NVIDIA L40S vs NVIDIA Tesla P100 PCIe 16 GB Comparison

FAQ

Q: How large is the performance gap between the NVIDIA L40S and the NVIDIA Tesla P100 PCIe 16 GB in the recorded database?

A: In the only shared benchmark, Geekbench OpenCL, the L40S scores 330,727 against 79,605 for the P100. That is a delta of 315.5%, meaning the L40S is roughly four times faster in this workload.

Q: Which GPU sits higher in the overall performance percentile ranking?

A: The L40S is in the 99th percentile among all GPUs, while the Tesla P100 PCIe 16 GB is in the 92nd percentile. The L40S also has an average benchmark score of 295,763 across two tests, whereas the P100's average is 79,605 from a single test.

Q: What kind of memory and capacity do these cards use?

A: The L40S uses 48 GB of GDDR6 on a 384-bit bus with 864.0 GB/s of bandwidth. The Tesla P100 PCIe 16 GB uses 16 GB of HBM2 on a 4096-bit bus with 732.2 GB/s of bandwidth.

Q: Are there any architectural features present on one card but absent on the other?

A: Yes. The L40S has 142 ray tracing cores and 568 tensor cores. The Tesla P100 has neither, as it is built on the Pascal architecture which predates both.

Q: Which card has a higher boost clock and what are the respective FP32 figures?

A: The L40S boosts to 2520 MHz and delivers 91.61 TFLOPS FP32. The P100 boosts to 1329 MHz and delivers 9.526 TFLOPS FP32. The L40S also matches its FP32 rate in FP16 at 91.61 TFLOPS (1:1), while the P100's FP16 is 19.05 TFLOPS (2:1).

Q: What is the launch MSRP of the Tesla P100 PCIe 16 GB, and does the L40S have one in the database?

A: The Tesla P100 PCIe 16 GB has a launch MSRP of 5,699 USD. The L40S entry does not include a launch MSRP field.

The Verdict

The data is unambiguous. For any compute workload represented by the Geekbench OpenCL test, the NVIDIA L40S is the superior choice by a wide margin. Its score of 330,727 is 315.5% higher than the P100's 79,605. The L40S also sits at the 99th percentile of all GPUs, versus the 92nd percentile for the P100. If the task is modern AI inference, ray tracing, or high-throughput FP32 compute, the L40S is the only rational pick.

The Tesla P100 PCIe 16 GB remains relevant only for legacy environments where Pascal-specific code paths or its HBM2 memory interface are required. Its 16 GB capacity is a quarter of the L40S's 48 GB, and its FP32 throughput is an order of magnitude lower. The P100 has no ray tracing or tensor core support, so any workload relying on those features cannot run on it at all.

For buyers with existing Pascal-based infrastructure who cannot migrate software stacks, the P100's 92nd percentile ranking shows it is still far from obsolete in raw OpenCL terms, beating the RTX 5090 by 0.3% in its nearest rival list. But for new deployments, the L40S's 99th percentile placement and 315.5% lead in the head-to-head test make the decision trivial.

Head-to-Head Benchmarks

The only recorded head-to-head test is Geekbench OpenCL. The L40S scores 330,727, while the P100 scores 79,605. The delta is 315.5% in favor of the L40S. This is not a marginal improvement; it is a generational leap. To put it in context, the L40S's nearest rivals include the H200 NVL at 334,891 (11.7% higher) and the MI300X at 317,994 (7% lower). The P100's nearest rivals are much closer to its own score, with the RTX 5090 at 79,842 (0.3% higher) and the Pro Vega 64X at 80,959 (1.7% higher).

In practical terms, the L40S's OpenCL result is 4.15 times the P100's result. That means a task taking one hour on the P100 would take roughly 14.5 minutes on the L40S, assuming perfect scaling. The L40S also has a second benchmark, Geekbench Vulkan, scoring 260,799, which the P100 does not have recorded. The L40S's average benchmark score across both tests is 295,763, while the P100's average is exactly its single OpenCL score of 79,605.

The win count is 1 for the L40S and 0 for the P100. There is no test in the database where the P100 outperforms the L40S. The closest the P100 comes is in its own rival group, where it trades blows with the RTX 5090, but that is irrelevant when comparing directly to the L40S.

Specification Differences

The two cards differ in nearly every measurable specification. The L40S uses 76,300 million transistors on a 609 mm² die, while the P100 uses 15,300 million transistors on a 610 mm² die. The die sizes are nearly identical, but the transistor density is 125.3M per mm² for the L40S versus 25.1M per mm² for the P100. This is a direct consequence of the process node: the L40S is on 5 nm, the P100 on 16 nm, both from TSMC.

Clock speeds diverge significantly. The L40S has a base clock of 1110 MHz and a boost of 2520 MHz. The P100 has a base of 1190 MHz and a boost of 1329 MHz. The P100 actually starts at a higher base clock, but the L40S's boost is 89.6% higher. Memory clocks also differ: the L40S runs at 2250 MHz with 18 Gbps effective, while the P100 runs at 715 MHz with 1430 Mbps effective.

Memory capacity and bandwidth favor the L40S. It has 48 GB of GDDR6 on a 384-bit bus, yielding 864.0 GB/s. The P100 has 16 GB of HBM2 on a 4096-bit bus, yielding 732.2 GB/s. The P100's bus is wider, but the L40S still achieves 18% more bandwidth due to faster memory clocks.

Compute unit counts are starkly different. The L40S has 18,176 shading units, 568 TMUs, and 192 ROPs. The P100 has 3,584 shading units, 224 TMUs, and 96 ROPs. The L40S has 5.07 times the shading units and 2.54 times the TMUs. Pixel rate is 483.8 GPixel/s for the L40S versus 127.6 GPixel/s for the P100. Texture rate is 1,431.4 GTexel/s versus 297.7 GTexel/s.

Power and connectivity also differ. The L40S has a TDP of 300 W with a 1x 16-pin connector and a suggested PSU of 700 W. The P100 has a TDP of 250 W with a 1x 8-pin connector and a suggested PSU of 600 W. The L40S uses PCIe 4.0 x16, while the P100 uses PCIe 3.0 x16. The L40S has display outputs (1x HDMI 2.1, 3x DisplayPort 1.4a), while the P100 has no outputs.

Architecture Differences

The L40S is built on the Ada Lovelace architecture with the AD102 chip, classified in the Server Ada (Lxx) generation. The P100 is built on the Pascal architecture with the GP100 chip, classified in the Tesla Pascal (Pxx) generation. The L40S was released on 2022-10-12, while the P100 was released on 2016-06-19. The L40S's predecessor is Server Ampere and its successor is Server Hopper. The P100's predecessor is Tesla Maxwell and its successor is Tesla Volta.

The most critical architectural difference is the presence of dedicated hardware on the L40S. It has 142 ray tracing cores and 568 tensor cores. The P100 has none of either. This means the L40S can accelerate ray-traced rendering and tensor-based AI workloads natively, while the P100 must rely on shader-based computation for such tasks. The L40S's FP16 performance is 91.61 TFLOPS at a 1:1 ratio with FP32, which is typical of Ada Lovelace's unified compute pipeline. The P100's FP16 is 19.05 TFLOPS at a 2:1 ratio, meaning it can only reach that rate by pairing two FP32 cores, a less flexible arrangement.

API support differs as well. The L40S supports DirectX 12 Ultimate (12_2), Vulkan 1.4, and OpenGL 4.6. The P100 supports DirectX 12 (12_1), Vulkan 1.3, and OpenGL 4.6. The L40S's higher DirectX feature level and newer Vulkan version reflect its modern architecture. Both cards are marked as end-of-life in production status, but the L40S's release date is six years later, which explains the massive transistor density advantage: 125.3M per mm² versus 25.1M per mm².

The physical dimensions are similar: both are 267 mm (10.5 inches) long. The L40S is 111 mm (4.4 inches) tall, while the P100's height is not recorded. Both are dual-slot cards. The L40S's 5 nm process allows 76,300 million transistors on a die nearly identical in size to the P100's 610 mm², demonstrating how far fabrication technology advanced between 2016 and 2022.