NVIDIA L40 vs NVIDIA Tesla V100 SXM2 16 GB Comparison

Head-to-Head Benchmarks

The recorded benchmark data shows a decisive overall advantage for the NVIDIA L40 in the only two tests where both accelerators have comparable scores. Across the two head-to-head tests, the L40 secures 2 wins, while the Tesla V100 SXM2 16 GB records no wins. The largest margin appears in the Geekbench OpenCL test, where the L40 scores 330,926 against the V100's 87,454, a striking 278.4% difference. This is not a marginal improvement; it is a generational leap in raw compute throughput, reflecting the substantial architectural and process node advancements between the two products.

In the Geekbench Vulkan test, the L40 again leads, scoring 237,295 versus 141,336 for the V100, a 67.9% advantage. While this margin is smaller than the OpenCL gap, it still places the L40 firmly ahead in graphics and compute API workloads. The delta percentages confirm that the L40's superiority is consistent across both APIs, but the larger OpenCL gap suggests that the L40's advantages are most pronounced in compute-heavy, non-graphics workloads, which aligns with its modern architecture and significantly higher FP32 throughput.

The average benchmark score, which combines all recorded tests, reinforces this disparity. The L40 achieves an average score of 284,111, while the V100 SXM2 16 GB manages 114,395. This places the L40 in the 99th percentile of all GPUs in the database, while the V100 sits at the 95th percentile. Although both are high-performing parts, the L40's average score is roughly 148% higher, a difference that separates an older, albeit capable, accelerator from a modern, high-end server part.

Where Each One Wins

Based on the available benchmark wins, the NVIDIA L40 is the clear winner in every measured category. The data shows it wins both the OpenCL and Vulkan tests, which are representative of general-purpose compute and graphics API performance. For users prioritizing compute-heavy tasks such as AI training, scientific simulation, or rendering that leverage OpenCL, the L40's 278.4% lead in that specific test is a definitive factor. Similarly, for applications relying on Vulkan for graphics or compute, the L40's 67.9% lead ensures a substantial performance buffer.

The Tesla V100 SXM2 16 GB, despite being an end-of-life product, still holds a place in the 95th percentile of all GPUs. Its strengths lie not in winning these specific benchmarks, but in its historical role and its specific hardware features. The V100's HBM2 memory, with a 4096-bit bus and 897.0 GB/s bandwidth, provides a different memory architecture than the L40's GDDR6 solution. While the L40's 864.0 GB/s bandwidth is close, the V100's HBM2 design historically offered advantages in certain memory-latency-sensitive workloads. However, in the recorded database results, this does not translate into a benchmark win. The V100 also offers FP16 performance at a 2:1 ratio (31.33 TFLOPS), whereas the L40 offers FP16 at a 1:1 ratio (90.52 TFLOPS). This means the L40's FP16 performance is significantly higher, making it the superior choice for mixed-precision AI workloads. Ultimately, the data indicates that for any workload represented by these two benchmarks, the L40 is the better performer.

Architecture Differences

The NVIDIA L40 and Tesla V100 SXM2 16 GB are products of different architectural eras, which explains their significant performance disparity. The L40 is built on the Ada Lovelace architecture, using the AD102 chip, fabricated on a 5 nm process at TSMC. This advanced node allows for a massive transistor count of 76,300 million on a 609 mm² die, yielding a transistor density of 125.3M per mm². In contrast, the V100 uses the Volta architecture with the GV100 chip, built on a 12 nm process, also at TSMC. This older node houses 21,100 million transistors on a larger 815 mm² die, resulting in a much lower density of 25.9M per mm². The L40's modern node is a primary driver of its higher clock speeds and efficiency.

The compute core configurations differ drastically. The L40 features 18,176 shading units, 568 texture mapping units (TMUs), and 192 render output units (ROPs). It also includes 142 ray tracing (RT) cores and 568 tensor cores. The V100, by contrast, has 5,120 shading units, 320 TMUs, and 128 ROPs, with no RT cores and 640 tensor cores. The L40's higher shading unit count and inclusion of RT cores give it a substantial lead in raw shader throughput and enable real-time ray tracing, a feature the V100 lacks entirely. The tensor core count is higher on the V100 (640 vs. 568), but the L40's newer tensor core design and higher FP32 throughput more than compensate.

Memory subsystems also differ. The L40 uses 48 GB of GDDR6 memory on a 384-bit bus, delivering 864.0 GB/s of bandwidth. The V100 uses 16 GB of HBM2 memory on a 4096-bit bus, delivering 897.0 GB/s. While the V100 has a slight bandwidth advantage, the L40 offers three times the memory capacity, which is critical for large models and datasets that cannot fit in the V100's 16 GB frame buffer. Clock speeds further separate them: the L40 has a base clock of 735 MHz and a boost clock of 2490 MHz, while the V100 has a higher base clock of 1312 MHz but a much lower boost clock of 1530 MHz. The L40's higher boost clock, enabled by the 5 nm process, is a key factor in its performance lead.

FAQ

Q: Which GPU has a higher average benchmark score?

A: The NVIDIA L40 has a significantly higher average benchmark score of 284,111, compared to the Tesla V100 SXM2 16 GB's 114,395. This places the L40 in the 99th percentile of all GPUs, while the V100 sits in the 95th percentile.

Q: How large is the performance gap in the Geekbench OpenCL test?

A: The L40 wins the OpenCL test with a score of 330,926 against the V100's 87,454, representing a 278.4% advantage for the L40. This is the single largest performance delta recorded in the head-to-head comparison.

Q: Does the Tesla V100 SXM2 16 GB have any advantages in memory bandwidth?

A: Yes, the V100's HBM2 memory provides 897.0 GB/s of bandwidth on a 4096-bit bus, which is slightly higher than the L40's 864.0 GB/s on a 384-bit bus. However, the L40 offers 48 GB of memory compared to the V100's 16 GB.

Q: What is the difference in FP32 (single-precision) performance?

A: The NVIDIA L40 delivers 90.52 TFLOPS of FP32 performance, while the Tesla V100 SXM2 16 GB delivers 15.67 TFLOPS. This represents a massive difference, with the L40 offering over five times the single-precision throughput.

Q: Which GPU supports real-time ray tracing?

A: The NVIDIA L40 supports real-time ray tracing with its 142 dedicated RT cores. The Tesla V100 SXM2 16 GB does not have any RT cores, so it lacks this hardware capability entirely.

Q: What are the production statuses of these two GPUs?

A: Both the NVIDIA L40 and the Tesla V100 SXM2 16 GB are listed as end-of-life products in the database. The L40 was released in October 2022, while the V100 was released in June 2017.

Specification Differences

The following specifications differ between the NVIDIA L40 and the Tesla V100 SXM2 16 GB.

| Specification | NVIDIA L40 | NVIDIA Tesla V100 SXM2 16 GB |

| :--- | :--- | :--- |

| Architecture | Ada Lovelace | Volta |

| Chip | AD102 | GV100 |

| Process Node | 5 nm | 12 nm |

| Transistors | 76,300 million | 21,100 million |

| Die Size | 609 mm² | 815 mm² |

| Transistor Density | 125.3M / mm² | 25.9M / mm² |

| Base Clock | 735 MHz | 1312 MHz |

| Boost Clock | 2490 MHz | 1530 MHz |

| Memory Clock | 2250 MHz (18 Gbps effective) | 876 MHz (1752 Mbps effective) |

| Memory Size | 48 GB | 16 GB |

| Memory Type | GDDR6 | HBM2 |

| Memory Bus Width | 384 bit | 4096 bit |

| Memory Bandwidth | 864.0 GB/s | 897.0 GB/s |

| Shading Units | 18176 | 5120 |

| TMUs | 568 | 320 |

| ROPs | 192 | 128 |

| RT Cores | 142 | None |

| Tensor Cores | 568 | 640 |

| Pixel Rate | 478.1 GPixel/s | 195.8 GPixel/s |

| Texture Rate | 1,414.3 GTexel/s | 489.6 GTexel/s |

| FP32 Performance | 90.52 TFLOPS | 15.67 TFLOPS |

| FP16 Performance | 90.52 TFLOPS (1:1) | 31.33 TFLOPS (2:1) |

| TDP | 300 W | 250 W |

| Slot Width | Dual-slot | SXM Module |

| Power Connectors | 1x 16-pin | None |

| Suggested PSU | 700 W | 600 W |

| Bus Interface | PCIe 4.0 x16 | PCIe 3.0 x16 |

| Display Outputs | 4x DisplayPort 1.4a | No outputs |

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

| Dimensions (Length) | 267 mm (10.5 inches) | Not specified |

| Dimensions (Height) | 111 mm (4.4 inches) | Not specified |

| Release Date | 2022-10-12 | 2017-06-20 |

| Predecessor | Server Ampere | Tesla Pascal |

| Successor | Server Hopper | Tesla Turing |