AMD Radeon R9 M390X vs NVIDIA Tesla K80 Comparison

The AMD Radeon R9 M390X and NVIDIA Tesla K80 represent two very different philosophies from the same 28 nm era, and the benchmark data shows a clear, if narrow, overall victory for the NVIDIA part. The Tesla K80 edges out the Radeon in the only shared test, but the M390X holds its own in the broader performance percentile rankings, making this a nuanced comparison of a mobile graphics chip against a dual-slot compute accelerator.

Head-to-Head Benchmarks

The only directly comparable benchmark result in the data is the Geekbench Vulkan test, and it tells a straightforward story. The NVIDIA Tesla K80 scores 19,111 points, while the AMD Radeon R9 M390X trails at 16,951 points. This represents a delta of -11.3% for the AMD part, meaning the Tesla K80 is roughly 11% faster in this specific API workload. The result is decisive enough to give the NVIDIA card the sole head-to-head win, but it does not paint a complete picture of their relative capabilities.

Looking at the average benchmark scores, the gap narrows significantly. The Radeon R9 M390X posts an average score of 20,662, while the Tesla K80 averages 18,866. That is a reversal of the Vulkan result, with the AMD card coming out ahead by approximately 9.5% when considering all their respective benchmark runs. This discrepancy highlights how API-specific performance can be; the Vulkan test favors the Tesla’s raw compute throughput, while the average across other workloads suggests the Radeon has better overall balance in the tasks it was tested on.

The percentile rankings further contextualize this. The Radeon R9 M390X sits in the 66th percentile of all GPUs, while the Tesla K80 falls to the 63rd percentile. This means the AMD card outperforms a slightly larger share of the GPU landscape, even though it loses the direct Vulkan comparison. The nearest rivals for each card reinforce this positioning: the Radeon’s closest competitor is the AMD Radeon RX 5600 XT, which scores 20,713 against the M390X’s 20,662, a negligible delta of -0.2%. On the NVIDIA side, the Tesla K80 is nearly tied with the RTX 2000 Ada Generation, which scores 18,954 versus 18,866, a delta of -0.5%. These tight margins indicate both cards are well-anchored in their respective performance tiers.

Where Each One Wins

The NVIDIA Tesla K80 wins in raw compute-oriented tasks, as evidenced by its Vulkan benchmark victory and its superior fill rates. It delivers 42.85 GPixel/s of pixel throughput and 171.4 GTexel/s of texture throughput, both of which are substantially higher than the Radeon’s 23.14 GPixel/s and 92.54 GTexel/s. For workloads that hammer the rasterizer or texture units—such as certain scientific visualizations or heavy compute shaders—the Tesla has a clear advantage. Its FP32 performance of 4.113 TFLOPS also outpaces the Radeon’s 2.961 TFLOPS, making it the stronger choice for general-purpose compute tasks that rely on single-precision floating-point math.

The AMD Radeon R9 M390X wins in overall versatility and efficiency. Its higher average benchmark score and better percentile ranking suggest it handles a wider range of workloads more consistently. The card also supports FP16 at a 1:1 ratio, delivering 2.961 TFLOPS in half-precision, whereas the Tesla K80 has no listed FP16 capability. This makes the Radeon a more future-proof option for applications that leverage mixed-precision or half-precision arithmetic, which is increasingly common in machine learning inference and certain graphics effects. The Radeon also consumes far less power, with a TDP of 75 W compared to the Tesla’s 300 W, making it suitable for mobile or low-power deployments where the Tesla’s dual-slot, 8-pin-powered design would be impractical.

Architecture Differences

The architectural divide is stark. The AMD Radeon R9 M390X is built on GCN 3.0, codenamed Amethyst, and belongs to the Gem System R9 M300 generation. It is fabricated on a 28 nm process at TSMC, with 5,000 million transistors packed into a 366 mm² die, yielding a transistor density of 13.7M per mm². The NVIDIA Tesla K80 uses the older Kepler 2.0 architecture, specifically the GK210 chip, and belongs to the Tesla Kepler Kxx generation. It also uses a 28 nm TSMC process but packs 7,100 million transistors into a much larger 561 mm² die, resulting in a lower density of 12.7M per mm². This means NVIDIA chose to scale up the chip physically rather than shrink it, which contributed to its higher power draw.

The memory configurations differ significantly. The Radeon features 4 GB of GDDR5 on a 256-bit bus, delivering 160.0 GB/s of bandwidth. The Tesla K80 doubles down with 12 GB of GDDR5 on a 384-bit bus, achieving 240.6 GB/s. That larger memory pool is a clear advantage for datasets that exceed 4 GB, a common scenario in compute workloads. The Radeon’s memory clock is listed at 1250 MHz (5 Gbps effective), while the Tesla runs at 1253 MHz (5 Gbps effective), so the bandwidth difference comes entirely from the wider bus.

Shader and texture resources also favor the Tesla. It has 2,496 shading units, 208 TMUs, and 48 ROPs, versus the Radeon’s 2,048 shading units, 128 TMUs, and 32 ROPs. The Tesla’s higher counts directly explain its superior fill rates and FP32 throughput. Clock speeds tell a different story: the Radeon has no listed base or boost clock, but the Tesla runs at a modest 562 MHz base and 824 MHz boost. Despite the lower clocks, the Tesla’s sheer resource count gives it the edge in raw throughput.

Feature support shows a split. The Radeon supports DirectX 12 (12_0), OpenGL 4.6, and Vulkan 1.2.170. The Tesla supports DirectX 12 (11_1), a slightly older feature level, but matches OpenGL 4.6 and has a newer Vulkan version at 1.2.175. The Radeon offers display outputs (portable device dependent), while the Tesla has no display outputs at all, confirming its role as a headless compute accelerator. The Radeon is an MXM module with no power connectors, while the Tesla is a dual-slot card requiring a 1x 8-pin connector and a suggested 700 W PSU.

FAQ

Q: Which GPU has higher raw FP32 compute performance?

A: The NVIDIA Tesla K80 leads with 4.113 TFLOPS, compared to the AMD Radeon R9 M390X’s 2.961 TFLOPS.

Q: Does the AMD card support half-precision FP16 math?

A: Yes, the Radeon R9 M390X delivers 2.961 TFLOPS in FP16 at a 1:1 ratio. The Tesla K80 has no listed FP16 performance.

Q: How do their memory bandwidths compare?

A: The Tesla K80 offers 240.6 GB/s over a 384-bit bus, while the Radeon R9 M390X provides 160.0 GB/s over a 256-bit bus.

Q: What is the average benchmark score difference?

A: The Radeon R9 M390X averages 20,662, which is about 9.5% higher than the Tesla K80’s average of 18,866.

Q: Which card is more power-efficient?

A: The Radeon R9 M390X has a TDP of 75 W, far lower than the Tesla K80’s 300 W, making it better suited for mobile or constrained environments.

Q: Can the Tesla K80 output to a display?

A: No, the Tesla K80 has no display outputs, whereas the Radeon R9 M390X’s outputs are portable device dependent.

Specification Differences

| Specification | AMD Radeon R9 M390X | NVIDIA Tesla K80 |

|---------------|---------------------|------------------|

| Architecture | GCN 3.0 | Kepler 2.0 |

| Chip | Amethyst | GK210 |

| Generation | Gem System (R9 M300) | Tesla Kepler (Kxx) |

| Transistors | 5,000 million | 7,100 million |

| Die Size | 366 mm² | 561 mm² |

| Transistor Density | 13.7M / mm² | 12.7M / mm² |

| Base Clock | Not listed | 562 MHz |

| Boost Clock | Not listed | 824 MHz |

| Memory Size | 4 GB | 12 GB |

| Memory Bus Width | 256 bit | 384 bit |

| Memory Bandwidth | 160.0 GB/s | 240.6 GB/s |

| Shading Units | 2048 | 2496 |

| TMUs | 128 | 208 |

| ROPs | 32 | 48 |

| Pixel Rate | 23.14 GPixel/s | 42.85 GPixel/s |

| Texture Rate | 92.54 GTexel/s | 171.4 GTexel/s |

| FP32 | 2.961 TFLOPS | 4.113 TFLOPS |

| FP16 | 2.961 TFLOPS (1:1) | Not listed |

| TDP | 75 W | 300 W |

| Slot Width | MXM Module | Dual-slot |

| Power Connectors | None | 1x 8-pin |

| Suggested PSU | Not listed | 700 W |

| Display Outputs | Portable Device Dependent | No outputs |

| DirectX Support | 12 (12_0) | 12 (11_1) |

| Vulkan Support | 1.2.170 | 1.2.175 |

| Release Date | 2015-05-04 | 2014-11-16 |

| Predecessor | Solar System | Tesla Fermi |

| Successor | Polaris Mobile | Tesla Maxwell |

| Avg Benchmark Score | 20,662 | 18,866 |

| Percentile | 66th | 63rd |