NVIDIA L20 vs NVIDIA Tesla V100 SXM2 16 GB Comparison

The Verdict

The data presents a clear generational divide. The NVIDIA L20 is the decisive winner in every recorded benchmark, holding a 119.5% higher average benchmark score (251,147 vs 114,395). The L20 sits in the 99th percentile of all GPUs, while the Tesla V100 SXM2 16 GB sits in the 95th. The L20’s nearest rivals include the NVIDIA L40 and RTX 6000 Ada Generation, which are 11.6% and 12.6% ahead, respectively. The V100’s nearest rivals are much closer: the RTX A5500 Mobile is only 0.4% ahead, and the RTX 4000 SFF Ada Generation trails by 2.3%.

The data suggests the L20 is the pick for anyone running modern compute workloads that can leverage its newer architecture and massive 48 GB memory pool. The V100, despite being an older design, still holds a place for legacy or specific HPC tasks where its HBM2 memory and high-bandwidth bus are sufficient. However, from a pure performance standpoint, the L20 is the stronger card. The V100 is an end-of-life product, while the L20 is active. For new deployments, the L20 is the data-backed choice. For maintaining existing infrastructure, the V100 remains a known quantity, but it is not competitive on raw score.

Architecture Differences

The two GPUs come from different architectural eras. The L20 uses the AD102 chip on the Ada Lovelace architecture, built on a 5 nm process at TSMC. The V100 uses the GV100 chip on the older Volta architecture, built on a 12 nm process. The L20 packs 76,300 million transistors into a 609 mm² die, yielding a density of 125.3 million transistors per square millimeter. The V100, in contrast, has 21,100 million transistors spread across a much larger 815 mm² die, resulting in a density of only 25.9 million per square millimeter. This is a massive density advantage for the L20, which highlights the manufacturing process leap.

The L20’s feature set is modern: it includes 92 RT cores and 368 tensor cores. The V100 has no RT cores at all, but it does have 640 tensor cores. The L20 supports DirectX 12 Ultimate (12_2), while the V100 is limited to DirectX 12 (12_1). Both support OpenGL 4.6 and Vulkan 1.4. The L20 has four DisplayPort 1.4a outputs, while the V100 has no display outputs. The V100 is an SXM module, whereas the L20 is a dual-slot card with a 16-pin power connector. The L20 is also a much higher-frequency part, with a base clock of 1440 MHz and a boost of 2520 MHz, compared to the V100’s 1312 MHz base and 1530 MHz boost.

Where Each One Wins

Based on the recorded data, the L20 wins every single benchmark category. In the Geekbench OpenCL test, the L20 scores 274,276, which is a staggering 213.6% higher than the V100’s 87,454. In the Geekbench Vulkan test, the L20 scores 228,018, which is 61.3% higher than the V100’s 141,336. The data shows no scenario where the V100 comes out ahead. However, the V100 does have a niche. Its memory system has a 4096-bit bus, which is significantly wider than the L20’s 384-bit bus, and it uses HBM2 memory. The L20 uses GDDR6. The V100’s memory bandwidth is 897.0 GB/s, which is slightly higher than the L20’s 864.0 GB/s. For workloads that are extremely bandwidth-sensitive and not compute-bound, the V100 could arguably hold its own, but the benchmark scores do not reflect a win.

The L20’s FP32 throughput is 59.35 TFLOPS, which is 3.79 times that of the V100’s 15.67 TFLOPS. In FP16, the L20 achieves 59.35 TFLOPS (at a 1:1 ratio), while the V100 achieves 31.33 TFLOPS (at a 2:1 ratio). The L20 also has a massive advantage in texture rate, 927.4 GTexel/s versus 489.6 GTexel/s, and pixel rate, 322.6 GPixel/s versus 195.8 GPixel/s. The L20 is also a much more modern part, released in November 2023, while the V100 was released in June 2017. The L20 is still in active production, while the V100 is end-of-life. If a workload absolutely needs the highest FP32 or FP16 throughput, or if it needs the 48 GB memory capacity, the L20 is the only option. If a workload is bound by a very specific legacy CUDA library that only supports Volta, the V100 would be the required choice.

FAQ

Q: Which GPU has a higher average benchmark score?

A: The NVIDIA L20 has a much higher average benchmark score of 251,147, compared to the NVIDIA Tesla V100 SXM2 16 GB’s 114,387. The L20’s score is 119.6% higher.

Q: How does the L20 perform in the Geekbench OpenCL test?

A: The L20 scores 274,276 in Geekbench OpenCL, which is 213.6% higher than the V100’s score of 87,454.

Q: How does the L20 compare to the V100 in the Vulkan benchmark?

A: The L20 scores 228,018 in Geekbench Vulkan, which is 61.3% higher than the V100’s score of 141,336. The L20 wins this test.

Q: What are the memory differences between the two cards?

A: The L20 has 48 GB of GDDR6 memory on a 384-bit bus, with a bandwidth of 864.0 GB/s. The V100 has 16 GB of HBM2 memory on a 4096-bit bus, with a bandwidth of 897.0 GB/s.

Q: Which card has more shading units?

A: The L20 has 11,776 shading units, while the V100 has 5,120. The L20 also has 368 TMUs and 128 ROPs, while the V100 has 320 TMUs and 128 ROPs.

Q: What is the difference in FP32 performance?

A: The L20 has an FP32 throughput of 59.35 TFLOPS, which is nearly four times the V100’s 15.67 TFLOPS.

Head-to-Head Benchmarks

The data shows a complete sweep for the L20. In the first head-to-head test, Geekbench OpenCL, the L20 scores 274,276. The V100 scores 87,454. This is a delta of 213.6%, meaning the L20 more than triples the performance of the V100. This is a massive gap, likely due to the architectural improvements. The L20’s transistor density is 125.3 million per square millimeter, compared to 25.9 million for the V100, which speaks to the efficiency of the newer process.

The second test, Geekbench Vulkan, shows a closer but still dominant margin. The L20 scores 228,018, while the V100 scores 141,336. The delta is 61.3%. While the L20 still wins, the Vulkan test is relatively more favorable to the V100. This could be due to the V100’s higher memory bandwidth (897.0 GB/s versus 864.0 GB/s) or its 4096-bit bus, which can be beneficial in certain memory access patterns. The V100 also has a 2:1 FP16 ratio, reaching 31.33 TFLOPS, which is a higher ratio than the L20’s 1:1. However, the L20 still has a raw FP16 throughput of 59.35 TFLOPS, which is higher than the V100’s 31.33 TFLOPS.

The L20’s score in OpenCL is 213.6% higher, while its score in Vulkan is 61.3% higher. The fact that the OpenCL gap is so much larger than the Vulkan gap suggests that the V100’s older architecture is more competitive in the Vulkan API, but it is still not enough to overcome the L20’s raw power. The L20 also has 368 tensor cores, which are dedicated to AI and ray tracing workloads. The V100 has 640 tensor cores, but it lacks RT cores entirely. The L20’s 92 RT cores give it a hardware-accelerated ray tracing ability that the V100 simply does not have. The data shows that the L20 is the undisputed winner in every recorded benchmark, with a 2-0 win record.

Specification Differences

| Specification | NVIDIA L20 | NVIDIA Tesla V100 SXM2 16 GB |

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

| Chip | AD102 | GV100 |

| 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 | 1440 MHz | 1312 MHz |

| Boost Clock | 2520 MHz | 1530 MHz |

| Memory Clock | 2250 MHz (18 Gbps effective) | 876 MHz (1752 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 | 897.0 GB/s |

| Shading Units | 11776 | 5120 |

| TMUs | 368 | 320 |

| ROPs | 128 | 128 |

| RT Cores | 92 | None (null) |

| Tensor Cores | 368 | 640 |

| Pixel Rate | 322.6 GPixel/s | 195.8 GPixel/s |

| Texture Rate | 927.4 GTexel/s | 489.6 GTexel/s |

| FP16 | 59.35 TFLOPS (1:1) | 31.33 TFLOPS (2:1) |

| FP32 | 59.35 TFLOPS | 15.67 TFLOPS |

| TDP | 275 W | 250 W |

| Slot Width | Dual-slot | SXM Module |

| Power Connectors | 1x 16-pin | None |

| 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) |

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

| Release Date | 2023-11-15 | 2017-06-20 |

| Predecessor | Server Ampere | Tesla Pascal |

| Successor | Server Hopper | Tesla Turing |

| Percentile | 99th | 95th |

| Average Benchmark Score | 251,147 | 114,395 |