NVIDIA L20 vs NVIDIA Tesla P100 PCIe 16 GB Comparison

Head-to-Head Benchmarks

The benchmark data presents a decisive picture in favor of the NVIDIA L20. The only directly comparable measurement, Geekbench OpenCL, shows the L20 scoring 274,276 points against the Tesla P100 PCIe 16 GB’s 79,605 points. That is a 244.5% advantage, meaning the L20 delivers more than triple the compute throughput in this workload. This is not a marginal generational step; it is a fundamental shift in performance class.

Looking at the broader database context, the L20’s average benchmark score of 251,147 places it in the 99th percentile of all GPUs. Its nearest rivals include the NVIDIA L40 at 284,111 (11.6% faster) and the NVIDIA RTX 6000 Ada Generation at 287,237 (12.6% faster), while it sits comfortably ahead of the NVIDIA PG506-232 by 11.6% and the AMD Radeon PRO W7900D by 14.2%. The Tesla P100, by contrast, records an average score of 79,605, which places it in the 92nd percentile. Its closest competitors are the NVIDIA Tesla P100 PCIe 12 GB at 79,396 (0.3% behind), the NVIDIA GeForce RTX 5090 at 79,842 (0.3% ahead), the AMD Radeon RX 6850M XT at 78,940 (0.8% behind), and the AMD Radeon Pro Vega 64X at 80,959 (1.7% ahead). The P100’s competitive set is entirely different, clustered in a much lower performance band.

In short, the head-to-head data is one-sided. The L20 wins the only available confrontation, and the margin is enormous. The P100 has no countervailing benchmark victory in the recorded data.

Architecture Differences

The two cards are separated by multiple architectural generations. The L20 is built on the Ada Lovelace architecture (chip AD102), while the Tesla P100 is based on the older Pascal architecture (chip GP100). This is not a minor revision; the design philosophies and capabilities are drastically different. The L20 is fabricated on a 5 nm process at TSMC, whereas the P100 uses a 16 nm process, also at TSMC. This process shrink allows the L20 to pack 76,300 million transistors onto a 609 mm² die, achieving a transistor density of 125.3 million per square millimeter. The P100, by contrast, contains only 15,300 million transistors on a slightly larger 610 mm² die, yielding a density of 25.1 million per square millimeter. The L20 has nearly five times the transistor density.

The compute configuration is equally divergent. The L20 houses 11,776 shading units, 368 TMUs, and 128 ROPs. The P100 has 3,584 shading units, 224 TMUs, and 96 ROPs. The L20 also brings dedicated hardware that the P100 lacks entirely: 92 RT cores for ray tracing and 368 tensor cores for AI acceleration. The P100 has no RT cores and no tensor cores in the recorded specifications. The L20’s raw compute figures reflect this: it delivers 59.35 TFLOPS for FP32 and the same 59.35 TFLOPS for FP16 (at a 1:1 ratio). The P100 manages 9.526 TFLOPS for FP32 and 19.05 TFLOPS for FP16 (at a 2:1 ratio). The L20 is roughly 6.2 times faster in FP32 and over 3 times faster in FP16.

Memory subsystem design also differs. The L20 uses 48 GB of GDDR6 memory on a 384 bit bus, achieving a bandwidth of 864.0 GB/s. The P100 uses only 16 GB of HBM2 memory, but on a much wider 4096 bit bus, delivering 732.2 GB/s. While the P100’s HBM2 offers a wide bus, the L20’s GDDR6 still provides higher total bandwidth. The memory clock speeds are not directly comparable due to different memory types, but the effective data rates are recorded as 18 Gbps for the L20 and 1430 Mbps for the P100. The interface standard also changed: the L20 uses PCIe 4.0 x16, while the P100 uses PCIe 3.0 x16. Display outputs are another major difference: the L20 has 4x DisplayPort 1.4a, while the P100 has no display outputs at all, indicating a pure compute accelerator role for the older card.

FAQ

Q: Which GPU has the higher average benchmark score, and by how much?

A: The NVIDIA L20 has an average benchmark score of 251,147, which is significantly higher than the Tesla P100’s 79,605. The L20 also sits in the 99th percentile of all GPUs, while the P100 is in the 92nd percentile.

Q: Does the Tesla P100 have any ray tracing or tensor core hardware?

A: No. The recorded specifications for the Tesla P100 list no RT cores and no tensor cores. The L20, in contrast, includes 92 RT cores and 368 tensor cores, making it much more capable for modern workloads that rely on these features.

Q: What is the memory capacity and bandwidth difference between the two?

A: The L20 has 48 GB of GDDR6 memory with a bandwidth of 864.0 GB/s. The P100 has 16 GB of HBM2 memory with a bandwidth of 732.2 GB/s. The L20 has both more total capacity and higher bandwidth.

Q: Which card has a higher FP16 throughput?

A: The L20 achieves 59.35 TFLOPS for FP16 (at a 1:1 ratio with FP32). The P100 achieves 19.05 TFLOPS for FP16 (at a 2:1 ratio). The L20 is more than 3 times faster in this metric.

Q: How do the power connectors differ?

A: The L20 uses a single 16-pin power connector, while the P100 uses a single 8-pin connector. Both have a suggested power supply of 600 W, and both are dual-slot cards.

Q: Which card is still in production?

A: The L20 is listed as “Active” in production status, while the Tesla P100 is listed as “End-of-life.”

The Verdict

The data points to a lopsided comparison. For any workload that benefits from raw compute throughput, modern feature support, or large memory capacity, the NVIDIA L20 is the clear choice. Its benchmark score is 244.5% higher than the P100’s in OpenCL, and it holds a 99th percentile ranking versus the P100’s 92nd. The L20 offers over 6 times the FP32 throughput, over 3 times the FP16 throughput, and 3 times the memory capacity. It also brings RT cores and tensor cores to the table, features that the P100 simply cannot provide.

The Tesla P100 has no recorded wins in this comparison. Its strengths are relative to other older cards, not against the L20. For example, it is only 0.3% slower than the GeForce RTX 5090 in the database, which shows it holds up well among its own generation’s peers. But when placed against the L20, every measurable metric favors the newer card. The only advantage the P100 might have is its PCIe 3.0 interface, which is more compatible with older systems, and its lower power draw of 250 W versus the L20’s 275 W. But these are minor points against a 244.5% performance deficit.

For users considering a new purchase, the L20 is the only rational choice from a performance standpoint. For those needing to maintain legacy systems, the P100 remains a functional compute accelerator, but it is strictly outclassed in every benchmark and specification comparison available in the database.

Specification Differences

| Specification | NVIDIA L20 | NVIDIA Tesla P100 PCIe 16 GB |

| :--- | :--- | :--- |

| Architecture | Ada Lovelace | Pascal |

| Generation | Server Ada (Lxx) | Tesla Pascal (Pxx) |

| Process Node | 5 nm | 16 nm |

| Transistors | 76,300 million | 15,300 million |

| Die Size | 609 mm² | 610 mm² |

| Transistor Density | 125.3M / mm² | 25.1M / mm² |

| Base Clock | 1440 MHz | 1190 MHz |

| Boost Clock | 2520 MHz | 1329 MHz |

| Memory Clock | 2250 MHz (18 Gbps effective) | 715 MHz (1430 Mbps effective) |

| Memory Size | 48 GB | 16 GB |

| Memory Type | GDDR6 | HBM2 |

| Memory Bus Width | 384 bit | 4096 bit |

| Memory Bandwidth | 864.0 GB/s | 732.2 GB/s |

| Shading Units | 11776 | 3584 |

| TMUs | 368 | 224 |

| ROPs | 128 | 96 |

| RT Cores | 92 | N/A |

| Tensor Cores | 368 | N/A |

| Pixel Rate | 322.6 GPixel/s | 127.6 GPixel/s |

| Texture Rate | 927.4 GTexel/s | 297.7 GTexel/s |

| FP32 Performance | 59.35 TFLOPS | 9.526 TFLOPS |

| FP16 Performance | 59.35 TFLOPS (1:1) | 19.05 TFLOPS (2:1) |

| TDP | 275 W | 250 W |

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

| Bus Interface | PCIe 4.0 x16 | PCIe 3.0 x16 |

| Display Outputs | 4x DisplayPort 1.4a | No outputs |

| DirectX Support | 12 Ultimate (12_2) | 12 (12_1) |

| Vulkan Support | 1.4 | 1.3 |

| Dimensions (Length) | 267 mm (10.5 inches) | 267 mm (10.5 inches) |

| Production Status | Active | End-of-life |

| Release Date | 2023-11-15 | 2016-06-19 |

| Launch MSRP | Not available | 5,699 USD |

| Predecessor | Server Ampere | Tesla Maxwell |

| Successor | Server Hopper | Tesla Volta |

| Average Benchmark Score | 251,147 | 79,605 |

| Percentile vs All GPUs | 99 | 92 |