NVIDIA L40S vs NVIDIA Tesla P100 PCIe 12 GB Comparison

The NVIDIA L40S and the NVIDIA Tesla P100 PCIe 12 GB represent two very different eras of NVIDIA's data-center strategy, and the recorded data reflects that gap plainly. The L40S, an Ada Lovelace part released in late 2022, sits in the 99th percentile of all GPUs in the database with an average benchmark score of 295763. The Tesla P100 PCIe 12 GB, a Pascal-based card from mid-2016, occupies the 92nd percentile with an average score of 79396. In the one direct head-to-head measurement available, the Geekbench OpenCL test, the L40S wins by 316.6 percent. Both cards are now end-of-life, and both were built for compute rather than gaming, but the data shows they serve very different performance envelopes.

The Verdict

The data makes the L40S the clear choice wherever raw throughput, memory capacity, or modern feature support matters. It delivers 91.61 TFLOPS of FP32 compute against the P100's 9.526 TFLOPS, packs 48 GB of GDDR6 against 12 GB of HBM2, and includes 142 RT cores and 568 tensor cores that the P100 simply does not have. Its 864.0 GB/s of memory bandwidth also exceeds the P100's 549.1 GB/s despite the older card's exotic HBM2 subsystem. For inference, rendering, or any workload that benefits from tensor or ray tracing hardware, the recorded data leaves no ambiguity.

The P100 still holds niche relevance. It is the more economical card to power: a 250 W TDP with a single 8-pin connector and a 600 W suggested PSU, versus 300 W, a 16-pin connector, and a 700 W suggested PSU for the L40S. It carries a launch MSRP of 4,599 USD, though that figure says nothing about current conditions. Notably, the P100 achieves double-rate FP16 at 19.05 TFLOPS thanks to its 2:1 FP16 ratio, while the L40S runs FP16 at 1:1 with FP32 at 91.61 TFLOPS, so the older card's half-precision throughput is a historical strength even though it is absolutely outclassed today. If a workload is FP16-bound and the hardware already exists in inventory, the P100 remains functional; for anything new, the L40S is the data-backed answer.

FAQ

Q: How large is the performance gap between the L40S and the Tesla P100 PCIe 12 GB?

A: In the Geekbench OpenCL head-to-head, the L40S scored 330727 against the P100's 79396, a 316.6 percent advantage. Average benchmark scores show the same story: 295763 versus 79396.

Q: Which card has more memory?

A: The L40S has 48 GB of GDDR6 on a 384-bit bus with 864.0 GB/s of bandwidth. The P100 has 12 GB of HBM2 on a 3072-bit bus with 549.1 GB/s of bandwidth. The L40S offers four times the capacity and substantially more bandwidth despite using conventional GDDR6.

Q: Do both cards support ray tracing and AI acceleration hardware?

A: No. The L40S includes 142 RT cores and 568 tensor cores. The P100 has neither, as these dedicated units did not exist in the Pascal Tesla generation.

Q: How do the two cards compare against their nearest rivals?

A: The L40S sits 3 percent ahead of the RTX 6000 Ada Generation (287237) and 4.1 percent ahead of the L40 (284111), while trailing the AMD Instinct MI300X (317994) by 7 percent and the H200 NVL (334891) by 11.7 percent. The P100 is within a tight cluster: 0.3 percent behind the Tesla P100 PCIe 16 GB (79605), 0.6 percent behind the GeForce RTX 5090 (79842), 0.6 percent ahead of the Radeon RX 6850M XT (78940), and 1.9 percent behind the Radeon Pro Vega 64X (80959).

Q: Which card has lower power requirements?

A: The P100, with a 250 W TDP, a single 8-pin power connector, and a 600 W suggested PSU. The L40S draws 300 W through a 16-pin connector with a 700 W suggested PSU.

Q: Are both cards still in production?

A: No. Both are listed as end-of-life. The L40S was released on October 12, 2022, and the P100 on June 19, 2016.

Architecture Differences

The architectural distance between these two cards is measured in process nodes, transistor budgets, and feature sets. The L40S is built on TSMC's 5 nm process, packing 76,300 million transistors into a 609 mm² die, yielding a density of 125.3M transistors per square millimeter. The P100 uses TSMC's 16 nm process with 15,300 million transistors on a 610 mm² die, a density of 25.1M per square millimeter. The dies are nearly identical in area, yet the Ada part carries roughly five times the transistors, which explains most of the performance difference before clocks are even considered.

Clock behavior differs sharply. The P100 runs a base clock of 1190 MHz with a modest boost to 1329 MHz, while the L40S starts at 1110 MHz and boosts to 2520 MHz. The Ada card nearly doubles its clock under load; the Pascal card barely moves. Memory clocks follow the same pattern in relative terms: 18 Gbps effective GDDR6 on the L40S versus 1430 Mbps effective HBM2 on the P100.

The generational lineage also differs in intent. The P100 belongs to the Tesla Pascal generation (Pxx), succeeded by Tesla Volta and preceded by Tesla Maxwell, a line focused on scientific compute with HBM2 memory and no display outputs. The L40S belongs to the Server Ada generation (Lxx), preceded by Server Ampere and succeeded by Server Hopper, and it does include display outputs: one HDMI 2.1 and three DisplayPort 1.4a, marking it as a hybrid compute-and-visualization card. The software support reflects the gap too. The L40S reports DirectX 12 Ultimate (12_2), Vulkan 1.4, and OpenGL 4.6, while the P100 reports DirectX 12 (12_1), Vulkan 1.3, and the same OpenGL 4.6.

Specification Differences

• Chip and architecture: AD102 on Ada Lovelace (L40S) versus GP100 on Pascal (P100).
• Process node: 5 nm versus 16 nm, both fabbed at TSMC.
• Transistors and density: 76,300 million at 125.3M/mm² versus 15,300 million at 25.1M/mm²; die sizes are 609 mm² and 610 mm².
• Shading hardware: 18176 shading units, 568 TMUs, 192 ROPs, 142 RT cores, and 568 tensor cores on the L40S; 3584 shading units, 224 TMUs, 96 ROPs, and no RT or tensor cores on the P100.
• Throughput rates: 483.8 GPixel/s and 1,431.4 GTexel/s for the L40S against 127.6 GPixel/s and 297.7 GTexel/s for the P100.
• Compute: 91.61 TFLOPS FP32 and 91.61 TFLOPS FP16 (1:1) versus 9.526 TFLOPS FP32 and 19.05 TFLOPS FP16 (2:1).
• Memory: 48 GB GDDR6, 384-bit, 864.0 GB/s versus 12 GB HBM2, 3072-bit, 549.1 GB/s.
• Power: 300 W TDP, 1x 16-pin, 700 W suggested PSU versus 250 W TDP, 1x 8-pin, 600 W suggested PSU.
• Interface: PCIe 4.0 x16 versus PCIe 3.0 x16.
• Display: 1x HDMI 2.1 and 3x DisplayPort 1.4a versus no outputs at all.
• Physical: both dual-slot and 267 mm long; the L40S also lists a height of 111 mm.
• Release dates: October 2022 versus June 2016, with the L40S following Server Ampere and the P100 following Tesla Maxwell.

Head-to-Head Benchmarks

Only one direct head-to-head result exists in the database, and it is decisive. In geekbench_opencl, the L40S scored 330727 while the Tesla P100 PCIe 12 GB scored 79396, a 316.6 percent margin for the L40S. That is not a generational step; it is a multiplication of capability. To put it in context, the P100's own score is closely matched by cards like the Tesla P100 PCIe 16 GB (79605) and the GeForce RTX 5090 (79842) in its rival cluster, meaning the L40S outperforms the P100's entire neighborhood by a factor of roughly four.

The L40S supplements that OpenCL result with a Geekbench Vulkan score of 260799, a test the P100 does not have a recorded result for. Its average benchmark score of 295763 places it in the 99th percentile of all GPUs in the database. Within its rival cluster, the L40S leads the RTX 6000 Ada Generation by 3 percent and the L40 by 4.1 percent, sits 7 percent behind the AMD Instinct MI300X, and 11.7 percent behind the NVIDIA H200 NVL. That positioning matters: the L40S competes near the top of the professional segment, while the P100's 92nd-percentile average of 79396 reflects a card that was once flagship-class but now trades blows with a tightly bunched group of rivals spread within about 2 percent of each other.

The underlying hardware explains the benchmark gap. The L40S brings roughly five times the shading units, more than double the texture rate (1,431.4 versus 297.7 GTexel/s), nearly four times the pixel rate (483.8 versus 127.6 GPixel/s), and almost ten times the FP32 compute (91.61 versus 9.526 TFLOPS). The P100's HBM2 subsystem was its signature advantage at launch, but even there the L40S's 864.0 GB/s exceeds the P100's 549.1 GB/s by a wide margin while offering four times the memory capacity. The single P100 counterpoint is FP16 throughput, where its 2:1 ratio yields 19.05 TFLOPS, double its FP32 rate; the L40S nonetheless eclipses that figure at 91.61 TFLOPS FP16. With one win for the L40S and none for the P100 across recorded head-to-head tests, the data renders a clean verdict.