AMD Radeon HD 8870M vs NVIDIA GeForce GTX 660 Comparison

# Head-to-Head Benchmarks

The sole direct benchmark comparison available between these two GPUs is Geekbench OpenCL, and the results are decisive. The NVIDIA GeForce GTX 660 scores 11,347, while the AMD Radeon HD 8870M manages 8,462. That translates to a 34.1% advantage for the NVIDIA card — a substantial margin that dwarfs the differences typically seen between cards in adjacent performance tiers. For context, the GTX 660 sits at the 45th percentile among all GPUs in the database, while the HD 8870M rests just below at the 43rd percentile; despite being only two percentile points apart globally, the raw compute gap in this specific workload is anything but marginal.

The GTX 660's nearest rivals in average benchmark score include the NVIDIA TITAN V CEO Edition at 9,037 (a mere 0.2% behind) and the GeForce GTX 560 at 9,058 (0.4% behind). On the AMD side, the Radeon 550X comes within 1.2% and the Radeon Pro WX 5100 trails by 1.8%. The HD 8870M's nearest competitors tell a similar story: the GeForce MX330 is 0.1% ahead, the Radeon 880M is 0.3% ahead, and the GTX 675MX is 0.4% behind. The Intel Arc A380 leads the HD 8870M by 1.1%. What these neighbor comparisons reveal is that both cards are clustered in a dense field of mid-range performers, yet the head-to-head OpenCL gap between them is unusually wide for cards that occupy such similar overall percentile positions.

Looking at the broader benchmark picture, the GTX 660 has three recorded scores: 4,305 in Geekbench Metal, 11,347 in OpenCL, and 11,415 in Vulkan. Its average across all benchmarks is 9,022. The HD 8870M, by contrast, has only one recorded benchmark — that same OpenCL test at 8,462 — which also serves as its average. The absence of Vulkan and Metal data for the AMD card is itself informative: the GTX 660 demonstrates versatility across multiple compute APIs, whereas the HD 8870M's profile is built entirely on a single OpenCL result. This limited data set means the 34.1% delta is the only quantitative head-to-head signal available, but it is a strong one.

# The Verdict

From the data alone, the choice is unambiguous for compute-heavy workloads: the NVIDIA GeForce GTX 660 wins the only direct comparison, and it wins by 34.1%. The GTX 660 also offers higher peak compute throughput — 1.981 TFLOPS FP32 versus 992.0 GFLOPS for the HD 8870M — which is exactly double the AMD card's figure. The average benchmark scores reinforce this hierarchy: 9,022 versus 8,462, a 6.6% advantage for NVIDIA even when accounting for the GTX 660's additional Vulkan and Metal results pulling its average upward.

That said, the HD 8870M is not without its own positioning. It sits within 0.1% of the GeForce MX330 and 0.3% of the Radeon 880M, meaning it holds its own against a set of newer mobile and entry-level discrete GPUs. Its percentile ranking at 43 is only two points shy of the GTX 660's 45, suggesting that in workloads not captured by OpenCL — particularly those that favor GCN architecture characteristics — the gap could narrow or even reverse. The data does not contain any such counter-example, however.

The practical verdict: if the workload is OpenCL compute, the GTX 660 is the clear pick. If the workload is unknown or varied, the GTX 660's broader benchmark coverage (three tests versus one) provides more confidence in consistent performance. The HD 8870M's single data point leaves too much uncertainty for a recommendation in its favor, despite its respectable showing against its own nearest rivals.

# Architecture Differences

The two GPUs come from fundamentally different design philosophies. The GTX 660 uses the GK106 chip built on NVIDIA's Kepler architecture, fabricated by TSMC on a 28 nm process. It packs 2,540 million transistors into a 221 mm² die, yielding a transistor density of 11.5 million per square millimeter. The HD 8870M uses AMD's Venus chip based on Graphics Core Next 1.0, also on TSMC's 28 nm node, but with 1,500 million transistors on a much smaller 123 mm² die, achieving a higher density of 12.2 million per square millimeter. AMD fits more transistors per area but has fewer total resources to work with.

The compute resource allocation diverges sharply. The GTX 660 fields 960 shading units, 80 texture mapping units, and 24 ROPs. The HD 8870M counters with 640 shading units, 40 TMUs, and 16 ROPs. These are not proportional differences — NVIDIA has 50% more shaders, 100% more TMUs, and 50% more ROPs. The pixel rate tells the story: 20.64 GPixel/s for the GTX 660 versus 12.40 GPixel/s for the HD 8870M. Texture rate is even more lopsided: 82.56 GTexel/s versus 31.00 GTexel/s — a 2.66x advantage for NVIDIA.

Clock behavior also reflects architectural intent. The GTX 660 runs at a 980 MHz base with a 1032 MHz boost, while the HD 8870M operates at 725 MHz base and 775 MHz boost. The NVIDIA card is clocked roughly 35% higher at base and 33% higher at boost. Kepler's design favors higher clocks and wider execution resources, while GCN 1.0 leans on a more modest configuration with lower clock targets. The FP32 throughput difference — 1.981 TFLOPS versus 992.0 GFLOPS — is exactly 2x, which tracks the combination of more shaders and higher clocks.

Memory architecture adds another layer of divergence. The GTX 660 uses a 192-bit bus with 2 GB GDDR5 at 6 Gbps effective, producing 144.2 GB/s of bandwidth. The HD 8870M has a narrower 128-bit bus with 2 GB GDDR5 at 4.5 Gbps effective, yielding 72.00 GB/s. That is exactly half the bandwidth. Both support PCIe 3.0 x16 and DirectX 12, though the GTX 660 is rated at DirectX 12 (11_0) while the HD 8870M reaches 12 (11_1). Both support OpenGL 4.6; Vulkan support is nearly identical at 1.2.175 versus 1.2.170.

# Specification Differences

The published specifications differ across nearly every measurable dimension. Process node and foundry are identical (28 nm, TSMC), but transistor count, die size, and density all favor different interpretations: the GTX 660 has more transistors (2,540M vs 1,500M), a larger die (221 mm² vs 123 mm²), but lower density (11.5M/mm² vs 12.2M/mm²). Clock speeds are higher on the GTX 660 for both base (980 vs 725 MHz) and boost (1032 vs 775 MHz), and memory runs faster at 6 Gbps effective versus 4.5 Gbps.

Memory bus width is 192-bit versus 128-bit, and bandwidth is 144.2 GB/s versus 72.00 GB/s — a clean 2x difference. Shading units are 960 versus 640, TMUs are 80 versus 40, and ROPs are 24 versus 16. Pixel rate (20.64 vs 12.40 GPixel/s), texture rate (82.56 vs 31.00 GTexel/s), and FP32 (1.981 TFLOPS vs 992.0 GFLOPS) all favor the GTX 660 by wide margins.

The GTX 660 has a TDP of 140 W, is dual-slot, requires a single 6-pin power connector, and suggests a 300 W PSU. The HD 8870M has no TDP, slot width, power connector, or PSU specification listed — consistent with its mobile-oriented design. The GTX 660 offers display outputs (2x DVI, 1x HDMI 1.4a, 1x DisplayPort 1.2); the HD 8870M lists none. The GTX 660 is 241 mm (9.5 inches) long; the HD 8870M has no dimensions. Release dates differ by roughly seven months: September 5, 2012 for the GTX 660, March 31, 2013 for the HD 8870M. The GTX 660 launched at 229 USD MSRP; the HD 8870M has no launch price on record.

# FAQ

Q: Which GPU wins the Geekbench OpenCL benchmark?

A: The NVIDIA GeForce GTX 660 scores 11,347 versus the AMD Radeon HD 8870M's 8,462, giving NVIDIA a 34.1% lead in this test.

Q: How do their average benchmark scores compare?

A: The GTX 660 has an average score of 9,022 across three tests (Metal, OpenCL, Vulkan), while the HD 8870M averages 8,462 from its single OpenCL result.

Q: What is the FP32 compute difference?

A: The GTX 660 delivers 1.981 TFLOPS, exactly double the HD 8870M's 992.0 GFLOPS.

Q: Do both cards support the same APIs?

A: Both support DirectX 12 and OpenGL 4.6. The GTX 660 supports Vulkan 1.2.175 and the HD 8870M supports Vulkan 1.2.170. The GTX 660's DirectX 12 rating is 11_0, while the HD 8870M's is 11_1.

Q: What memory bandwidth does each card provide?

A: The GTX 660 offers 144.2 GB/s over a 192-bit bus, while the HD 8870M provides 72.00 GB/s over a 128-bit bus — exactly half.

Q: How do the cards rank among all GPUs?

A: The GTX 660 sits at the 45th percentile, while the HD 8870M sits at the 43rd percentile.

# Where Each One Wins

The GTX 660 wins the only direct benchmark comparison, and the win is decisive at 34.1% in OpenCL. It also wins every raw specification battle that matters for compute throughput: double the FP32, double the memory bandwidth, 50% more shading units, double the TMUs, 50% more ROPs, and significantly higher clock speeds across the board. For any workload that scales with raw compute or memory throughput — OpenCL compute, texture-heavy rendering, pixel-fill operations — the data points entirely to the GTX 660.

The HD 8870M's case rests on its positioning among its nearest rivals rather than direct comparison. It sits within 0.3% of the Radeon 880M and 0.1% of the GeForce MX330, suggesting it remains competitive with much newer entry-level parts. Its higher transistor density (12.2M/mm² versus 11.5M/mm²) indicates a more compact, potentially more power-efficient design — though no TDP is listed to confirm. For scenarios where the GTX 660's 140 W TDP and dual-slot footprint are prohibitive, the HD 8870M's unspecified power profile and mobile heritage could make it the more practical choice in constrained environments.

The use-case split is therefore clear: desktop performance computing favors the GTX 660 without qualification, while the HD 8870M's appeal is limited to situations where the GTX 660's physical requirements cannot be met. The data does not support any scenario where the HD 8870M outperforms the GTX 660 in a direct, measurable way — it simply occupies a different niche with lower absolute capability.