NVIDIA L40S vs NVIDIA Tesla V100 PCIe 16 GB Comparison

NVIDIA L40S and NVIDIA Tesla V100 PCIe 16 GB are two data-center accelerators separated by five years of architecture evolution. The benchmark data places the L40S in the 99th percentile of all GPUs, while the V100 sits in the 96th percentile—a gap that translates into decisive performance differences across every tested workload. The L40S wins both head-to-head benchmark comparisons, with the largest margin appearing in Vulkan compute, where it leads by 130.7%.

Head-to-Head Benchmarks

The Geekbench OpenCL test shows the L40S scoring 330,727 points against the V100’s 163,063 points. That is a delta of 102.8%, meaning the L40S delivers more than double the OpenCL performance of the older accelerator. To contextualize this, the L40S’s nearest rival in average benchmark score is the AMD Instinct MI300X at 317,994, which the L40S trails by 7%; the V100’s nearest rival is the AMD Instinct MI100 at 139,035, which the V100 trails by 0.7%. The L40S’s OpenCL score alone exceeds the V100’s average benchmark score (138,063) by roughly 140%, underscoring how far the newer architecture has advanced.

In Geekbench Vulkan, the L40S scores 260,799 versus the V100’s 113,062, a delta of 130.7%. This is the L40S’s largest winning margin in any tested workload. The Vulkan result is particularly telling because the V100 predates broad Vulkan adoption in compute environments; its score lands below its own OpenCL result, while the L40S maintains high performance across both APIs. The L40S’s Vulkan score is 88.9% of its OpenCL score, whereas the V100’s Vulkan score is only 69.3% of its OpenCL score—suggesting the newer architecture handles diverse compute interfaces more consistently.

Aggregating the two tests, the L40S achieves an average benchmark score of 295,763, while the V100 averages 138,063. The L40S’s average is 114.2% higher. In the nearest-rival context, the L40S outperforms the NVIDIA RTX 6000 Ada Generation (287,237) by 3% and the NVIDIA L40 (284,111) by 4.1%, but falls behind the NVIDIA H200 NVL (334,891) by 11.7%. The V100, meanwhile, sits in a much tighter cluster: it leads the NVIDIA Tesla V100 SXM2 32 GB (137,731) by just 0.2%, trails the AMD Instinct MI100 by 0.7%, and beats the AMD Radeon PRO V620 (136,472) by 1.2% and the AMD Radeon Pro W6800X Duo (135,774) by 1.7%.

The per-test deltas are consistent: the L40S wins OpenCL by 102.8% and Vulkan by 130.7%. No test favors the V100. The closest the V100 comes to competitiveness is in OpenCL, where its deficit is still larger than a full doubling of performance. In Vulkan, the gap widens by nearly 28 percentage points, indicating that the L40S’s architectural advantages are amplified in newer compute APIs.

The Verdict

The data presents a clear hierarchy: the NVIDIA L40S is the superior performer in every benchmark test included in this analysis. For workloads that rely on OpenCL or Vulkan compute, the L40S delivers between 2.03x and 2.31x the performance of the V100 PCIe 16 GB. The 99th-percentile ranking of the L40S versus the 96th-percentile ranking of the V100 reinforces this gap; the L40S competes with the fastest accelerators available, while the V100 sits near the boundary of top-tier performance.

Buyers should choose the L40S if their priority is raw compute throughput and modern API support. Its benchmark scores place it within 3% of the NVIDIA RTX 6000 Ada Generation and within 4.1% of the NVIDIA L40, while remaining 11.7% behind the NVIDIA H200 NVL—a flagship that the L40S approaches but does not match. For organizations that already run V100 deployments and need to upgrade, the L40S offers a straightforward performance multiplier without requiring a shift in vendor ecosystem.

The V100 PCIe 16 GB remains a functional accelerator, but its benchmark results place it in a different performance class. Its average score of 138,063 is nearly identical to the AMD Instinct MI100 (139,035) and the NVIDIA Tesla V100 SXM2 32 GB (137,731), indicating that it is competitive with its contemporary peers. However, against the L40S, the V100’s 16 GB memory capacity and older architecture result in consistent losses. The V100 is only a rational choice in scenarios where its specific features—such as HBM2 memory or its particular tensor core design—are required, though the benchmark data does not capture those specialized workloads.

FAQ

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

A: The L40S scores 330,727 in Geekbench OpenCL, while the V100 scores 163,063. This represents a 102.8% delta, meaning the L40S is more than twice as fast.

Q: What is the average benchmark score difference between the two GPUs?

A: The L40S has an average benchmark score of 295,763, and the V100 has an average of 138,063. The L40S’s average is 114.2% higher.

Q: How does the L40S compare to its nearest rivals?

A: The L40S is 3% ahead of the NVIDIA RTX 6000 Ada Generation (287,237) and 4.1% ahead of the NVIDIA L40 (284,111), but 7% behind the AMD Instinct MI300X (317,994) and 11.7% behind the NVIDIA H200 NVL (334,891).

Q: How does the V100 compare to its nearest rivals?

A: The V100 leads the NVIDIA Tesla V100 SXM2 32 GB (137,731) by 0.2% and the AMD Radeon PRO V620 (136,472) by 1.2%, but trails the AMD Instinct MI100 (139,035) by 0.7%.

Q: Which GPU has a higher percentile ranking?

A: The L40S ranks in the 99th percentile of all GPUs, while the V100 ranks in the 96th percentile.

Q: Did the V100 win any benchmark tests?

A: No. The L40S won both head-to-head tests: Geekbench OpenCL and Geekbench Vulkan. The V100 recorded zero wins.

Specification Differences

The two GPUs diverge sharply on memory configuration. The L40S packs 48 GB of GDDR6 on a 384-bit bus, yielding 864.0 GB/s of bandwidth. The V100 offers 16 GB of HBM2 on a 4096-bit bus, achieving 897.0 GB/s of bandwidth. Despite having one-third the capacity, the V100’s HBM2 provides slightly higher bandwidth—a 3.8% advantage in raw throughput. Memory clock rates differ accordingly: the L40S runs at 2250 MHz (18 Gbps effective), while the V100 runs at 876 MHz (1752 Mbps effective).

Compute resources show a massive disparity. The L40S features 18,176 shading units, 568 TMUs, and 192 ROPs, against the V100’s 5,120 shading units, 320 TMUs, and 128 ROPs. Pixel and texture rates reflect this: the L40S achieves 483.8 GPixel/s and 1,431.4 GTexel/s, while the V100 manages 176.6 GPixel/s and 441.6 GTexel/s. The L40S’s FP32 throughput is 91.61 TFLOPS, and its FP16 throughput is also 91.61 TFLOPS (1:1 ratio). The V100 delivers 14.13 TFLOPS FP32 and 28.26 TFLOPS FP16 (2:1 ratio). The L40S also includes 142 RT cores and 568 tensor cores, while the V100 has 640 tensor cores and no RT cores.

Physical and interface specifications differ as well. The L40S uses a PCIe 4.0 x16 interface and a single 16-pin power connector; the V100 uses PCIe 3.0 x16 and dual 8-pin connectors. Both are dual-slot cards with a 300 W TDP and a 700 W suggested PSU. The L40S includes display outputs (1x HDMI 2.1, 3x DisplayPort 1.4a), whereas the V100 has no display outputs. The L40S measures 267 mm in length and 111 mm in height; V100 dimensions are not provided. The L40S supports DirectX 12 Ultimate (12_2), while the V100 supports DirectX 12 (12_1); both support OpenGL 4.6 and Vulkan 1.4.

Architecture Differences

The L40S is built on the Ada Lovelace architecture using the AD102 chip, fabricated on a 5 nm process at TSMC. It integrates 76,300 million transistors on a 609 mm² die, yielding a transistor density of 125.3M per mm². The V100 uses the Volta architecture with the GV100 chip, fabricated on a 12 nm process at TSMC. It contains 21,100 million transistors on a larger 815 mm² die, resulting in a density of 25.9M per mm²—roughly one-fifth the density of the L40S.

Clock behavior reflects the process advantage. The L40S has a base clock of 1110 MHz and a boost clock of 2520 MHz. The V100 has a base clock of 1245 MHz and a boost clock of 1380 MHz. While the V100 starts at a higher base clock, the L40S’s boost clock exceeds it by 82.6%, illustrating the newer architecture’s ability to sustain higher frequencies under load.

The L40S’s generation is listed as "Server Ada (Lxx)," with a predecessor of "Server Ampere" and a successor of "Server Hopper." The V100 belongs to "Tesla Volta (Vxx)," with a predecessor of "Tesla Pascal" and a successor of "Tesla Turing." The L40S was released on 2022-10-12, while the V100 was released on 2017-06-20. Both are marked as end-of-life in production status.

Feature differences extend to ray tracing and API support. The L40S includes 142 RT cores for hardware-accelerated ray tracing, a capability entirely absent from the V100. The L40S’s tensor core count (568) is lower than the V100’s (640), but the newer tensor cores operate within a much higher overall compute throughput. The L40S also supports DirectX 12 Ultimate, which includes features like mesh shaders and variable rate shading, whereas the V100 is limited to DirectX 12 (12_1).

Where Each One Wins

The L40S wins in every benchmark category measured here. In OpenCL, its 330,727 score represents a 102.8% advantage, making it the clear choice for general-purpose GPU compute workloads that leverage OpenCL. In Vulkan, its 260,799 score extends the lead to 130.7%, positioning it as the better option for applications that use Vulkan for compute or rendering tasks. The L40S’s 48 GB memory capacity also supports larger datasets and models than the V100’s 16 GB, which is critical for deep learning training or inference with large batch sizes.

The V100 retains a single quantitative advantage: memory bandwidth. Its 897.0 GB/s exceeds the L40S’s 864.0 GB/s by 3.8%, which could benefit memory-bandwidth-bound workloads that fit within its 16 GB capacity. However, the benchmark data does not include a test that isolates memory bandwidth, so this advantage is theoretical rather than demonstrated. The V100’s higher base clock (1245 MHz vs. 1110 MHz) also gives it a marginal edge in lightly threaded scenarios, but its lower boost clock and vastly smaller compute resource pool negate this in sustained workloads.

For users with existing V100 deployments that rely on PCIe 3.0 compatibility, the V100’s older bus interface may integrate more easily into legacy systems. The L40S’s PCIe 4.0 interface requires a compatible motherboard or riser, though it is backward compatible. The V100’s lack of display outputs is a non-issue for headless compute, but the L40S’s inclusion of HDMI and DisplayPort outputs adds flexibility for hybrid workloads that need visual output.

In summary, the L40S is the dominant choice for any modern compute workload, offering 2x to 2.3x performance across the tested benchmarks. The V100 is a legacy product whose only clear advantages are its slightly higher memory bandwidth and its compatibility with older PCIe generations. Organizations prioritizing raw throughput, larger memory, or modern API support should select the L40S; those constrained by legacy infrastructure or specific HBM2 requirements might still consider the V100, but the benchmark data offers no performance-based reason to prefer it.