AMD Radeon Pro Vega II Duo vs NVIDIA L40S Comparison

# NVIDIA L40S vs AMD Radeon Pro Vega II Duo: Benchmark Database Analysis

The data is unequivocal: the NVIDIA L40S wins every recorded head-to-head benchmark against the AMD Radeon Pro Vega II Duo. Across the two workloads measured, the L40S leads by a wide margin, with the Geekbench OpenCL score showing a 236% advantage and the Vulkan score showing a 162.6% advantage. The AMD part is not without merit, however. It posts a higher percentile ranking relative to all GPUs, and it lands closer to the L40S in average score than any other rival in the database. The short version is this: pick the L40S for raw compute dominance, pick the AMD for a different platform and a different performance profile.

Where Each One Wins

The recorded data splits the two parts cleanly along application lines. The NVIDIA L40S wins both head-to-head tests in the database, with Geekbench OpenCL at 330727 versus 98432 and Geekbench Vulkan at 260799 versus 99296. That is a dominant sweep, with zero wins for the AMD card in the direct comparisons.

However, the AMD Radeon Pro Vega II Duo is not without its own strengths. Its aggregate benchmark score across all recorded tests puts it in the 94th percentile of all GPUs, well above the L40S's 99th percentile. Indeed, the AMD card's average score is 106750, which is within 0.6% of the AMD Radeon Pro W6600 and within 2.6% of the AMD Radeon Pro Vega II. It sits in a tight cluster of high-end AMD parts, meaning it is not an outlier in its own ecosystem.

The L40S sits in a different performance class entirely. Its average score of 295763 places it 3% above the NVIDIA RTX 6000 Ada Generation, 4.1% above the NVIDIA L40, and it even leads the AMD Instinct MI300X by 7%. It trails only the NVIDIA H200 NVL in this peer group, by 11.7%. The data shows a clear hierarchy: the L40S is a top-tier compute part, while the AMD part is a strong mid-to-high-range part.

Architecture Differences

The two parts come from completely different design lineages, and these differences explain the benchmark results.

The NVIDIA L40S is built on the Ada Lovelace architecture, using the AD102 chip fabricated on a 5 nm process at TSMC. It packs 76,300 million transistors onto a 609 mm² die, for a transistor density of 125.3M per mm². The AMD Radeon Pro Vega II Duo uses the Vega 20 chip, GCN 5.1, on a 7 nm TSMC process. It has 13,230 million transistors on a 331 mm² die, which is 40.0M per mm². This is the first major split: the NVIDIA part has roughly three times the transistor density of the AMD part.

The compute configuration is just as different. The L40S has 18,176 shading units, 568 tensor cores, 142 ray tracing cores, and 192 ROPs. It has 568 texture mapping units. The AMD has 4,096 shaders and 256 texture mapping units, and 64 ROPs. The NVIDIA part has more than four times the shaders, more than twice the texture units, and three times the ROPs. The tensor core and RT core field is blank for the AMD part, as those features are not present in that architecture.

Memory and clocks are quite different. The L40S has 48 GB of GDDR6 memory on a 384-bit bus, which provides 864.0 GB/s of bandwidth. Its memory clock is 2250 MHz with an effective data rate of 18 Gbps. The AMD card has 32 GB of HBM2 memory on a 4096-bit bus, with 1.02 TB/s of bandwidth. Its memory clock is 1000 MHz, with an effective data rate of 2 Gbps. The NVIDIA part has 50% more capacity and the AMD part has a much wider bus, but the measured bandwidth is still 18% higher on the NVIDIA.

The compute ceilings follow the same pattern. The L40S delivers 91.61 TFLOPS FP32 and the same 91.61 TFLOPS for FP16 (1:1 ratio). The AMD card delivers 14.09 TFLOPS FP32 and 28.18 TFLOPS FP16 (2:1 ratio). The NVIDIA part is 6.5 times higher in FP32 and 3.25 times higher in FP16.

Other differences are in the platform. The L40S is a dual-slot card with a 16-pin power connector, and the database recommends a 700 W power supply. It is a PCIe 4.0 x16 interface and has 1x HDMI and 3x DisplayPort 1.4a outputs. The AMD card is a quad-slot design, with an Apple MPX bus interface and 1x HDMI 2.0b and 4x Thunderbolt outputs. The NVIDIA card is designed for server use, while the AMD is a Mac-oriented part. The L40S supports DirectX 12 Ultimate (12_2), OpenGL 4.6, Vulkan 1.4. The AMD supports DirectX 12 (12_1), OpenGL 4.6, Vulkan 1.3.

Head-to-Head Benchmarks

The first head-to-head benchmark in the database is the Geekbench OpenCL test. Here the NVIDIA L40S scored 330,727, against the AMD's 98,432. That is a 236% lead for the NVIDIA part. This is not a close contest. The score of the NVIDIA is 3.36 times higher than that of the AMD part. The OpenCL test is largely a compute-bound workload, and the L40S's enormous FP32 throughput and memory bandwidth simply overwhelm the AMD card.

The second record is the Geekbench Vulkan test. The NVIDIA L40S scored 260,799, the AMD Radeon Pro Vega II Duo scored 99,296. The NVIDIA part is 162.6% higher. Again the NVIDIA part wins. The L40S's Vulkan performance, while not as high as its OpenCL score, is still in a different class than the AMD part. The AMD's Vulkan score is actually close to its OpenCL score, indicating that the AMD part is performing consistently across APIs, but it is simply outclassed in both.

FAQ

Q: Which GPU is the fastest in the database?

A: Based on the average benchmark score, the NVIDIA L40S is the fastest in this pair, with NVIDIA H200 NVL above it. The NVIDIA L40S scores 295,763 on average, placing it in the 99th percentile. That is well above the AMD Radeon Pro Vega II Duo's 106,750, which is in the 94th percentile.

Q: How much faster is the NVIDIA L40S in OpenCL?

A: It is 236% faster. Scores of 330,727 versus 98,432.

Q: What about the Vulkan performance?

A: The NVIDIA L40S scored 260,799, which is 162.6% higher than the AMD Radeon Pro Vega II Duo's 99,296.

Q: Does the AMD card win any benchmark?

A: No. The recorded head-to-head results show the NVIDIA L40S winning both tests, 2 wins, 0 for the AMD. The AMD does not beat the NVIDIA in any benchmark.

Q: How do the memory configurations differ?

A:** The NVIDIA has 48 GB GDDR6 on a 384-bit bus with 864.0 GB/s of bandwidth. The AMD has 32 GB HBM2 on a 4096-bit bus, with 1.02 TB/s of bandwidth. The NVIDIA bus is narrower but the memory runs faster, netting it 18% more bandwidth.

Q: What is the thermal and power profile?

A:** The NVIDIA L40S is rated at 300 W with a 700 W power supply recommended. The AMD Radeon Pro Vega II Duo has a 475W TDP, with an 850W power supply recommended. The NVIDIA is a dual-slot card, while the AMD is a quad-slot card.

The Verdict

The database's verdict is straightforward, and it is a resounding one. Based on the recorded data, the NVIDIA L40S is the part to buy for anyone who needs the best compute performance. It wins both head-to-head matchups by enormous margins (236% and 162.6%), sits at the 99th percentile of all GPUs, and offers a feature set (tensor cores, ray tracing cores, 48 GB of memory, modern APIs) that the AMD part simply does not have. The L40S is in the class of the NVIDIA H200 NVL, RTX 6000 Ada, and AMD Instinct MI300X.

The AMD Radeon R980 Pro Vega II Duo is a different beast. It is the only Mac-oriented part in this comparison, with an Apple MPX bus and Thunderbolt outputs, and it holds its own among other AMD parts. For a workstation or a Mac environment where the ecosystem and API support matter more than raw compute, the Vega II Duo makes sense. But if the goal is raw compute, and the database shows the L40S is the clear and overwhelming pick, the AMD part is a niche product for a different use case.

Specification Differences

| Feature | NVIDIA L40S | AMD Radeon Pro Vega II Duo |

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

| Chip | AD102 | Vega 20 |

| Architecture | Ada Lovelace | GCN 5.1 |

| Process | 5 nm | 7 nm |

| Foundry | TSMC | TSMC |

| Transistors | 76,300 million | 13,230 million |

| Die size | 609 mm² | 331 mm² |

| Transistor density | 125.3M / mm² | 40.0M / mm² |

| Base clock | 1110 MHz | 1400 MHz |

| Boost clock | 2520 MHz | 1720 MHz |

| Memory | 48 GB GDDR6 | 32 GB HBM2 |

| Memory bus | 384 bit | 4096 bit |

| Memory bandwidth | 864.0 GB/s | 1.02 TB/s |

| Shading units | 18176 | 4096 |

| Tensor cores | 568 | N/A |

| RT cores | 142 | N/A |

| Texture units | 568 | 256 |

| ROPs | 192 | 64 |

| Pixel rate | 483.8 GPixel/s | 110.1 GPixel/s |

| Texture rate | 1,431.4 GTexel/s | 440.3 GTexel/s |

| FP32 | 91.61 TFLOPS | 14.09 TFLOPS |

| FP16 | 91.61 TFLOPS (1:1) | 28.18 TFLOPS (2:1) |

| TDP | 300 W | 475 W |

| Slot width | Dual-slot | Quad-slot |

| Power connectors | 1x 16-pin | None (Apple MPX) |

| Recommended PSU | 700 W | 850 W |

| Bus interface | PCIe 4.0 x16 | Apple MPX |

| Display outputs | 1x HDMI 2.1, 3x DisplayPort 1.4a | 1x HDMI 2.0b, 4x Thunderbolt |

| DirectX | 12 Ultimate (12_2) | 12 (12_1) |

| OpenGL | 4.6 | 4.6 |

| Vulkan | 1.4 | 1.3 |

| Length | 267 mm | Not listed |

| Height | 111 mm | Not listed |

| Release date | 2022-10-12 | 2019-06-02 |

| Production status | End-of-life | End-of-life |

| Predecessor | Server Ampere | None |

| Successor | Server Hopper | None |

| Launch MSRP | None | 4.399 USD (stated once as launch MSRP) |