NVIDIA L20 vs NVIDIA Tesla P100 PCIe 12 GB Comparison

Head-to-Head Benchmarks

The recorded data contains a single direct benchmark comparison between the NVIDIA L20 and the NVIDIA Tesla P100 PCIe 12 GB, and the result is decisive. In the Geekbench OpenCL test, the L20 scores 274,276 points, while the Tesla P100 manages 79,396 points. That represents a delta of 245.5% in favor of the L20, meaning the newer card delivers roughly three and a half times the compute throughput in this particular workload. This is not a marginal victory; it is a generational gap made visible in a single metric.

Looking at the broader database averages, the L20 carries an average benchmark score of 251,147 across all recorded tests, placing it in the 99th percentile of all GPUs. The Tesla P100 PCIe 12 GB, by contrast, has an average score of 79,396, which puts it in the 92nd percentile. The percentile gap, while still favoring the L20, is narrower than the raw score gap might suggest, because the database includes a wide range of consumer and professional parts. Still, the L20 sits at the very top of the distribution, while the P100, despite its age, remains a respectable performer even by modern standards.

When the L20 is compared to its nearest rivals in the database, the picture sharpens further. The NVIDIA PG506-232 scores 225,124, which is 11.6% below the L20. The AMD Radeon PRO W7900D scores 219,827, which is 14.2% below. On the other side, the NVIDIA L40 scores 284,111, putting it 11.6% above the L20, and the NVIDIA RTX 6000 Ada Generation scores 287,237, which is 12.6% above. So the L20 sits in a tight cluster of high-end professional cards, slightly behind the top Ada offerings but clearly ahead of the previous-generation and competing workstation parts.

For the Tesla P100 PCIe 12 GB, the nearest rivals tell a different story. The Tesla P100 PCIe 16 GB scores 79,605, which is a mere 0.3% higher, confirming that the 12 GB and 16 GB variants are nearly identical in compute performance. The NVIDIA GeForce RTX 5090 scores 79,842, which is 0.6% higher, and the AMD Radeon RX 6850M XT scores 78,940, which is 0.6% lower. The AMD Radeon Pro Vega 64X scores 80,959, which is 1.9% higher. In other words, the P100's performance level is closely matched by a handful of very different cards, ranging from a modern consumer flagship to a laptop GPU and an older workstation part. Its 92nd percentile placement reflects this clustering: it is not an outlier, but rather a card that sits in a crowded performance band.

The head-to-head data leaves no ambiguity about the winner. The L20 wins the only recorded benchmark, and it does so by a margin that dwarfs any other comparison in the database. Even the largest delta among the L20's nearest rivals, the 14.2% gap to the AMD Radeon PRO W7900D, is trivial next to the 245.5% gap between the L20 and the P100. This is not a close contest by any measure.

Architecture Differences

The two cards come from different eras of NVIDIA's design philosophy, and the architecture gap explains nearly all of the performance difference. The L20 is built on the Ada Lovelace architecture, using the AD102 chip, and is fabricated on a 5 nm process at TSMC. The Tesla P100 uses the Pascal architecture, built on the GP100 chip, and is fabricated on a 16 nm process, also at TSMC. The process node difference alone, from 16 nm to 5 nm, represents a massive jump in transistor density and efficiency.

The transistor counts reflect this shift. The L20 packs 76,300 million transistors onto a 609 mm² die, giving it a transistor density of 125.3 million per square millimeter. The P100 has 15,300 million transistors on a 610 mm² die, which works out to 25.1 million per square millimeter. The die sizes are nearly identical, at 609 mm² versus 610 mm², but the L20 crams five times as many transistors into that space. This is the clearest possible illustration of what five generations of process improvement can achieve.

The memory subsystems are also fundamentally different. The L20 uses 48 GB of GDDR6 memory on a 384-bit bus, delivering 864.0 GB/s of bandwidth. The P100 uses 12 GB of HBM2 memory on a much wider 3072-bit bus, but delivers only 549.1 GB/s due to its older memory technology and lower clock speeds. The L20's memory runs at 2250 MHz, or 18 Gbps effective, while the P100's memory runs at 715 MHz, or 1430 Mbps effective. The L20 has four times the capacity and roughly 57% more bandwidth, despite using a narrower bus.

The compute resources are equally lopsided. The L20 has 11,776 shading units, 368 texture mapping units, and 128 render output units. The P100 has 3,584 shading units, 224 TMUs, and 96 ROPs. The L20 also brings dedicated hardware that the P100 lacks entirely: 92 ray tracing cores and 368 tensor cores. The P100 has no RT cores and no tensor cores, reflecting its pre-ray-tracing, pre-AI-accelerator design. In terms of raw throughput, the L20 delivers 59.35 TFLOPS of FP32 performance and the same 59.35 TFLOPS of FP16 performance, at a 1:1 ratio. The P100 delivers 9.526 TFLOPS of FP32 and 19.05 TFLOPS of FP16, at a 2:1 ratio. Even in FP16, where the P100 is comparatively stronger, the L20 more than triples its output.

The clock speeds also favor the L20. The L20 has a base clock of 1440 MHz and a boost clock of 2520 MHz. The P100 has a base clock of 1190 MHz and a boost clock of 1329 MHz. The L20's boost clock is nearly double the P100's, which, combined with the massive core count difference, explains the overwhelming benchmark margin.

Power consumption is similar on paper, with the L20 rated at 275 W and the P100 at 250 W, but the efficiency story is entirely different. The L20 delivers 59.35 TFLOPS of FP32 performance within its 275 W envelope, while the P100 delivers 9.526 TFLOPS within 250 W. That is a roughly sixfold improvement in performance per watt, a direct consequence of the process node and architecture advancements. Both cards are dual-slot designs and both recommend a 600 W power supply, but the L20 uses a single 16-pin connector while the P100 uses a single 8-pin connector.

The interface and output capabilities also differ. The L20 uses PCIe 4.0 x16 and has four DisplayPort 1.4a outputs. The P100 uses PCIe 3.0 x16 and has no display outputs at all, making it a pure compute accelerator. The L20 supports DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4. The P100 supports DirectX 12 (12_1), OpenGL 4.6, and Vulkan 1.3. The L20 is listed as an active production part, released in November 2023, while the P100 is end-of-life, released in June 2016.

The Verdict

The data points to a clear conclusion: the NVIDIA L20 is the superior card in every measured category. It wins the only head-to-head benchmark by 245.5%, and its architecture, memory capacity, compute resources, and feature set are all generations ahead of the Tesla P100 PCIe 12 GB. The L20's 99th percentile ranking versus the P100's 92nd percentile underscores the gap, even accounting for the fact that the P100 still outperforms a majority of GPUs in the database.

For workloads that rely on FP32 compute, ray tracing, tensor operations, or large memory footprints, the L20 is the obvious choice. Its 48 GB of GDDR6 memory is four times the P100's 12 GB of HBM2, and its 864.0 GB/s bandwidth is substantially higher. The presence of 92 RT cores and 368 tensor cores opens up capabilities that the P100 cannot offer at all. The L20 is also a more modern platform, with PCIe 4.0 support, display outputs, and a current production status.

The Tesla P100 PCIe 12 GB, however, is not without merit in the database context. Its 92nd percentile placement shows that it remains a capable compute card, particularly for FP64 or mixed-precision workloads where its Pascal architecture's 2:1 FP16 ratio might be leveraged. It is also a lower-power card at 250 W, and its HBM2 memory, while smaller and slower, still provides respectable bandwidth. For legacy applications or environments where the P100 is already deployed, the data suggests it can still hold its own against a range of modern parts, as evidenced by its tight clustering with the RTX 5090 and the RX 6850M XT.

That said, for any new deployment or any workload where performance is the primary concern, the L20 is the only rational pick. The P100's era has passed, and the benchmark data reflects a generational chasm that no amount of software optimization can bridge. The L20 is not just faster; it is faster by a factor that makes the comparison almost academic.

FAQ

Q: How much faster is the NVIDIA L20 than the Tesla P100 PCIe 12 GB in Geekbench OpenCL?

A: The L20 scores 274,276 points compared to the P100's 79,396 points, a delta of 245.5% in favor of the L20.

Q: What is the memory capacity difference between the two cards?

A: The L20 has 48 GB of GDDR6 memory, while the Tesla P100 PCIe 12 GB has 12 GB of HBM2 memory. The L20 also has higher bandwidth at 864.0 GB/s versus 549.1 GB/s.

Q: Does the Tesla P100 support ray tracing or tensor cores?

A: No. The P100 has no RT cores and no tensor cores. The L20 has 92 ray tracing cores and 368 tensor cores.

Q: What are the production statuses of the two cards?

A: The NVIDIA L20 is listed as active, with a release date of November 2023. The Tesla P100 PCIe 12 GB is end-of-life, with a release date of June 2016.

Q: How do the cards compare in terms of transistor count and process node?

A: The L20 uses a 5 nm process and packs 76,300 million transistors on a 609 mm² die. The P100 uses a 16 nm process and has 15,300 million transistors on a 610 mm² die.

Q: What is the average benchmark score for each card?

A: The L20 has an average benchmark score of 251,147, placing it in the 99th percentile of all GPUs. The P100 has an average score of 79,396, placing it in the 92nd percentile.

Where Each One Wins

The NVIDIA L20 wins in every directly measured benchmark and in nearly every architectural category. Its advantages are most pronounced in raw compute throughput, where its 59.35 TFLOPS of FP32 performance dwarfs the P100's 9.526 TFLOPS. It also wins decisively in memory capacity, bandwidth, and feature support. The 48 GB memory pool is suited for large models or datasets that would not fit in the P100's 12 GB. The tensor cores enable AI and deep learning workloads that the P100 cannot accelerate. The RT cores enable ray-traced rendering, which is entirely absent on the P100. The L20's support for DirectX 12 Ultimate and Vulkan 1.4 also makes it more versatile for modern graphics APIs, and its display outputs allow for direct visualization, something the P100 cannot do.

The Tesla P100 PCIe 12 GB retains a narrow set of advantages, though the data does not show any benchmark where it beats the L20. Its 250 W power draw is slightly lower than the L20's 275 W, which could matter in power-constrained environments. Its 3072-bit memory bus is wider than the L20's 384-bit bus, though the L20's faster memory clocks more than compensate. The P100's FP16 performance of 19.05 TFLOPS, while lower than the L20's 59.35 TFLOPS, is achieved at a 2:1 ratio, which may be relevant for specific mixed-precision workflows. The P100 also uses a standard 8-pin power connector, which is more universally compatible than the L20's 16-pin connector. Finally, the P100's end-of-life status and older architecture mean it may be available at lower cost in the secondary market, though the database does not track pricing for the L20.

In practical terms, the L20 is the card for anyone building a new system or upgrading for modern workloads. The P100 is a legacy part that remains surprisingly competitive in its percentile ranking, but it cannot match the L20's performance, features, or efficiency. The data is unambiguous: this is not a contest between equals, but a demonstration of how far GPU architecture has advanced in seven years.