NVIDIA Tesla P100 PCIe 12 GB vs NVIDIA Tesla P40 Comparison

# NVIDIA Tesla P100 PCIe 12 GB vs NVIDIA Tesla P40

The Verdict

The NVIDIA Tesla P100 PCIe 12 GB is the clear benchmark winner, taking the only head-to-head test by a substantial 28% margin. In Geekbench OpenCL, the P100 scores 79396 against the P40's 62017, a decisive gap that reflects the P100's HBM2 memory advantage and higher transistor count. The P100 also sits at the 92nd percentile of all GPUs, while the P40 ranks at the 89th percentile.

For compute-heavy workloads where OpenCL performance matters most, the P100 is the superior choice. It delivers 9.526 TFLOPS of FP32 performance and 19.05 TFLOPS of FP16 performance (2:1 ratio), making it dramatically stronger for mixed-precision work. The P40's FP16 output of 183.7 GFLOPS (1:64 ratio) is a non-starter for anyone needing half-precision throughput.

However, the P40 is not without merit. It offers 24 GB of GDDR5 memory versus the P100's 12 GB of HBM2, which could matter for models or datasets that require more capacity than bandwidth. The P40 also has a higher boost clock (1531 MHz vs 1329 MHz), more shading units (3840 vs 3584), and higher pixel and texture rates. For raw rasterization throughput, the P40 leads — but these are compute-focused Tesla cards with no display outputs.

The verdict is straightforward: choose the P100 for compute performance and memory bandwidth, choose the P40 only if 24 GB capacity and higher FP32 peak throughput (11.76 TFLOPS) outweigh the P100's 28% OpenCL advantage. The P100's launch MSRP was 4,599 USD; the P40's was 5,699 USD.

FAQ

Q: Which card is faster in OpenCL?

A: The P100 wins decisively, scoring 79396 versus the P40's 62017 — a 28% advantage.

Q: Does the P40 have more memory?

A: Yes, the P40 has 24 GB of GDDR5 memory, double the P100's 12 GB of HBM2.

Q: Which card has higher FP32 throughput?

A: The P40 leads with 11.76 TFLOPS, compared to the P100's 9.526 TFLOPS.

Q: How do they compare in memory bandwidth?

A: The P100 is far ahead at 549.1 GB/s, versus the P40's 347.1 GB/s.

Q: Are both cards the same size?

A: Both are dual-slot, 267 mm (10.5 inches) long, and require a 600 W power supply.

Q: Which card supports Vulkan 1.4?

A: Only the P40 supports Vulkan 1.4; the P100 is capped at Vulkan 1.3.

Architecture Differences

Both cards are built on NVIDIA's Pascal architecture at TSMC's 16 nm process node, and both belong to the Tesla Pascal (Pxx) generation. However, they use different chips with distinct design priorities.

The P100 uses the GP100 chip, which is a massive 610 mm² die containing 15,300 million transistors. This translates to a transistor density of 25.1M per mm². The GP100 is designed around HBM2 memory, which explains its 3072-bit memory bus and 549.1 GB/s bandwidth. The P100's FP16 capability is 19.05 TFLOPS at a 2:1 ratio, indicating a hardware path for half-precision compute — a key feature for deep learning workloads.

The P40 uses the GP102 chip, a smaller 471 mm² die with 11,800 million transistors, also at 25.1M per mm² density. The GP102 is a more conventional graphics-oriented chip, paired with GDDR5 memory on a 384-bit bus. Its FP16 performance is severely limited at 183.7 GFLOPS with a 1:64 ratio, meaning FP16 is effectively emulated rather than hardware-accelerated. This is the single biggest architectural differentiator: the P100 was built for compute-heavy, mixed-precision workloads, while the P40 was adapted from a gaming-class GPU for server use.

Both cards have no ray tracing cores and no tensor cores — these are pre-Volta products. They share the same PCIe 3.0 x16 bus interface and have no display outputs. Both are end-of-life products, with the P100 released in June 2016 and the P40 in September 2016.

Specification Differences

The two cards differ in nearly every key specification except TDP (both 250 W), slot width (dual-slot), bus interface (PCIe 3.0 x16), and length (267 mm / 10.5 inches).

| Specification | P100 PCIe 12 GB | P40 |

|---|---|---|

| Chip | GP100 | GP102 |

| Process Node | 16 nm | 16 nm |

| Transistors | 15,300 million | 11,800 million |

| Die Size | 610 mm² | 471 mm² |

| Base Clock | 1190 MHz | 1303 MHz |

| Boost Clock | 1329 MHz | 1531 MHz |

| Memory Size | 12 GB | 24 GB |

| Memory Type | HBM2 | GDDR5 |

| Memory Bus | 3072 bit | 384 bit |

| Memory Bandwidth | 549.1 GB/s | 347.1 GB/s |

| Shading Units | 3584 | 3840 |

| TMUs | 224 | 240 |

| ROPs | 96 | 96 |

| Pixel Rate | 127.6 GPixel/s | 147.0 GPixel/s |

| Texture Rate | 297.7 GTexel/s | 367.4 GTexel/s |

| FP32 | 9.526 TFLOPS | 11.76 TFLOPS |

| FP16 | 19.05 TFLOPS (2:1) | 183.7 GFLOPS (1:64) |

| Memory Clock | 715 MHz (1430 Mbps effective) | 1808 MHz (7.2 Gbps effective) |

| Vulkan Support | 1.3 | 1.4 |

| Power Connector | 1x 8-pin | 8-pin EPS |

| Height | Not specified | 111 mm (4.4 inches) |

| Launch MSRP | 4,599 USD | 5,699 USD |

The P40 has higher clocks across the board, more shading units and TMUs, and double the memory capacity, but the P100 counters with a wider memory bus, higher bandwidth, and vastly superior FP16 throughput.

Head-to-Head Benchmarks

The only direct comparison available is the Geekbench OpenCL test, and it is a landslide. The P100 scores 79396, while the P40 manages 62017. That is a 28% delta in favor of the P100, a margin large enough to be considered a generational gap rather than a minor variance.

Context from the nearest rivals reinforces the P100's position. The P100 sits within 0.3% of the Tesla P100 PCIe 16 GB (score 79605) and within 0.6% of the GeForce RTX 5090 (score 79842). It outperforms the AMD Radeon RX 6850M XT by 0.6% and only trails the Radeon Pro Vega 64X by 1.9%. In short, the P100's OpenCL score places it among top-tier compute hardware.

The P40, by contrast, is clustered with mid-range and workstation cards. It leads the AMD Radeon Pro WX 9100 by 1.4%, trails the AMD Radeon VII by 1.4%, and ties closely with the NVIDIA CMP 30HX and AMD Radeon RX 9060 XT LP, both of which it beats by 2%. The P40 also has a Vulkan score of 68172, which is higher than its OpenCL score — suggesting it handles graphics-style compute better than raw compute workloads.

The data shows a clear hierarchy: the P100 is approximately 22% faster than the P40 in average benchmark score (79396 vs 65095), and 28% faster in the direct head-to-head OpenCL test.

Where Each One Wins

NVIDIA Tesla P100 PCIe 12 GB wins in:

• OpenCL compute performance: 79396 vs 62017, a 28% lead.
• Memory bandwidth: 549.1 GB/s vs 347.1 GB/s — a 58% advantage that matters for memory-bound kernels.
• FP16 compute: 19.05 TFLOPS vs 183.7 GFLOPS — ideal for deep learning inference and training where half-precision is used.
• Transistor count: 15,300 million vs 11,800 million, reflecting a more complex compute-oriented design.
• Overall benchmark percentile: 92nd vs 89th among all GPUs.

NVIDIA Tesla P40 wins in:

• Memory capacity: 24 GB vs 12 GB — double the VRAM for large models or datasets that fit within GDDR5's bandwidth constraints.
• FP32 compute: 11.76 TFLOPS vs 9.526 TFLOPS — a 23% raw single-precision throughput advantage.
• Pixel rate: 147.0 GPixel/s vs 127.6 GPixel/s.
• Texture rate: 367.4 GTexel/s vs 297.7 GTexel/s.
• Boost clock: 1531 MHz vs 1329 MHz, and base clock of 1303 MHz vs 1190 MHz.
• Vulkan support: 1.4 vs 1.3 on the P100.

In practice, the P100 is the better accelerator for heterogeneous compute, especially tasks that leverage FP16 or require high memory bandwidth. The P40's strengths are in capacity-bound scenarios and workloads that favor higher FP32 throughput, such as certain rendering or scientific simulations that do not use half-precision. Given that both cards are end-of-life and carry no display outputs, the choice should be driven by the specific compute profile: bandwidth and FP16 point to the P100; capacity and FP32 peak point to the P40. For most general compute tasks, the P100's 28% OpenCL lead makes it the default pick.