NVIDIA CMP 40HX vs NVIDIA Tesla P100 PCIe 16 GB Comparison

The NVIDIA CMP 40HX and the NVIDIA Tesla P100 PCIe 16 GB represent two very different answers to the question of what a GPU should do. The CMP 40HX is a Turing-era product built for a single, specific task, while the Tesla P100 is a Pascal-era compute workhorse. The benchmark data shows a clear single win for the CMP 40HX, but the underlying specifications tell a story of divergent design goals that go far beyond a single score.

Where Each One Wins

The head-to-head data contains only one benchmark result, and it belongs to the NVIDIA CMP 40HX. In the Geekbench OpenCL test, the CMP 40HX scores 93,395 against the Tesla P100's 79,605, a 17.3% advantage. This is a significant margin, placing the CMP 40HX ahead of its rival in raw compute throughput as measured by this test. Its average benchmark score of 85,637 across all tests further reinforces this position, placing it in the 93rd percentile of all GPUs.

However, the Tesla P100's single OpenCL score of 79,605 is not a complete picture of its capabilities. It sits in the 92nd percentile of all GPUs, just one point behind the CMP 40HX. The data shows the P100 is not far behind in this specific test, and its nearest rivals include the NVIDIA GeForce RTX 5090, against which it is only 0.3% slower. This suggests that in the OpenCL workload, the P100 is competitive with much newer hardware, even if it loses to the CMP 40HX. The CMP 40HX's win is in raw compute speed, but the P100's architecture with 16 GB of HBM2 memory and a 4096-bit bus suggests its strengths may lie in memory-bound tasks where capacity and bandwidth are paramount, a dimension not captured by the single OpenCL score.

Architecture Differences

The two cards are built on fundamentally different architectures from different generations. The CMP 40HX uses the TU106 chip based on the Turing architecture, manufactured on a 12 nm process at TSMC. It packs 10,800 million transistors into a 445 mm² die. In contrast, the Tesla P100 uses the GP100 chip based on the older Pascal architecture, built on a 16 nm process, also at TSMC. The P100's die is significantly larger at 610 mm² and contains 15,300 million transistors. Interestingly, the transistor density is nearly identical: 24.3M per mm² for the CMP 40HX and 25.1M per mm² for the P100.

The feature sets diverge sharply. The CMP 40HX is a Turing part with 36 RT cores and 288 tensor cores, reflecting its newer architecture. The Tesla P100 has no RT cores and no tensor cores, as these features were introduced after its release. The CMP 40HX also supports DirectX 12 Ultimate (12_2) and Vulkan 1.4, while the P100 is limited to DirectX 12 (12_1) and Vulkan 1.3. Both support OpenGL 4.6. The CMP 40HX's memory is 8 GB of GDDR6 on a 256-bit bus, while the P100 uses 16 GB of HBM2 on a massive 4096-bit bus. This is the core architectural difference: a newer, feature-rich GPU with less memory versus an older, memory-focused GPU with double the capacity and a much wider interface.

FAQ

Q: Which GPU has more shading units?

A: The Tesla P100 has 3,584 shading units, while the CMP 40HX has 2,304. The P100 also has more texture mapping units (224 vs 144) and more raster operation units (96 vs 64).

Q: How do their memory bandwidths compare?

A: The Tesla P100 has significantly higher memory bandwidth at 732.2 GB/s, thanks to its 4096-bit HBM2 interface. The CMP 40HX offers 448.0 GB/s over a 256-bit GDDR6 bus.

Q: Does the CMP 40HX support hardware ray tracing?

A: Yes, the CMP 40HX includes 36 RT cores and 288 tensor cores as part of the Turing architecture. The Tesla P100 has no RT or tensor cores.

Q: Which card has a higher FP32 performance?

A: The Tesla P100 leads in FP32 with 9.526 TFLOPS, versus the CMP 40HX's 7.603 TFLOPS. The P100 also leads in FP16 with 19.05 TFLOPS compared to 15.21 TFLOPS.

Q: What is the power draw difference?

A: The CMP 40HX has a TDP of 185 W, while the Tesla P100 has a TDP of 250 W. The CMP 40HX also has a lower suggested PSU of 450 W, versus 600 W for the P100.

Q: Are these cards still in production?

A: No, both are end-of-life products. The CMP 40HX was released in February 2021, while the Tesla P100 was released in June 2016 and has a successor named Tesla Volta.

Specification Differences

The key specification differences are stark. The CMP 40HX is a 12 nm Turing chip with 10,800 million transistors on a 445 mm² die, while the Tesla P100 is a 16 nm Pascal chip with 15,300 million transistors on a 610 mm² die. Clock speeds differ, with the CMP 40HX running at a 1470 MHz base and 1650 MHz boost, compared to the P100's 1190 MHz base and 1329 MHz boost. Memory is a major split: 8 GB GDDR6 at 14 Gbps effective on a 256-bit bus for the CMP 40HX, versus 16 GB HBM2 at 1430 Mbps effective on a 4096-bit bus for the P100. The CMP 40HX has 36 RT cores and 288 tensor cores, while the P100 has none. The P100 has more shading units (3,584 vs 2,304), TMUs (224 vs 144), and ROPs (96 vs 64). The CMP 40HX has a dual-slot width, 1x 8-pin power connector, and uses a PCIe 1.0 x4 interface, while the P100 also has a dual-slot width and 1x 8-pin power connector but uses a PCIe 3.0 x16 interface. The CMP 40HX is 229 mm long, while the P100 is 267 mm long. Both have no display outputs. The CMP 40HX supports DirectX 12 Ultimate and Vulkan 1.4, while the P100 supports DirectX 12 (12_1) and Vulkan 1.3. The P100 has a predecessor (Tesla Maxwell) and successor (Tesla Volta), whereas the CMP 40HX does not list either.

Head-to-Head Benchmarks

The only direct benchmark comparison available is the Geekbench OpenCL test, where the NVIDIA CMP 40HX wins decisively. It scores 93,395 against the Tesla P100's 79,605, a delta of 17.3%. This is a substantial lead that cannot be dismissed. Looking at the competitive landscape, the CMP 40HX's average score of 85,637 puts it 4.4% ahead of the AMD Radeon PRO W6600 and 5.8% ahead of the AMD Radeon Pro Vega 64X, while being just 1.7% behind the AMD Radeon PRO W7600. The Tesla P100's average score of 79,605 places it 0.3% ahead of the Tesla P100 PCIe 12 GB and 1.7% behind the AMD Radeon Pro Vega 64X. The P100 is also remarkably close to the NVIDIA GeForce RTX 5090, which scores 79,842, a 0.3% difference. This suggests that while the CMP 40HX wins this particular test, the P100 is not obsolete; it remains within striking distance of modern flagships in this workload. The CMP 40HX's victory in OpenCL is likely tied to its higher boost clock and newer architecture, but the P100's 16 GB of HBM2 memory could be a deciding factor in memory-intensive applications where the CMP 40HX's 8 GB would be insufficient.

The Verdict

The data points to a clear choice based on workload. The NVIDIA CMP 40HX is the winner in the single benchmark test, delivering 17.3% higher OpenCL performance. It is a Turing-generation card with ray tracing cores, tensor cores, and modern API support, all within a lower 185 W TDP. For tasks that rely on raw compute throughput and benefit from newer features, the CMP 40HX is the better option. The Tesla P100, however, offers double the memory capacity (16 GB vs 8 GB), a much wider 4096-bit memory bus, and higher bandwidth at 732.2 GB/s. It also has more shading units, TMUs, and ROPs, which contribute to its higher FP32 and FP16 performance. The P100's age is a factor, but its memory configuration makes it a candidate for workloads where memory capacity and bandwidth are the limiting factors. The choice is binary: the CMP 40HX for faster compute in a single benchmark, or the Tesla P100 for larger memory capacity and higher theoretical throughput. The data does not support a universal winner, only a situational one.