NVIDIA P106-090 vs NVIDIA Tesla M2090 Comparison

FAQ

Q: Which GPU has the higher average benchmark score?

A: The NVIDIA P106-090 has an average benchmark score of 13470, while the NVIDIA Tesla M2090 scores 13075. This places the P106-090 at the 54th percentile of all GPUs, one percentile point ahead of the Tesla M2090’s 53rd percentile.

Q: How large is the performance gap in the only shared benchmark?

A: In the Geekbench OpenCL test, the P106-090 scores 21304 versus the Tesla M2090’s 13075. That is a 62.9% advantage for the P106-090, making it the clear winner in the sole head-to-head comparison.

Q: Which GPU has more memory and a wider memory bus?

A: The Tesla M2090 has 6 GB of GDDR5 memory on a 384-bit bus, double the 3 GB of the P106-090. However, the P106-090’s 192-bit bus still achieves higher bandwidth at 192.2 GB/s versus 177.4 GB/s.

Q: What are the transistor and die size differences?

A: The P106-090 packs 4,400 million transistors on a 200 mm² die using a 16 nm process, yielding a density of 22.0M transistors per mm². The Tesla M2090 uses 3,000 million transistors on a much larger 520 mm² die at 40 nm, with a density of just 5.8M per mm².

Q: Do either of these cards support display outputs?

A: No. Both the P106-090 and the Tesla M2090 list “No outputs” as their display configuration. They are compute-only accelerators.

Q: Which GPU has the higher power consumption and what power connectors does it require?

A: The Tesla M2090 has a 250 W TDP and needs both a 6-pin and an 8-pin power connector, with a suggested 600 W PSU. The P106-090 draws only 75 W, uses a single 6-pin connector, and suggests a 250 W PSU.

Architecture Differences

The P106-090 and Tesla M2090 come from entirely different architectural eras. The P106-090 is built on NVIDIA’s Pascal architecture using the GP106 chip, fabricated on a 16 nm process at TSMC. The Tesla M2090 belongs to the older Fermi 2.0 architecture with the GF110 chip, made on a 40 nm process, also at TSMC. This process gap is stark: 16 nm versus 40 nm, which directly influences transistor density. The P106-090 crams 4,400 million transistors into 200 mm², while the Tesla M2090 spreads 3,000 million transistors across 520 mm². The density figures tell the story — 22.0M transistors per mm² for Pascal versus 5.8M for Fermi.

The shading unit counts differ as well. The P106-090 has 768 shading units, 48 TMUs, and 48 ROPs. The Tesla M2090 has 512 shading units but more TMUs at 64, with the same 48 ROPs. Neither card features ray tracing cores or tensor cores, as both predate those technologies. Clock behavior also diverges: the P106-090 has explicit base and boost clocks of 1354 MHz and 1531 MHz, while the Tesla M2090 lists no base or boost clock at all, only its memory clock of 924 MHz (3.7 Gbps effective). The P106-090’s memory runs at 2002 MHz (8 Gbps effective).

API support differs meaningfully. The P106-090 supports DirectX 12 (12_1), OpenGL 4.6, and Vulkan 1.4. The Tesla M2090 supports DirectX 12 (11_0) and OpenGL 4.6, but has no Vulkan support listed. The bus interfaces also contrast: the P106-090 uses a modest PCIe 1.0 x1 connection, while the Tesla M2090 uses the far wider PCIe 2.0 x16 interface. Both are dual-slot cards with no display outputs, and both are end-of-life products. The P106-090 launched in 2017 within the Mining GPUs generation, while the Tesla M2090 launched in 2011 as part of the Tesla Fermi (x20xx) line, with a predecessor simply named "Tesla" and a successor in "Tesla Kepler."

Head-to-Head Benchmarks

Only one benchmark appears in both GPUs’ result sets, and it is a decisive one. In Geekbench OpenCL, the P106-090 scores 21304 against the Tesla M2090’s 13075. The delta is 62.9% in favor of the P106-090. That is not a marginal win; it is a dominant margin that dwarfs the differences seen in their nearest-rival comparisons.

To frame this properly, consider where each card sits among its nearest rivals. The P106-090’s average score of 13470 is within 0.3% of the GeForce GTX 570 (13515) and 0.5% of the Radeon Pro 555 (13407). It is essentially a dead heat with those cards. The Tesla M2090’s average score of 13075 puts it 0.7% ahead of the GeForce GTX 1660 SUPER (12986) and 1.1% ahead of the Radeon RX 580 (12928), but 0.9% behind the GeForce GTX 950 (13189). So the Tesla M2090 is competitive with mid-range GPUs from later generations, but the P106-090’s OpenCL result is in a different league entirely.

The wins tally reflects this: the P106-090 claims 1 win in the head-to-head column, the Tesla M2090 claims 0. No benchmark in the shared set favors the Tesla. The P106-090 also benefits from three benchmark results in total — including 3DMark Steel Nomad DX12 (509) and Geekbench Vulkan (18596) — whereas the Tesla M2090 has only the single OpenCL result. Those additional benchmarks are not directly comparable to the Tesla, but they show the P106-090 has broader testing coverage and a higher ceiling in synthetic workloads.

The gap in OpenCL is particularly telling because both cards are compute-oriented with no display outputs. The P106-090’s 2.352 TFLOPS of FP32 performance versus the Tesla M2090’s 1,332.2 GFLOPS explains much of the delta. Even though the Tesla has more memory (6 GB vs 3 GB) and a wider bus (384-bit vs 192-bit), the P106-090’s newer architecture and higher clocks push its bandwidth higher as well — 192.2 GB/s versus 177.4 GB/s. In every measurable performance metric, the Pascal part wins.

The Verdict

The data is unambiguous: the NVIDIA P106-090 is the superior GPU for compute workloads. It wins the only shared benchmark by 62.9%, has a higher average benchmark score (13470 vs 13075), and sits at a higher percentile rank (54 vs 53). The architectural advantages are overwhelming — a 16 nm process versus 40 nm, 4,400 million transistors versus 3,000 million, and nearly double the FP32 throughput. For anyone choosing between these two for OpenCL-based tasks, the P106-090 is the clear pick.

The Tesla M2090 is not without merit. It offers double the memory capacity (6 GB vs 3 GB) and a wider 384-bit memory bus, which could matter for workloads that exceed the P106-090’s 3 GB frame buffer. It also has more TMUs (64 vs 48), which could assist in texture-heavy operations. But its 177.4 GB/s bandwidth is actually lower than the P106-090’s 192.2 GB/s, negating the bus-width advantage. The Tesla’s 250 W TDP and 600 W suggested PSU are also far more demanding than the P106-090’s 75 W and 250 W requirements.

For power efficiency, the P106-090 wins by a wide margin. It delivers significantly more performance while drawing one-third the power. The Tesla M2090’s only practical advantage is memory capacity. If a workload requires more than 3 GB, the Tesla is the only choice. Otherwise, the P106-090 beats it on speed, efficiency, and API support — notably Vulkan, which the Tesla lacks entirely.

The verdict is straightforward: the P106-090 wins on performance, efficiency, and modern features. The Tesla M2090 is a legacy product with a capacity edge but a speed disadvantage. For new deployments, the P106-090 is the rational selection.

Specification Differences

The two GPUs differ across nearly every specification field. The P106-090 uses the GP106 chip on a 16 nm process with 4,400 million transistors on a 200 mm² die. The Tesla M2090 uses the GF110 chip on a 40 nm process with 3,000 million transistors on a 520 mm² die. Transistor density is 22.0M per mm² versus 5.8M per mm².

Clock speeds: the P106-090 has a base clock of 1354 MHz and a boost of 1531 MHz, with memory at 2002 MHz (8 Gbps effective). The Tesla M2090 has no base or boost clock listed, and its memory runs at 924 MHz (3.7 Gbps effective).

Memory: the P106-090 has 3 GB of GDDR5 on a 192-bit bus with 192.2 GB/s bandwidth. The Tesla M2090 has 6 GB of GDDR5 on a 384-bit bus with 177.4 GB/s bandwidth.

Compute units: the P106-090 has 768 shading units, 48 TMUs, and 48 ROPs. The Tesla M2090 has 512 shading units, 64 TMUs, and 48 ROPs. Pixel rate is 73.49 GPixel/s for the P106-090 versus 20.83 GPixel/s for the Tesla. Texture rate is 73.49 GTexel/s versus 41.66 GTexel/s. FP32 is 2.352 TFLOPS versus 1,332.2 GFLOPS. The P106-090 also has an FP16 rating of 36.74 GFLOPS (1:64), while the Tesla has none.

Power: the P106-090 has a 75 W TDP with a single 6-pin connector and a 250 W suggested PSU. The Tesla M2090 has a 250 W TDP with a 6-pin and 8-pin connector and a 600 W suggested PSU.

Bus and outputs: the P106-090 uses PCIe 1.0 x1; the Tesla uses PCIe 2.0 x16. Both have no display outputs. API support: the P106-090 has DirectX 12 (12_1), OpenGL 4.6, and Vulkan 1.4. The Tesla has DirectX 12 (11_0) and OpenGL 4.6, but no Vulkan.

Dimensions are nearly identical: the P106-090 is 250 mm (9.8 inches) long, the Tesla is 248 mm (9.8 inches). Both are dual-slot, end-of-life products. The P106-090 released in 2017 within the Mining GPUs generation; the Tesla M2090 released in 2011 with a predecessor "Tesla" and successor "Tesla Kepler."

Where Each One Wins

NVIDIA P106-090 wins on raw compute performance. Its Geekbench OpenCL score of 21304 is 62.9% higher than the Tesla’s 13075. The FP32 throughput of 2.352 TFLOPS versus 1,332.2 GFLOPS confirms this. For any OpenCL-heavy workload — GPU compute, scientific simulation, or machine learning inference — the P106-090 is faster by a wide margin.

NVIDIA P106-090 wins on power efficiency. At 75 W with a 250 W suggested PSU, it delivers superior performance at one-third the power draw of the Tesla M2090’s 250 W and 600 W suggested PSU. This makes it far easier to integrate into existing systems without power supply upgrades.

NVIDIA P106-090 wins on modern features. It supports Vulkan 1.4, while the Tesla has no Vulkan support. Its DirectX 12 (12_1) is newer than the Tesla’s DirectX 12 (11_0). The 16 nm process and higher transistor density also indicate a more advanced design with better architectural efficiency.

NVIDIA Tesla M2090 wins on memory capacity. With 6 GB versus 3 GB, the Tesla can handle larger datasets that exceed the P106-090’s memory limit. This is critical for workloads like large matrix operations or rendering scenes that require more than 3 GB of frame buffer.

NVIDIA Tesla M2090 wins on memory bus width. The 384-bit bus is double the P106-090’s 192-bit bus. Although the P106-090’s higher clock speed gives it more bandwidth overall, the wider bus may provide better latency characteristics for certain access patterns.

NVIDIA Tesla M2090 wins on texture units. Its 64 TMUs outnumber the P106-090’s 48. This could benefit texture-heavy compute tasks, though the P106-090’s texture rate of 73.49 GTexel/s still far exceeds the Tesla’s 41.66 GTexel/s.

Use-case split: Choose the P106-090 for general compute, OpenCL acceleration, power-constrained environments, or any workload needing Vulkan support. Choose the Tesla M2090 only when memory capacity above 3 GB is a hard requirement, and accept its substantial speed and efficiency penalties.