X also benefits from a newer architecture (CDNA 4.0 versus CDNA 3.0), a smaller 3 nm process node, and a higher boost clock of 2200 MHz versus 2100 MHz. Its larger die of 2380 mm² and higher transistor count of 185,000 million indicate a more complex and recent design.

There are no benchmark results in the database to validate real-world performance. The percentile rankings for both cards sit at 50, and the average benchmark scores are zero. Therefore, these conclusions rest entirely on the theoretical specifications recorded in the data. Users requiring maximum FP32 or FP16 throughput should look to the MI308X. Users requiring maximum memory capacity for large model residency should look to the MI350X. The choice hinges on whether compute density or memory capacity takes priority in the target workload.