NVIDIA L40S vs NVIDIA Tesla V100 PCIe 32 GB Comparison

NVIDIA L40S and NVIDIA Tesla V100 PCIe 32 GB represent two distinct eras of NVIDIA's data center GPU lineup, with the L40S built on the Ada Lovelace architecture and the V100 based on the older Volta architecture. The benchmark data shows a decisive performance gap between the two, with the L40S leading in every recorded test, though the V100 retains relevance in specific legacy workloads due to its unique memory technology and high tensor core count.

Head-to-Head Benchmarks

The Geekbench OpenCL test shows the NVIDIA L40S scoring 330,727 points against the Tesla V100's 168,763 points, resulting in a 96% delta in favor of the newer card. This is the larger of the two benchmark gaps, reflecting the L40S's substantial advantage in raw compute throughput. In Geekbench Vulkan, the L40S scores 260,799 while the V100 manages 131,847, a 97.8% delta — the closest margin between the two, yet still a commanding lead for the L40S.

The average benchmark score reinforces this hierarchy. The L40S posts an average of 295,763 across all tests, while the V100 averages 150,305. This places the L40S in the 99th percentile of all GPUs, whereas the V100 sits in the 96th percentile. The gap between the two cards is roughly the same as the difference between the V100 and its nearest rival, the AMD Instinct MI100, which trails the V100 by 8.1%.

Context from the nearest rivals list provides additional perspective. The L40S's average score of 295,763 puts it 3% ahead of the NVIDIA RTX 6000 Ada Generation (287,237) and 4.1% ahead of the NVIDIA L40 (284,111). However, it falls 7% behind the AMD Instinct MI300X (317,994) and 11.7% behind the NVIDIA H200 NVL (334,891). The V100's average of 150,305 is 1.1% below the NVIDIA A10G (151,963) and 6.5% below the AMD Radeon Pro W6800X (160,671), while sitting 8.1% above the AMD Instinct MI100 (139,035).

Where Each One Wins

The L40S wins both head-to-head benchmark tests, giving it a clean 2-0 record in this comparison. Its strength lies in raw compute performance, driven by a much higher FP32 throughput of 91.61 TFLOPS compared to the V100's 14.13 TFLOPS. The L40S also delivers superior graphics-oriented capabilities, as evidenced by its Vulkan score being nearly double that of the V100, and it supports modern APIs including DirectX 12 Ultimate (12_2) and Vulkan 1.4.

The Tesla V100's wins are more situational and derived from architectural characteristics rather than benchmark scores. It features 640 tensor cores, which is 72 more than the L40S's 568, suggesting potential advantages in certain dense matrix operations despite its lower overall FP32 throughput. The V100's HBM2 memory, with a 4096-bit bus width, provides 897.0 GB/s of bandwidth — slightly higher than the L40S's 864.0 GB/s from its 384-bit GDDR6 interface. This bandwidth advantage could benefit memory-bound workloads that fit within the V100's 32 GB capacity.

The V100 also presents a lower power profile at 250 W TDP compared to the L40S's 300 W, and it requires a 600 W suggested PSU versus the L40S's 700 W suggestion. For deployments constrained by power delivery or thermal envelopes, the V100 offers a more modest footprint, though the L40S compensates with significantly higher performance per watt in the benchmark data.

The Verdict

The data clearly favors the NVIDIA L40S for virtually all modern workloads. Its 96% lead in OpenCL and 97.8% lead in Vulkan represent overwhelming performance advantages that translate directly to reduced processing times and higher throughput. The L40S's 48 GB of GDDR6 memory, while offering lower raw bandwidth than the V100's HBM2, provides 50% more capacity, which is critical for large models and datasets that exceed 32 GB.

The L40S should be the choice for organizations running compute-intensive applications, machine learning training, or real-time ray tracing workloads, given its 142 RT cores (the V100 has none) and support for DirectX 12 Ultimate. Its 99th percentile ranking among all GPUs, combined with a 4.1% lead over the NVIDIA L40 and a 3% edge over the RTX 6000 Ada Generation, positions it as a top-tier accelerator within its immediate competitive set.

The Tesla V100 PCIe 32 GB remains a viable option only for specific scenarios where its unique characteristics align with workload requirements. Its 640 tensor cores could be advantageous for certain inference tasks or older deep learning frameworks optimized for Volta's tensor core design. The slightly higher memory bandwidth (897.0 GB/s vs 864.0 GB/s) may benefit certain memory-bound kernels, and the lower 250 W TDP makes it easier to deploy in power-constrained environments. However, with an average benchmark score that is roughly half of the L40S, the V100's use cases are increasingly narrow, and its 96th percentile ranking — while still strong — reflects its age relative to modern accelerators.

FAQ

Q: How much faster is the NVIDIA L40S compared to the Tesla V100 in OpenCL?

A: The L40S scores 330,727 in Geekbench OpenCL versus 168,763 for the V100, a 96% performance delta in favor of the L40S.

Q: Does the Tesla V100 have any benchmark where it beats the L40S?

A: No. In the two head-to-head benchmark tests (Geekbench OpenCL and Geekbench Vulkan), the L40S wins both, with the V100 trailing by 96% and 97.8% respectively.

Q: What is the memory capacity difference between the two cards?

A: The L40S has 48 GB of GDDR6 memory, while the V100 has 32 GB of HBM2 memory. The L40S provides 50% more capacity, but the V100 offers slightly higher bandwidth at 897.0 GB/s versus 864.0 GB/s.

Q: Which card has more tensor cores?

A: The Tesla V100 has 640 tensor cores, which is 72 more than the L40S's 568 tensor cores. However, the L40S's newer architecture delivers far higher overall FP32 performance.

Q: How do the two cards compare in terms of power requirements?

A: The V100 has a 250 W TDP and suggests a 600 W PSU, while the L40S has a 300 W TDP and suggests a 700 W PSU. The L40S uses a single 16-pin power connector, whereas the V100 uses two 8-pin connectors.

Q: What is the direct comparison between the L40S and its closest rival?

A: The L40S's average benchmark score of 295,763 is 3% higher than the NVIDIA RTX 6000 Ada Generation (287,237) and 4.1% higher than the NVIDIA L40 (284,111), but it trails the AMD Instinct MI300X by 7% and the NVIDIA H200 NVL by 11.7%.

Architecture Differences

The architectural divide between these two GPUs is generational. The L40S uses the AD102 chip on a 5 nm process from TSMC, packing 76,300 million transistors into a 609 mm² die, achieving a transistor density of 125.3 million per square millimeter. The V100 uses the GV100 chip on a 12 nm process, containing 21,100 million transistors on a larger 815 mm² die, yielding a density of just 25.9 million per square millimeter. This density difference explains the L40S's massive performance advantage despite its smaller physical footprint.

The L40S's Ada Lovelace architecture brings 18,176 shading units, 568 TMUs, and 192 ROPs, alongside 142 RT cores for ray tracing — a feature entirely absent from the V100. The V100's Volta architecture provides 5,120 shading units, 320 TMUs, and 128 ROPs, with 640 tensor cores compared to the L40S's 568. The L40S pairs its FP32 throughput of 91.61 TFLOPS with equal FP16 performance (91.61 TFLOPS at a 1:1 ratio), whereas the V100's FP32 of 14.13 TFLOPS doubles to 28.26 TFLOPS in FP16 (2:1 ratio).

Memory architectures differ fundamentally. The L40S uses 48 GB of GDDR6 on a 384-bit bus, while the V100 uses 32 GB of HBM2 on a 4096-bit bus. The L40S's memory operates at 2250 MHz (18 Gbps effective), whereas the V100's runs at 876 MHz (1752 Mbps effective). The L40S supports PCIe 4.0 x16, while the V100 is limited to PCIe 3.0 x16. The L40S includes display outputs (1x HDMI 2.1 and 3x DisplayPort 1.4a), while the V100 has no display outputs, reflecting its pure compute orientation.

Specification Differences

The two cards diverge on nearly every measurable specification. The L40S has a base clock of 1110 MHz and a boost clock of 2520 MHz, while the V100 runs at 1230 MHz base and 1380 MHz boost — the V100's higher base clock is overshadowed by the L40S's much higher boost capability. The L40S achieves a pixel rate of 483.8 GPixel/s and a texture rate of 1,431.4 GTexel/s, compared to the V100's 176.6 GPixel/s and 441.6 GTexel/s.

Power and physical specifications show the L40S at 300 W TDP with a dual-slot design, measuring 267 mm in length and 111 mm in height, using a single 16-pin power connector. The V100 also uses a dual-slot design at 250 W TDP but requires two 8-pin connectors, and its dimensions are not listed in the available data. The L40S was released on 2022-10-12, while the V100 launched earlier on 2018-03-26; both are now end-of-life products. The L40S's predecessor is Server Ampere and its successor is Server Hopper, while the V100's predecessor is Tesla Pascal and its successor is Tesla Turing. In API support, the L40S offers DirectX 12 Ultimate (12_2) compared to the V100's DirectX 12 (12_1), with both supporting OpenGL 4.6 and Vulkan 1.4.