AMD Instinct MI308X vs AMD Radeon 680M Comparison

AMD Instinct MI308X and AMD Radeon 680M occupy opposite ends of the AMD accelerator spectrum. The MI308X is a data center compute module built on CDNA 3.0, while the 680M is a mobile integrated graphics processor based on RDNA 2.0. The recorded data shows no shared benchmark suite between the two, as the MI308X has no benchmark scores in the database, while the 680M has three recorded tests. This analysis separates the two by architectural purpose, measured performance where available, and specification differences.

Where Each One Wins

The AMD Instinct MI308X wins in raw compute throughput and memory capacity. Its FP32 output is recorded at 81.72 TFLOPS, and its FP16 output matches at 81.72 TFLOPS with a 1:1 ratio. The Radeon 680M delivers 3.379 TFLOPS FP32 and 6.758 TFLOPS FP16 with a 2:1 ratio. The MI308X exceeds the 680M by a factor of over 24 in FP32 compute, and by over 12 in FP16 compute. This makes the MI308X the clear choice for dense numerical workloads such as large-scale matrix operations, deep learning training, or scientific simulation where the data path is fully utilized.

The Radeon 680M wins in accessibility and feature completeness for graphics rendering. It supports DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4, while the MI308X records N/A for all three APIs. The 680M has 32 ROPs and a pixel rate of 70.40 GPixel/s, whereas the MI308X records 0 ROPs and 0 MPixel/s. The 680M also includes 12 ray tracing cores, which the MI308X does not list. For any workload that requires rasterization, ray tracing, or standard graphics APIs, the 680M is the functional option. The MI308X has no display outputs, while the 680M relies on portable device dependent outputs.

The 680M also wins in the database benchmark results. Its average benchmark score is 15270, with a percentile rank of 57 against all GPUs. The MI308X has no recorded benchmarks, an average score of 0, and a percentile rank of 50. The 680M’s nearest rivals in the database are the NVIDIA GeForce GTX 580 with an average score of 15283 (0.1% higher), the NVIDIA GeForce RTX 2060 with 15290 (0.1% higher), the NVIDIA GeForce RTX 3050 OEM with 15199 (0.5% lower), and the AMD Radeon RX 7600 with 15171 (0.7% lower). These deltas are very small, indicating the 680M sits in a tightly packed performance band among those discrete GPUs for the tests recorded.

Architecture Differences

The two chips come from different architectures and process nodes. The MI308X uses CDNA 3.0, designed for compute acceleration, while the 680M uses RDNA 2.0, a graphics-oriented architecture. The MI308X is fabricated on a 5 nm process at TSMC, while the 680M uses a 6 nm process at TSMC. Transistor counts differ massively: the MI308X has 153,000 million transistors on a 1017 mm² die, while the 680M has 13,100 million transistors on a 208 mm² die. Transistor density reflects this: the MI308X reaches 150.4 million transistors per mm², the 680M reaches 63.0 million per mm².

Core configurations diverge sharply. The MI308X has 19,456 shading units, 1,216 texture mapping units, and 0 ROPs. The 680M has 768 shading units, 48 TMUs, and 32 ROPs. The MI308X has no recorded ray tracing cores, while the 680M includes 12. Texture rate for the MI308X is 2,553.6 GTexel/s, versus 105.6 GTexel/s for the 680M. The MI308X’s pixel rate is 0 MPixel/s, which aligns with its lack of ROPs; the 680M’s pixel rate is 70.40 GPixel/s.

Memory architecture is another fundamental split. The MI308X uses 192 GB of HBM3 on a 8192-bit bus, with a memory clock of 1300 MHz (5.2 Gbps effective) and bandwidth of 5.32 TB/s. The 680M uses system shared memory, with type, bus width, and bandwidth all listed as system dependent. This means the 680M’s memory performance depends on the host platform, while the MI308X has dedicated high-bandwidth memory. The MI308X’s 8192-bit bus is the widest in the database for these products.

Clock behavior differs. The MI308X has a base clock of 1000 MHz and a boost clock of 2100 MHz. The 680M has a base clock of 2000 MHz and a boost clock of 2200 MHz. Despite the 680M’s higher base clock, the MI308X’s massive core count and memory bandwidth dominate throughput. The MI308X has a TDP of 750 W, while the 680M has a TDP of 50 W. Power connectors are listed as none for both, but the MI308X suggests a 1150 W power supply, while the 680M has no suggested PSU. The MI308X uses an OAM Module slot width, the 680M is an IGP. Bus interface: the MI308X uses PCIe 5.0 x16, the 680M uses PCIe 4.0 x8.

Release dates are close: the MI308X launched on 2023-12-05, the 680M on 2023-01-02. The 680M has a production status of Active and a successor listed as Navi III IGP, while the MI308X has no production status, no successor, and its predecessor is Radeon Instinct. The 680M’s predecessor is Vega II IGP.

The Verdict

From the recorded data, the AMD Instinct MI308X is the compute powerhouse. Its FP32 and FP16 throughput of 81.72 TFLOPS each, combined with 192 GB HBM3 memory and 5.32 TB/s bandwidth, makes it suited for large-scale numerical processing. The 680M cannot approach this compute density; its FP32 is 3.379 TFLOPS and its memory is system dependent. The MI308X’s 1:1 FP16 ratio also indicates it handles half-precision workloads at full rate, whereas the 680M’s 2:1 ratio halves its FP16 throughput relative to FP32.

The AMD Radeon 680M is the graphics-capable option. It has a full graphics pipeline with 32 ROPs, 70.40 GPixel/s pixel rate, 12 ray tracing cores, and support for DirectX 12 Ultimate, OpenGL 4.6, and Vulkan 1.4. The MI308X has none of these features, recording no ROPs, no pixel rate, and no API support. The 680M also has measured benchmark scores: 378 in 3DMark Steel Nomad DX12, 23468 in Geekbench OpenCL, and 21965 in Geekbench Vulkan. Its average benchmark score of 15270 places it at the 57th percentile, and it performs within 0.7% of the NVIDIA GeForce GTX 580, RTX 2060, RTX 3050 OEM, and AMD Radeon RX 7600 in the database. The MI308X has no benchmark scores, so its real-world performance cannot be compared directly to those rivals.

For a system that needs graphics output, ray tracing, and standard API support, the data points to the 680M. For a system that needs maximum raw compute and dedicated high-bandwidth memory, the data points to the MI308X. The MI308X has no display outputs, making it unusable for any visual output task. The 680M depends on portable device outputs, meaning it works in mobile systems. The power envelope also separates them: 750 W for the MI308X versus 50 W for the 680M, a 15-fold difference. The MI308X requires a 1150 W suggested PSU, which is not applicable to the 680M.

FAQ

Q: Which GPU has higher FP32 compute performance?

A: The AMD Instinct MI308X has 81.72 TFLOPS FP32, while the AMD Radeon 680M has 3.379 TFLOPS FP32. The MI308X delivers over 24 times the FP32 throughput.

Q: Does the MI308X support DirectX 12 or Vulkan?

A: No. The MI308X records N/A for DirectX, OpenGL, and Vulkan. The Radeon 680M supports DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4.

Q: What memory does each product use?

A: The MI308X uses 192 GB of HBM3 with an 8192-bit bus and 5.32 TB/s bandwidth. The 680M uses system shared memory, with its type, bus width, and bandwidth all system dependent.

Q: How does the 680M compare to its nearest rivals in benchmark scores?

A: The 680M has an average benchmark score of 15270. The NVIDIA GeForce GTX 580 scores 15283 (0.1% higher), the NVIDIA GeForce RTX 2060 scores 15290 (0.1% higher), the NVIDIA GeForce RTX 3050 OEM scores 15199 (0.5% lower), and the AMD Radeon RX 7600 scores 15171 (0.7% lower).

Q: What is the transistor count difference?

A: The MI308X has 153,000 million transistors, while the 680M has 13,100 million transistors. The MI308X also has a larger die at 1017 mm² versus 208 mm².

Q: Does either product have ray tracing cores?

A: The Radeon 680M has 12 ray tracing cores. The MI308X does not list any ray tracing cores in the database.

Head-to-Head Benchmarks

The database contains no direct head-to-head benchmark rows between the MI308X and the 680M. The MI308X has zero recorded benchmark entries, while the 680M has three. The largest wins for each side must be inferred from the specification data.

The MI308X’s dominant advantage is compute throughput. Its FP32 of 81.72 TFLOPS is over 24 times the 680M’s 3.379 TFLOPS. Its FP16 of 81.72 TFLOPS (1:1) is over 12 times the 680M’s 6.758 TFLOPS (2:1). Texture rate follows the same pattern: 2,553.6 GTexel/s versus 105.6 GTexel/s, a factor of 24.2. Memory bandwidth is even more lopsided: 5.32 TB/s versus system dependent, but the fixed HBM3 capacity of 192 GB versus system shared makes the MI308X’s memory path fully dedicated. The shading unit count of 19,456 versus 768 is a 25.3x difference. The MI308X’s transistor density of 150.4 million per mm² also exceeds the 680M’s 63.0 million per mm², indicating a tighter packing of logic.

The 680M’s wins are in graphics-specific features. It has 32 ROPs and a pixel rate of 70.40 GPixel/s, while the MI308X records 0 ROPs and 0 MPixel/s. The 680M has 12 ray tracing cores, which the MI308X lacks. The 680M supports DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4; the MI308X has N/A for all three. The 680M also has a higher base clock (2000 MHz versus 1000 MHz) and a slightly higher boost clock (2200 MHz versus 2100 MHz), though this does not compensate for the core count difference. The 680M’s power draw is 50 W versus 750 W for the MI308X, a 15x lower requirement.

In the measured benchmark data, the 680M’s scores are: 378 in 3DMark Steel Nomad DX12, 23468 in Geekbench OpenCL, and 21965 in Geekbench Vulkan. Its average benchmark score is 15270, and its nearest rival deltas are all under 1%. The GTX 580 and RTX 2060 each score 0.1% higher, the RTX 3050 OEM scores 0.5% lower, and the RX 7600 scores 0.7% lower. This indicates the 680M, despite being an integrated part, performs comparably to those discrete GPUs in the recorded tests. The MI308X has no such scores, so its benchmark position is undefined beyond a percentile rank of 50 against all GPUs.

Specification Differences

The two products differ in nearly every measured field. Process node: 5 nm for the MI308X, 6 nm for the 680M, both TSMC. Transistors: 153,000 million versus 13,100 million. Die size: 1017 mm² versus 208 mm². Transistor density: 150.4 million per mm² versus 63.0 million per mm². Base clock: 1000 MHz versus 2000 MHz. Boost clock: 2100 MHz versus 2200 MHz. Memory clock: 1300 MHz (5.2 Gbps effective) versus system shared. Memory size: 192 GB HBM3 versus system shared. Memory bus width: 8192 bit versus system shared. Memory bandwidth: 5.32 TB/s versus system dependent.

Shading units: 19,456 versus 768. TMUs: 1,216 versus 48. ROPs: 0 versus 32. Ray tracing cores: none listed versus 12. Pixel rate: 0 MPixel/s versus 70.40 GPixel/s. Texture rate: 2,553.6 GTexel/s versus 105.6 GTexel/s. FP32: 81.72 TFLOPS versus 3.379 TFLOPS. FP16: 81.72 TFLOPS (1:1) versus 6.758 TFLOPS (2:1). TDP: 750 W versus 50 W. Slot width: OAM Module versus IGP. Power connectors: none for both. Suggested PSU: 1150 W for the MI308X, none for the 680M. Bus interface: PCIe 5.0 x16 versus PCIe 4.0 x8. Display outputs: no outputs versus portable device dependent.

APIs: DirectX N/A versus 12 Ultimate (12_2), OpenGL N/A versus 4.6, Vulkan N/A versus 1.4. Release date: 2023-12-05 versus 2023-01-02. Production status: none listed versus Active. Predecessor: Radeon Instinct versus Vega II IGP. Successor: none versus Navi III IGP. The 680M has a recorded benchmark suite and average score of 15270, while the MI308X has none. Percentile rank: 50 for the MI308X, 57 for the 680M. The 680M has four nearest rivals with delta percentages from -0.1 to 0.7; the MI308X has no nearest rivals listed.