NVIDIA P106-090 vs NVIDIA Tesla K10 Comparison

The NVIDIA Tesla K10 and NVIDIA P106-090 are both end-of-life, no-display-output compute cards, but they represent two very different eras of GPU design. The Tesla K10 is a 2012-era dual-GPU Kepler monster, while the P106-090 is a 2017 mining-focused Pascal chip. Benchmark data reveals a clear split: the P106-090 dominates in raw compute scores, while the Tesla K10 relies on its architectural heritage. This analysis breaks down where each card wins, the fundamental design differences, and which workloads suit each GPU.

Where Each One Wins

The data is unambiguous in the single available head-to-head benchmark. The NVIDIA P106-090 wins the only direct comparison, taking the Geekbench OpenCL test with a score of 21,304 against the Tesla K10’s 14,029. That is a 34.1% lead for the P106-090, a massive margin that underscores the generational leap in compute efficiency between Kepler and Pascal.

The Tesla K10, however, is not without its own strengths, even if they don’t show up in the benchmark wins column. With zero wins in the head-to-head, the K10’s advantage is more about bandwidth and memory configuration than raw throughput. It offers 4 GB of GDDR5 on a 256-bit bus, delivering 160.0 GB/s of bandwidth. This is a broader memory pool than the P106-090’s 3 GB, and while the P106-090 has higher total bandwidth at 192.2 GB/s, the K10’s larger capacity could be more relevant for certain data-heavy workloads that need to hold more working set in VRAM.

For raw compute tasks like OpenCL workloads, the P106-090 is the clear winner. Its 2.352 TFLOPS of FP32 performance slightly edges out the K10’s 2.289 TFLOPS, and its newer Pascal architecture with support for DirectX 12 (12_1) and Vulkan 1.4 makes it more capable for modern APIs. The P106-090’s 73.49 GPixel/s pixel rate and 73.49 GTexel/s texture rate also show a balance that the K10 lacks; the K10’s 23.84 GPixel/s and 95.36 GTexel/s show a lopsided design favoring texturing over pixel output.

Architecture Differences

The architectural gap between these two cards is vast. The Tesla K10 is built on the GK104 chip using NVIDIA’s Kepler architecture, manufactured on a 28 nm process at TSMC. It packs 3,540 million transistors onto a 294 mm² die, yielding a transistor density of 12.0M per mm². The P106-090, in contrast, uses the GP106 chip with the Pascal architecture, built on a more advanced 16 nm process. It fits 4,400 million transistors into a smaller 200 mm² die, achieving a much higher density of 22.0M per mm². This density advantage is a direct result of the newer process node and is a key reason for the P106-090’s performance lead.

The core configurations differ dramatically. The Tesla K10 fields 1,536 shading units, 128 TMUs, and 32 ROPs. The P106-090 has fewer shading units at 768, fewer TMUs at 48, but more ROPs at 48. This means the K10 has a wider shader array and much higher texture throughput, while the P106-090 has better pixel processing capability per clock. The K10’s texture rate of 95.36 GTexel/s is nearly 30% higher than the P106-090’s 73.49 GTexel/s, but the P106-090’s pixel rate is over three times higher.

Clock speeds also tell a story. The Tesla K10 has no listed base or boost clock, with memory running at 1250 MHz (5 Gbps effective). The P106-090 runs at a base of 1354 MHz and boosts to 1531 MHz, with memory at 2002 MHz (8 Gbps effective). The P106-090’s higher clocks, combined with its newer architecture, help explain its compute dominance.

Power and interface differences are stark. The Tesla K10 is a 225 W card requiring a 550 W power supply and both a 6-pin and 8-pin power connector. The P106-090 sips power at just 75 W with a 250 W power supply recommendation and a single 6-pin connector. Both are dual-slot cards, but the K10 is longer at 272 mm compared to the P106-090’s 250 mm. The bus interface is another major split: the K10 uses PCIe 3.0 x16, while the P106-090 is hobbled with PCIe 1.0 x1, a limitation from its mining origins that could bottleneck data transfer in some scenarios.

Head-to-Head Benchmarks

The only direct benchmark comparison available is Geekbench OpenCL, and it is a decisive victory for the P106-090. The P106-090 scores 21,304, while the Tesla K10 scores 14,029, giving the P106-090 a 34.1% advantage. This is not a marginal win; it is a clear generational stomping. The P106-090’s score is also reflected in its other benchmark results, which include a 3DMark Steel Nomad DX12 score of 509 and a Geekbench Vulkan score of 18,596, both of which are not available for the K10.

Context from the nearest rivals puts these scores in perspective. The Tesla K10’s 14,029 OpenCL score places it just 0.9% behind the NVIDIA GeForce GTX 680, and 1.1% ahead of the AMD Radeon RX 570X. It also sits 1.5% ahead of the NVIDIA RTX A2000 Mobile and 1.6% ahead of the AMD Radeon 660M. The P106-090’s 21,304 score, by contrast, is in a different league. Its rivals include the NVIDIA GeForce GTX 570, which it beats by 0.3%, and the AMD Radeon Pro 555, which it leads by 0.5%. The P106-090 also sits 0.5% ahead of the AMD Radeon HD 7770M and 0.5% ahead of the AMD Radeon RX 9070 XT. This shows that the P106-090’s OpenCL performance is not just better than the K10; it is competitive with a much newer class of GPUs, despite the P106-090’s mining-focused design.

The average benchmark score further illustrates the gap. The Tesla K10 averages 14,029 across its single benchmark, while the P106-090 averages 13,470 across three benchmarks. This lower average for the P106-090 is due to the inclusion of its 3DMark result, which is a much lower score on a different scale. In the specific OpenCL test where they both compete, the P106-090 is the clear winner.

FAQ

Q: Which card has better OpenCL performance?

A: The NVIDIA P106-090 wins decisively. It scores 21,304 in Geekbench OpenCL, which is 34.1% higher than the Tesla K10’s 14,029.

Q: How do the memory configurations compare?

A: The Tesla K10 has 4 GB of GDDR5 on a 256-bit bus with 160.0 GB/s bandwidth. The P106-090 has 3 GB of GDDR5 on a 192-bit bus with higher bandwidth at 192.2 GB/s.

Q: What is the power consumption difference?

A: The Tesla K10 is rated at 225 W with a suggested 550 W power supply, requiring both a 6-pin and 8-pin connector. The P106-090 is much more efficient at 75 W with a 250 W power supply recommendation and a single 6-pin connector.

Q: Do these cards have display outputs?

A: No. Neither the Tesla K10 nor the P106-090 has any display outputs, making them unsuitable for standard desktop use.

Q: What is the architectural difference in process nodes?

A: The Tesla K10 uses a 28 nm process with 3,540 million transistors on a 294 mm² die. The P106-090 uses a 16 nm process with 4,400 million transistors on a smaller 200 mm² die.

Q: Which card supports newer APIs?

A: The P106-090 supports DirectX 12 (12_1) and Vulkan 1.4. The Tesla K10 is limited to DirectX 12 (11_0) and Vulkan 1.2.175, though both support OpenGL 4.6.

The Verdict

The choice between these two cards is straightforward based on the data. The NVIDIA P106-090 is the superior compute performer, with a 34.1% lead in OpenCL and a more modern architecture that supports newer APIs. Its 2.352 TFLOPS FP32 performance, higher pixel rate, and significantly lower power draw of 75 W make it the more practical and powerful option for raw compute tasks. The P106-090’s 16 nm process and higher transistor density deliver more performance per watt and per square millimeter.

The Tesla K10, however, retains niche appeal. Its 4 GB memory capacity is larger, and its 256-bit bus provides a solid 160.0 GB/s of bandwidth. For workloads that are more memory-capacity-sensitive than bandwidth-hungry, the K10’s extra gigabyte could matter. Its 1,536 shading units and 128 TMUs also give it a texture throughput advantage at 95.36 GTexel/s, which could be relevant for certain texture-heavy computations. But with zero benchmark wins and a lower OpenCL score, these advantages are theoretical rather than demonstrated.

For most users, the P106-090 is the pick. It is newer, faster, more efficient, and has a lower power footprint. The Tesla K10 is a relic of a dual-GPU era, and while its architecture is interesting, the data shows it cannot keep pace with even a mining-class Pascal chip. The P106-090’s PCIe 1.0 x1 interface is a potential bottleneck, but for compute workloads that fit within its 3 GB memory, the performance is there. The verdict is clear: the P106-090 wins on compute, efficiency, and modern feature support, while the Tesla K10 offers only a larger memory pool as a consolation prize.