NVIDIA CMP 40HX vs NVIDIA Tesla P40 Comparison

The NVIDIA CMP 40HX and NVIDIA Tesla P40 are both end-of-life, dual-slot, no-display-output accelerators, but they target entirely different workloads. The data shows a clear generational split: the CMP 40HX, built on Turing, wins both recorded head-to-head benchmarks, while the Tesla P40, built on Pascal, offers far more memory capacity. The CMP 40HX scores 93,395 in Geekbench OpenCL and 77,879 in Vulkan, placing it in the 93rd percentile of all GPUs, with an average score of 85,637. The Tesla P40 scores 62,017 and 68,172 respectively, sitting in the 89th percentile with an average of 65,095. The CMP 40HX leads by 50.6% in OpenCL and 14.2% in Vulkan, yet the P40’s 24 GB memory dwarfs the CMP 40HX’s 8 GB. This is a tale of compute efficiency versus memory capacity.

Where Each One Wins

The CMP 40HX wins decisively on raw compute benchmarks. In Geekbench OpenCL, it scores 93,395 versus the Tesla P40’s 62,017, a 50.6% advantage. This is not a marginal lead; it is a massive gap that suggests the Turing architecture’s feature set and higher clock speeds translate directly into compute throughput. The CMP 40HX also wins in Vulkan, scoring 77,879 against the P40’s 68,172, a 14.2% lead. If your workload is dominated by OpenCL or Vulkan compute, the CMP 40HX is the clear choice. It also has a higher boost clock at 1,650 MHz versus 1,531 MHz, and its GDDR6 memory runs at 14 Gbps effective, which helps memory-bound tasks.

The Tesla P40 wins on memory capacity and bandwidth efficiency per watt in a different sense. It offers 24 GB of GDDR5 memory, triple the CMP 40HX’s 8 GB. While the P40’s raw bandwidth of 347.1 GB/s is lower than the CMP 40HX’s 448.0 GB/s, the sheer capacity allows it to hold larger datasets, models, or framebuffers without spilling to system memory. For workloads that need to fit large working sets on the card, the P40 is the only option here. Additionally, the P40 has a wider 384-bit memory bus versus 256-bit, though the CMP 40HX’s faster memory clock compensates. The P40 also has more shading units (3,840 vs 2,304), more TMUs (240 vs 144), and more ROPs (96 vs 64), but its older Pascal architecture and lower clocks prevent it from converting that hardware into benchmark wins.

Architecture Differences

The CMP 40HX is built on the TU106 chip using Turing architecture, fabricated on TSMC’s 12 nm process. It packs 10,800 million transistors on a 445 mm² die, with a transistor density of 24.3M per mm². Turing brings modern features: 36 RT cores and 288 tensor cores, support for DirectX 12 Ultimate (12_2), and FP16 performance of 15.21 TFLOPS at a 2:1 ratio. This makes it a compute beast for workloads that leverage ray tracing or tensor operations, even if those are not typical for a mining card. Its base clock is 1,470 MHz, boosting to 1,650 MHz, and it uses 8 GB of GDDR6 on a 256-bit bus.

The Tesla P40 uses the GP102 chip with Pascal architecture, on TSMC’s 16 nm process. It has 11,800 million transistors on a 471 mm² die, with a slightly higher density of 25.1M per mm². Pascal is older, with no RT or tensor cores, and its FP16 performance is a mere 183.7 GFLOPS at a 1:64 ratio, meaning FP16 is effectively crippled for compute. The P40’s base clock is 1,303 MHz, boosting to 1,531 MHz, and it uses 24 GB of GDDR5 on a 384-bit bus. DirectX support is limited to 12 (12_1), not Ultimate. The P40 also uses an 8-pin EPS power connector versus the CMP 40HX’s standard 8-pin, and it requires a 600 W PSU versus 450 W.

The process node difference is notable: 12 nm versus 16 nm. Despite having fewer transistors, the CMP 40HX achieves higher clocks and better power efficiency, with a TDP of 185 W versus the P40’s 250 W. The CMP 40HX also supports PCIe 1.0 x4, which is odd and severely limits host transfer speeds, while the P40 uses PCIe 3.0 x16. For compute tasks that require frequent host-device transfers, the P40’s bus is vastly superior, though the CMP 40HX’s benchmark scores suggest it does not need that bandwidth to win.

The Verdict

Who should pick the NVIDIA CMP 40HX? Anyone running OpenCL or Vulkan compute workloads where raw score matters. The data is unambiguous: a 50.6% lead in OpenCL and a 14.2% lead in Vulkan. The CMP 40HX also offers RT and tensor cores, which are absent on the P40, making it future-proof for any application that starts using those features. Its higher boost clock, faster memory, and lower TDP (185 W vs 250 W) mean you get more performance per watt and per dollar of power supply. If your work fits within 8 GB of memory, the CMP 40HX is the superior accelerator.

Who should pick the NVIDIA Tesla P40? Anyone whose dataset exceeds 8 GB. The P40’s 24 GB memory is the defining feature. While it loses both benchmarks, it loses by a smaller margin in Vulkan (14.2%) than in OpenCL (50.6%), suggesting that memory capacity can partially compensate in some APIs. The P40 also has more raw compute hardware—3,840 shading units versus 2,304—and a wider 384-bit bus. For inference workloads that load large models, or rendering tasks that need big framebuffers, the P40’s capacity wins. It is also the only one with a solid PCIe 3.0 x16 interface, which matters for data transfer. The trade-off is clear: accept slower compute but gain triple the memory.

FAQ

Q: Which card has the higher average benchmark score?

A: The NVIDIA CMP 40HX has an average benchmark score of 85,637, which is 31.5% higher than the Tesla P40’s 65,095.

Q: How much faster is the CMP 40HX in Geekbench OpenCL?

A: The CMP 40HX scores 93,395 versus the P40’s 62,017, a 50.6% advantage.

Q: Does the Tesla P40 have any architectural advantage?

A: Yes. The P40 has 3,840 shading units, 240 TMUs, and 96 ROPs, all higher than the CMP 40HX’s 2,304, 144, and 64 respectively. However, the P40 lacks RT cores and tensor cores entirely.

Q: What is the memory capacity difference?

A: The Tesla P40 has 24 GB of GDDR5, while the CMP 40HX has 8 GB of GDDR6. The P40’s memory bus is 384-bit versus 256-bit, but the CMP 40HX has higher bandwidth at 448.0 GB/s versus 347.1 GB/s.

Q: Which card is more power-efficient per the data?

A: The CMP 40HX has a TDP of 185 W, while the Tesla P40 is rated at 250 W. The CMP 40HX also requires a 450 W PSU versus the P40’s 600 W suggestion.

Q: Are both cards still in production?

A: No. Both are marked as end-of-life. The CMP 40HX was released on 2021-02-24, while the Tesla P40 was released earlier on 2016-09-12.

Head-to-Head Benchmarks

The biggest win for the CMP 40HX is in Geekbench OpenCL. The CMP 40HX scores 93,395, while the Tesla P40 scores 62,017, yielding a delta of 50.6%. This is a dominant result. To put it in context, the CMP 40HX’s nearest rival, the AMD Radeon PRO W7600, scores 87,108, which is only 1.7% lower. The P40, meanwhile, sits near the AMD Radeon Pro WX 9100, which scores 64,212, a 1.4% difference. The gap between the CMP 40HX and P40 is larger than the gap between the CMP 40HX and its closest rival, indicating a generation leap.

The Vulkan benchmark is closer but still favors the CMP 40HX. The CMP 40HX scores 77,879, versus the P40’s 68,172, a 14.2% lead. This suggests that Vulkan is less sensitive to the architectural differences, or that the P40’s larger memory and wider bus help it close the gap. Still, the CMP 40HX wins. The CMP 40HX’s Vulkan score puts it in the 93rd percentile overall, while the P40’s score places it in the 89th. The delta in percentile is small, but the raw score difference is meaningful.

The CMP 40HX wins 2 out of 2 benchmarks. There is no test in the data where the Tesla P40 comes out ahead. However, the P40’s 24 GB memory is not a benchmark score; it is a capacity metric. In workloads that are memory-bound rather than compute-bound, the P40’s capacity could flip the practical outcome. The data does not include such a test, so the CMP 40HX remains the benchmark champion, but the P40 remains the capacity champion.

Specification Differences

The two cards differ on nearly every hardware specification. The CMP 40HX uses a TU106 chip with Turing architecture, while the P40 uses a GP102 chip with Pascal. Process nodes differ: 12 nm for the CMP 40HX, 16 nm for the P40. Transistor counts are close—10,800 million versus 11,800 million—but die sizes differ slightly at 445 mm² and 471 mm². Clocks favor the CMP 40HX: base 1,470 MHz versus 1,303 MHz, boost 1,650 MHz versus 1,531 MHz. Memory is a major split: 8 GB GDDR6 at 14 Gbps effective on a 256-bit bus, versus 24 GB GDDR5 at 7.2 Gbps on a 384-bit bus. Bandwidth is 448.0 GB/s versus 347.1 GB/s.

Compute units favor the P40 in raw count but not in efficiency. The P40 has 3,840 shading units, 240 TMUs, and 96 ROPs, versus the CMP 40HX’s 2,304, 144, and 64. But the P40 has no RT cores or tensor cores, while the CMP 40HX has 36 RT cores and 288 tensor cores. Pixel rates are 147.0 GPixel/s for the P40 versus 105.6 GPixel/s for the CMP 40HX; texture rates are 367.4 GTexel/s versus 237.6 GTexel/s. FP32 is 11.76 TFLOPS for the P40 versus 7.603 TFLOPS for the CMP 40HX, yet the CMP 40HX wins benchmarks—proof of architectural efficiency.

Power and interface differ sharply. The CMP 40HX has a TDP of 185 W, uses a 1x 8-pin connector, and suggests a 450 W PSU. The P40 has a TDP of 250 W, uses an 8-pin EPS connector, and suggests a 600 W PSU. The CMP 40HX’s bus is PCIe 1.0 x4, while the P40 uses PCIe 3.0 x16. Physical dimensions: the CMP 40HX is 229 mm long, 111 mm tall, and 35 mm wide; the P40 is 267 mm long and 111 mm tall, with no width specified. Both are dual-slot. API support differs: the CMP 40HX supports DirectX 12 Ultimate (12_2), while the P40 is limited to DirectX 12 (12_1). Both support OpenGL 4.6 and Vulkan 1.4. Release dates are far apart—2021-02-24 for the CMP 40HX, 2016-09-12 for the P40—and the P40 has a predecessor (Tesla Maxwell) and successor (Tesla Volta), while the CMP 40HX has neither listed.