NVIDIA GeForce GTX 660 vs NVIDIA Tesla C2070 Comparison

The NVIDIA Tesla C2070 and GeForce GTX 660 represent two distinct philosophies from the same manufacturer, separated by a generation and aimed at entirely different workloads. The data shows a fascinating reversal of expectations: the professional compute card, despite its workstation pedigree and larger memory pool, falls behind the consumer gaming card in raw compute benchmarks. This analysis examines the numbers to determine what each card truly excels at, and why the architecture gap matters more than the product category.

Head-to-Head Benchmarks

The only direct benchmark comparison available is the Geekbench OpenCL test, and it delivers a decisive verdict. The GeForce GTX 660 scores 11,347, while the Tesla C2070 trails at 9,716. That is a 14.4% advantage for the GTX 660, a significant margin that flips the expected hierarchy. The Tesla card, designed for scientific and professional compute, is outperformed by a card aimed at gamers in a general-purpose compute workload.

This result is even more striking when placed in context. The Tesla C2070's nearest rival in the database is the NVIDIA Tesla M10, which scores 9,724 — a mere 0.1% difference. The Quadro P4000 sits just 0.5% ahead at 9,665, and the AMD Radeon Pro WX 2100 is 0.7% ahead at 9,653. The C2070 is essentially clustered with these mid-range professional cards, all hovering within a 1% band. The GTX 660, meanwhile, sits in a different neighborhood entirely. Its nearest rivals include the NVIDIA TITAN V CEO Edition (9,037, just 0.2% behind) and the GeForce GTX 560 (9,058, 0.4% behind). The GTX 660's score places it 1.2% ahead of the AMD Radeon 550X and 1.8% ahead of the AMD Radeon Pro WX 5100.

The deltaPct of -14.4% in the head-to-head test confirms that the GTX 660 is not just marginally better — it is substantially faster in OpenCL. The data suggests that the Kepler architecture's efficiency gains over Fermi are real and measurable, not just theoretical marketing claims. The GTX 660 wins the only benchmark where both cards are tested, and it wins by a wide margin.

Where Each One Wins

The benchmark data shows only one win, going to the GeForce GTX 660. There are zero wins for the Tesla C2070 in head-to-head comparison. However, the specification sheet reveals where the Tesla card could claim supremacy in scenarios not captured by the OpenCL test.

The Tesla C2070 offers 6 GB of GDDR5 memory, three times the GTX 660's 2 GB. For workloads that require large datasets resident in VRAM — such as certain scientific simulations, rendering scenes with massive textures, or machine learning inference with large batch sizes — the Tesla's capacity is a clear advantage. The GTX 660 would run out of memory and be forced to spill to system RAM, which would cripple performance regardless of its compute advantage.

The Tesla also has a wider 384-bit memory bus compared to the GTX 660's 192-bit bus. Despite this, the bandwidth figures are nearly identical: 143.4 GB/s for the Tesla versus 144.2 GB/s for the GTX 660. The GTX 660 achieves this with a higher memory clock (6 Gbps effective versus 3 Gbps effective), which compensates for its narrower bus. In memory-bound workloads that favor bandwidth over capacity, the two cards are effectively tied.

The GTX 660 wins on raw compute throughput by a wide margin. Its FP32 performance is listed at 1.981 TFLOPS, versus 1,027.7 GFLOPS for the Tesla — a 92.7% advantage. This is the primary driver of its OpenCL victory. The GTX 660 also has more shading units (960 versus 448) and more texture mapping units (80 versus 56), which explains its superior texture rate of 82.56 GTexel/s versus 32.14 GTexel/s.

Architecture Differences

The two cards are built on fundamentally different architectures from different generations. The Tesla C2070 uses the GF100 chip, based on the Fermi architecture, manufactured on a 40 nm process at TSMC. The GTX 660 uses the GK106 chip, based on the Kepler architecture, also from TSMC but on a more advanced 28 nm node.

The process node difference is dramatic. The 40 nm Fermi chip packs 3,100 million transistors onto a 529 mm² die, yielding a transistor density of 5.9 million per square millimeter. The 28 nm Kepler chip packs 2,540 million transistors onto a much smaller 221 mm² die, achieving 11.5 million transistors per square millimeter — nearly double the density. This explains why the GTX 660 achieves higher performance with fewer transistors: the newer process allows for a more efficient design.

The architectural philosophy also differs. Fermi was NVIDIA's first generation to support compute capabilities that were ahead of its time, but it was power-hungry and complex. Kepler, released a generation later, focused on efficiency and higher clock speeds. The GTX 660's base clock of 980 MHz and boost clock of 1,032 MHz are significantly higher than the Tesla's memory clock of 747 MHz (the Tesla's base and boost clocks are not listed in the data, but its memory clock is lower). The GTX 660 also supports PCIe 3.0 x16, while the Tesla is limited to PCIe 2.0 x16 — a factor that affects data transfer rates between the GPU and host system.

The GTX 660 has a clear API advantage in Vulkan support, listing version 1.2.175, while the Tesla has no Vulkan support listed. Both cards support DirectX 12 (11_0) and OpenGL 4.6. The GTX 660 also has a lower TDP of 140 W versus 238 W for the Tesla, and requires only a single 6-pin power connector and a 300 W PSU, compared to the Tesla's 1x 6-pin + 1x 8-pin configuration and 550 W PSU recommendation.

FAQ

Q: Which card is faster in OpenCL compute?

A: The GeForce GTX 660 scores 11,347 in Geekbench OpenCL, while the Tesla C2070 scores 9,716. The GTX 660 is 14.4% faster in this test.

Q: Does the Tesla C2070 have any memory advantage?

A: Yes. The Tesla has 6 GB of GDDR5 memory compared to the GTX 660's 2 GB. The Tesla also has a wider 384-bit memory bus versus the GTX 660's 192-bit bus, although the effective bandwidth is nearly identical (143.4 GB/s versus 144.2 GB/s).

Q: Why does the GTX 660 have a higher FP32 throughput?

A: The GTX 660 lists 1.981 TFLOPS of FP32 performance versus 1,027.7 GFLOPS for the Tesla. This is driven by its 960 shading units (versus 448) and higher clock speeds, enabled by the more efficient 28 nm Kepler architecture.

Q: What is the transistor density difference between the two chips?

A: The Tesla's GF100 chip has a density of 5.9 million transistors per mm² on a 40 nm process, while the GTX 660's GK106 chip achieves 11.5 million per mm² on a 28 nm process. The Kepler chip is nearly twice as dense.

Q: Does the Tesla support Vulkan?

A: No Vulkan support is listed for the Tesla C2070. The GTX 660 lists Vulkan support at version 1.2.175.

Q: Which card has a higher pixel rate?

A: The GTX 660 has a pixel rate of 20.64 GPixel/s, while the Tesla C2070 achieves 16.07 GPixel/s. The GTX 660 also has a higher texture rate of 82.56 GTexel/s versus 32.14 GTexel/s.

The Verdict

The data paints a clear picture: the GeForce GTX 660 is the superior compute performer in the only benchmark where both are tested. Its 14.4% OpenCL lead, combined with nearly double the FP32 throughput, makes it the obvious choice for general-purpose compute workloads that fit within its 2 GB memory limit.

The Tesla C2070's justification lies entirely in its 6 GB memory capacity. For workloads that require more than 2 GB of VRAM, the GTX 660 is not an option — it will fail or degrade to system memory. The Tesla's wider memory bus and identical bandwidth (143.4 GB/s versus 144.2 GB/s) suggest it can handle large datasets without a bandwidth penalty. If the workload fits in 2 GB, the GTX 660 is faster in every measurable way. If it does not, the Tesla is the only viable choice.

The architecture comparison is instructive. The Kepler architecture on 28 nm delivers more than double the transistor density (11.5M vs 5.9M per mm²) and uses 98 W less power (140 W vs 238 W). The GTX 660 also supports PCIe 3.0 and Vulkan, making it more modern in platform compatibility. The Tesla's sole advantages are memory capacity and the professional Tesla branding, which historically implies different driver validation and support — though the data cannot quantify that.

For a user with workloads under 2 GB, the GTX 660 is the clear winner. For large-scale data processing that exceeds 2 GB, the Tesla C2070 is the only choice, despite its lower compute throughput. The GTX 660's launch MSRP was 229 USD, which reflects its consumer positioning, while the Tesla's pricing is not listed.

Specification Differences

| Specification | NVIDIA Tesla C2070 | NVIDIA GeForce GTX 660 |

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

| Architecture | Fermi | Kepler |

| Chip | GF100 | GK106 |

| Process Node | 40 nm | 28 nm |

| Transistors | 3,100 million | 2,540 million |

| Die Size | 529 mm² | 221 mm² |

| Transistor Density | 5.9M / mm² | 11.5M / mm² |

| Base Clock | Not listed | 980 MHz |

| Boost Clock | Not listed | 1032 MHz |

| Memory Clock | 747 MHz (3 Gbps effective) | 1502 MHz (6 Gbps effective) |

| Memory Size | 6 GB | 2 GB |

| Memory Bus Width | 384 bit | 192 bit |

| Memory Bandwidth | 143.4 GB/s | 144.2 GB/s |

| Shading Units | 448 | 960 |

| TMUs | 56 | 80 |

| ROPs | 48 | 24 |

| Pixel Rate | 16.07 GPixel/s | 20.64 GPixel/s |

| Texture Rate | 32.14 GTexel/s | 82.56 GTexel/s |

| FP32 | 1,027.7 GFLOPS | 1.981 TFLOPS |

| TDP | 238 W | 140 W |

| Power Connectors | 1x 6-pin + 1x 8-pin | 1x 6-pin |

| Suggested PSU | 550 W | 300 W |

| Bus Interface | PCIe 2.0 x16 | PCIe 3.0 x16 |

| Display Outputs | 1x DVI | 2x DVI, 1x HDMI 1.4a, 1x DisplayPort 1.2 |

| Vulkan Support | Not listed | 1.2.175 |

| Release Date | 2011-07-24 | 2012-09-05 |

| Predecessor | Tesla | GeForce 500 |

| Successor | Tesla Kepler | GeForce 700 |