NVIDIA L40 vs NVIDIA Tesla V100S PCIe 32 GB Comparison

# Head-to-Head Benchmarks

The benchmark data shows a single head-to-head comparison between the NVIDIA L40 and the NVIDIA Tesla V100S PCIe 32 GB, with the L40 emerging as the clear victor. In Geekbench OpenCL, the L40 scores 330,926 against the V100S's 194,415, a delta of 70.2% in favor of the newer card. This is not a marginal improvement; it represents a substantial generational leap in raw compute throughput.

Contextualizing the L40's score against its nearest rivals reinforces its standing. The L40's average benchmark score of 284,111 places it 1.1% ahead of the NVIDIA RTX 6000 Ada Generation (287,237), 3.9% behind the NVIDIA L40S (295,763), and 13.1% ahead of the NVIDIA L20 (251,147). Notably, the L40 trails the AMD Instinct MI300X by 10.7% (317,994), but that rival sits in a different performance tier entirely. The L40's percentile rank of 99 among all GPUs places it at the very top of the distribution.

Meanwhile, the Tesla V100S's average benchmark score of 194,415 tells a different story. Its nearest rivals cluster tightly around it: the NVIDIA RTX 6000D scores 195,964 (0.8% higher), the NVIDIA A100 SXM4 40 GB scores 187,147 (3.9% lower), the NVIDIA RTX 5000 Ada Generation scores 184,664 (5.3% lower), and the NVIDIA A100 SXM4 80 GB scores 183,725 (5.8% lower). The V100S holds a 98th percentile rank, which is still elite, but its absolute score is only 58.7% of the L40's OpenCL result.

The 70.2% delta in OpenCL is the only direct benchmark comparison available, but it is decisive. The L40's FP32 compute of 90.52 TFLOPS versus the V100S's 16.35 TFLOPS explains the gap: the L40 delivers more than 5.5 times the single-precision throughput. Even accounting for the V100S's FP16 advantage (32.71 TFLOPS at 2:1 ratio versus the L40's 90.52 TFLOPS at 1:1), the L40 still leads by a factor of 2.8 in half-precision work.

# Architecture Differences

The two cards represent entirely different architectural eras. The L40 is built on Ada Lovelace (chip AD102), fabricated on TSMC's 5 nm process, while the V100S uses Volta (chip GV100) on a 12 nm process. This node disparity is stark: the L40 packs 76,300 million transistors into a 609 mm² die, yielding a transistor density of 125.3M per mm². The V100S, by contrast, contains 21,100 million transistors spread across a larger 815 mm² die, resulting in a density of just 25.9M per mm². The L40 crams nearly five times more transistors per square millimeter.

The memory subsystems diverge completely. The L40 uses 48 GB of GDDR6 on a 384-bit bus, delivering 864.0 GB/s of bandwidth. The V100S uses 32 GB of HBM2 on a massive 4096-bit bus, achieving 1.13 TB/s — the V100S actually wins on raw bandwidth by 30.8%. However, the L40's memory operates at 18 Gbps effective (2250 MHz) versus the V100S's 2.2 Gbps effective (1107 MHz), showing how the L40 compensates with much faster signaling.

Core configurations differ dramatically. The L40 has 18,176 shading units, 568 TMUs, and 192 ROPs, alongside 142 RT cores and 568 tensor cores. The V100S has 5,120 shading units, 320 TMUs, and 128 ROPs, with no RT cores and 640 tensor cores. The L40's shading unit count is 3.55 times higher, and its TMU count is 1.78 times higher. Interestingly, the V100S actually has more tensor cores (640 vs 568), though the L40's tensor cores are architecturally newer.

The L40 supports DirectX 12 Ultimate (12_2) and Vulkan 1.4, while the V100S is limited to DirectX 12 (12_1) with the same Vulkan 1.4 support. Both cards support OpenGL 4.6. The L40 also includes 4x DisplayPort 1.4a outputs, whereas the V100S has no display outputs at all, reflecting its compute-only design.

# FAQ

Q: Which card has higher FP32 compute performance?

A: The NVIDIA L40 delivers 90.52 TFLOPS FP32, which is 5.54 times the Tesla V100S's 16.35 TFLOPS. This is the primary driver of the L40's 70.2% OpenCL benchmark advantage.

Q: Does the Tesla V100S have any memory advantage?

A: Yes. The V100S provides 1.13 TB/s of bandwidth via HBM2 on a 4096-bit bus, which is 30.8% higher than the L40's 864.0 GB/s from GDDR6 on a 384-bit bus. However, the L40 has 50% more capacity (48 GB vs 32 GB).

Q: How do the two cards compare in ray tracing support?

A: The L40 includes 142 RT cores, enabling hardware-accelerated ray tracing. The V100S has no RT cores (null in the data), meaning it lacks dedicated ray tracing hardware entirely.

Q: What are the power requirements?

A: The L40 has a TDP of 300 W with a suggested PSU of 700 W and a single 16-pin connector. The V100S has a TDP of 250 W with a suggested PSU of 600 W and dual 8-pin connectors. Both are dual-slot cards.

Q: Which card is more recent?

A: The L40 was released on 2022-10-12, while the V100S was released on 2019-11-25, making the L40 roughly three years newer. Both are now end-of-life products.

Q: How does the L40 compare to its closest rival, the RTX 6000 Ada Generation?

A: The L40's average benchmark score of 284,111 is 1.1% lower than the RTX 6000 Ada Generation's 287,237. These cards are effectively performance peers, with the L40S (295,763) sitting 3.9% above the L40.

# Specification Differences

| Specification | NVIDIA L40 | NVIDIA Tesla V100S PCIe 32 GB |

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

| Architecture | Ada Lovelace | Volta |

| 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 | 1245 MHz |

| Boost Clock | 2490 MHz | 1597 MHz |

| Memory Clock | 2250 MHz / 18 Gbps effective | 1107 MHz / 2.2 Gbps effective |

| Memory Size | 48 GB | 32 GB |

| Memory Type | GDDR6 | HBM2 |

| Memory Bus Width | 384 bit | 4096 bit |

| Memory Bandwidth | 864.0 GB/s | 1.13 TB/s |

| Shading Units | 18,176 | 5,120 |

| TMUs | 568 | 320 |

| ROPs | 192 | 128 |

| RT Cores | 142 | None |

| Tensor Cores | 568 | 640 |

| Pixel Rate | 478.1 GPixel/s | 204.4 GPixel/s |

| Texture Rate | 1,414.3 GTexel/s | 511.0 GTexel/s |

| FP32 | 90.52 TFLOPS | 16.35 TFLOPS |

| FP16 | 90.52 TFLOPS (1:1) | 32.71 TFLOPS (2:1) |

| TDP | 300 W | 250 W |

| Power Connectors | 1x 16-pin | 2x 8-pin |

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

| Release Date | 2022-10-12 | 2019-11-25 |

| Dimensions | 267 mm x 111 mm | Not specified |

# The Verdict

The data points to a decisive conclusion: the NVIDIA L40 is the superior performer in virtually every compute metric that matters for modern workloads. Its 70.2% OpenCL lead, 5.54x FP32 advantage, 4.4x texture rate, and 2.34x pixel rate leave no ambiguity. The L40 also brings architectural modernity with RT cores, DirectX 12 Ultimate support, and PCIe 4.0, none of which the V100S can match.

The V100S retains exactly two advantages: memory bandwidth (1.13 TB/s vs 864.0 GB/s) and tensor core count (640 vs 568). Neither compensates for the L40's overwhelming compute superiority. The V100S's higher base clock (1245 MHz vs 735 MHz) is irrelevant given the L40's much higher boost clock (2490 MHz vs 1597 MHz).

For buyers choosing between these two end-of-life server cards, the L40 is the rational pick for any compute-intensive task. Its 48 GB GDDR6 capacity also exceeds the V100S's 32 GB HBM2, making it more suitable for large model inference. The V100S's higher bandwidth might benefit memory-bound workloads, but the L40's 50% capacity advantage and 5.5x compute headroom make it the better all-rounder.

# Where Each One Wins

NVIDIA L40 wins on:

• Raw compute: 90.52 TFLOPS FP32 versus 16.35 TFLOPS, a 5.54x margin.
• Texture and pixel throughput: 1,414.3 GTexel/s and 478.1 GPixel/s versus 511.0 GTexel/s and 204.4 GPixel/s.
• Memory capacity: 48 GB versus 32 GB, a 50% increase.
• Modern features: RT cores (142), DirectX 12 Ultimate, PCIe 4.0, and display outputs (4x DisplayPort 1.4a).
• Benchmark performance: 70.2% ahead in Geekbench OpenCL.

NVIDIA Tesla V100S PCIe 32 GB wins on:

• Memory bandwidth: 1.13 TB/s versus 864.0 GB/s, a 30.8% advantage.
• Tensor core count: 640 versus 568, a 12.7% edge.
• Power efficiency per TDP: 250 W versus 300 W, though the L40 delivers far more performance per watt.
• Lower system power requirement: 600 W suggested PSU versus 700 W.

The V100S also has a higher base clock (1245 MHz vs 735 MHz), but this is offset by the L40's substantially higher boost clock. For workloads heavily dependent on memory bandwidth — such as certain sparse matrix operations or data movement patterns — the V100S could theoretically perform competitively. However, the L40's 5.5x FP32 advantage means any compute-bound task will favor it overwhelmingly. The V100S's edge in tensor core count is similarly muted because the L40's tensor cores are architecturally newer and operate at much higher throughput. The practical conclusion is that the L40 is the default choice for modern workloads, while the V100S remains a niche option for legacy software tuned specifically to Volta's HBM2 memory layout.