AMD Radeon HD 8850M vs NVIDIA GeForce GTX 670M Comparison

FAQ

Q: Which GPU has the higher OpenCL benchmark score?

A: The AMD Radeon HD 8850M records a Geekbench OpenCL score of 7447, while the NVIDIA GeForce GTX 670M scores 6513. The AMD part leads by 14.3% in this specific test.

Q: How do these two GPUs compare in transistor count and die size?

A: The NVIDIA GeForce GTX 670M uses 1,950 million transistors on a 332 mm² die, whereas the AMD Radeon HD 8850M uses 1,500 million transistors on a much smaller 123 mm² die. This means the AMD chip has a transistor density of 12.2M per mm² versus 5.9M per mm² for the NVIDIA chip.

Q: What are the memory configurations of each GPU?

A: The AMD Radeon HD 8850M has 2 GB of GDDR5 memory on a 128-bit bus, delivering 64.00 GB/s of bandwidth. The NVIDIA GeForce GTX 670M has 1536 MB of GDDR5 memory on a wider 192-bit bus, delivering 72.00 GB/s of bandwidth.

Q: Which GPU has a higher pixel fill rate?

A: The AMD Radeon HD 8850M achieves a pixel rate of 10.00 GPixel/s, which is higher than the NVIDIA GeForce GTX 670M's 8.372 GPixel/s. The AMD part has 16 ROPs compared to 24 ROPs on the NVIDIA part, but its higher clock speed compensates.

Q: What is the difference in shading units and texture units?

A: The AMD Radeon HD 8850M has 640 shading units and 40 texture mapping units. The NVIDIA GeForce GTX 670M has 336 shading units and 56 texture mapping units. Despite fewer shading units, the NVIDIA part has more TMUs.

Q: What API support differences exist between the two?

A: The AMD Radeon HD 8850M supports DirectX 12 (11_1), OpenGL 4.6, and Vulkan 1.2.170. The NVIDIA GeForce GTX 670M supports DirectX 12 (11_0) and OpenGL 4.6, but has no Vulkan support listed.

Architecture Differences

The AMD Radeon HD 8850M is built on the GCN 1.0 architecture, using the Venus chip manufactured on a 28 nm process at TSMC. This is a modern, efficient design that packs 1,500 million transistors into a compact 123 mm² die. The GCN architecture was designed with compute workloads in mind, which is reflected in the higher shading unit count of 640. The 28 nm node allows for a transistor density of 12.2M per mm², a significant advantage over the older process used by the competitor.

The NVIDIA GeForce GTX 670M uses the Fermi 2.0 architecture with the GF114 chip, fabricated on a 40 nm process also at TSMC. This older process node results in a substantially larger die size of 332 mm² to accommodate 1,950 million transistors. The transistor density is only 5.9M per mm², roughly half that of the AMD chip. Fermi 2.0 was designed with a different philosophy, favoring fewer, more complex streaming multiprocessors. This is evident in the 336 shading units, which are organized differently than the AMD design.

The memory subsystems diverge significantly. The AMD Radeon HD 8850M uses a 128-bit memory bus with 2 GB of GDDR5 memory clocked at 1000 MHz (4 Gbps effective), yielding 64.00 GB/s of bandwidth. The NVIDIA GeForce GTX 670M uses a wider 192-bit bus with 1536 MB of GDDR5 memory clocked at 750 MHz (3 Gbps effective), producing 72.00 GB/s of bandwidth. The NVIDIA part trades capacity for a wider bus and higher aggregate bandwidth.

Power delivery and form factor also differ. The NVIDIA GeForce GTX 670M has a TDP of 75 W, uses an MXM Module slot width, and requires no power connectors. It connects via an MXM-B (3.0) interface and has display outputs described as "Portable Device Dependent". The AMD Radeon HD 8850M does not list a TDP, slot width, or power connector requirements in the database, but it uses a standard PCIe 3.0 x16 bus interface. The NVIDIA part is clearly designed for modular laptop implementations, while the AMD part follows a more conventional mobile GPU interface.

Both GPUs are end-of-life products. The AMD Radeon HD 8850M was released on 2013-03-31, with the London as predecessor and Gem System as successor. The NVIDIA GeForce GTX 670M was released on 2012-03-21, with GeForce 500M as predecessor and GeForce 700M as successor. The NVIDIA part is a year older, which explains some of the architectural differences.

Head-to-Head Benchmarks

The only recorded head-to-head benchmark in the database is Geekbench OpenCL. In this test, the AMD Radeon HD 8850M scores 7447 points, while the NVIDIA GeForce GTX 670M scores 6513 points. This represents a 14.3% advantage for the AMD part. This is a substantial margin in a compute-oriented workload, which aligns with the architectural strengths of GCN 1.0. The AMD chip's 640 shading units provide a raw compute advantage, and the higher clock speeds of 575 MHz base and 625 MHz boost further contribute to the result.

The NVIDIA GeForce GTX 670M's FP32 throughput is 803.7 GFLOPS, which is remarkably close to the AMD Radeon HD 8850M's 800.0 GFLOPS. Despite this near parity in theoretical floating-point performance, the AMD part wins the actual benchmark by a clear margin. The difference may stem from driver optimizations, memory latency characteristics, or the efficiency of the GCN scheduler in OpenCL workloads. The AMD memory runs at 1000 MHz versus 750 MHz on the NVIDIA part, which could reduce memory latency penalties in compute tasks.

The texture and pixel rates tell a different story. The NVIDIA GeForce GTX 670M has a texture rate of 33.49 GTexel/s, which is 34% higher than the AMD Radeon HD 8850M's 25.00 GTexel/s. This comes from 56 TMUs on the NVIDIA part versus 40 on the AMD part. However, the AMD part has a higher pixel rate at 10.00 GPixel/s versus 8.372 GPixel/s, despite having only 16 ROPs versus 24. This is due to the AMD clock advantage and higher ROP efficiency per clock.

The benchmark data shows a single decisive win for the AMD Radeon HD 8850M. The wins tally is 1 for AMD and 0 for NVIDIA. The delta of 14.3% is substantial, placing the AMD part clearly ahead in the recorded compute test. For context, the AMD part's nearest rivals include the Intel UHD Graphics 750 at 7441 (0.1% behind), the AMD Radeon R7 350 at 7425 (0.3% behind), the NVIDIA GeForce GTX 1650 at 7472 (0.3% ahead), and the Intel Arc A310 at 7550 (1.4% ahead). The NVIDIA GeForce GTX 670M's nearest rivals are the NVIDIA GeForce GT 555M at 6493 (0.3% behind), the NVIDIA Quadro M5000M at 6481 (0.5% behind), the AMD Radeon Vega 10 Mobile at 6476 (0.6% behind), and the Intel UHD Graphics P750 at 6554 (0.6% ahead).

Specification Differences

| Specification | AMD Radeon HD 8850M | NVIDIA GeForce GTX 670M |

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

| Architecture | GCN 1.0 | Fermi 2.0 |

| Process Node | 28 nm | 40 nm |

| Transistors | 1,500 million | 1,950 million |

| Die Size | 123 mm² | 332 mm² |

| Transistor Density | 12.2M / mm² | 5.9M / mm² |

| Base Clock | 575 MHz | Not listed |

| Boost Clock | 625 MHz | Not listed |

| Memory Clock | 1000 MHz (4 Gbps effective) | 750 MHz (3 Gbps effective) |

| Memory Size | 2 GB | 1536 MB |

| Memory Bus Width | 128 bit | 192 bit |

| Memory Bandwidth | 64.00 GB/s | 72.00 GB/s |

| Shading Units | 640 | 336 |

| TMUs | 40 | 56 |

| ROPs | 16 | 24 |

| Pixel Rate | 10.00 GPixel/s | 8.372 GPixel/s |

| Texture Rate | 25.00 GTexel/s | 33.49 GTexel/s |

| FP32 | 800.0 GFLOPS | 803.7 GFLOPS |

| TDP | Not listed | 75 W |

| Slot Width | Not listed | MXM Module |

| Power Connectors | Not listed | None |

| Bus Interface | PCIe 3.0 x16 | MXM-B (3.0) |

| DirectX | 12 (11_1) | 12 (11_0) |

| Vulkan | 1.2.170 | Not listed |

| Release Date | 2013-03-31 | 2012-03-21 |

| Predecessor | London | GeForce 500M |

| Successor | Gem System | GeForce 700M |

Where Each One Wins

The AMD Radeon HD 8850M is the clear winner in compute-oriented workloads, as demonstrated by its 14.3% lead in Geekbench OpenCL. This makes it the preferable choice for tasks that leverage general-purpose GPU computing, such as OpenCL-accelerated applications, video encoding with compute shaders, or physics simulations. The GCN architecture was explicitly designed with compute efficiency in mind, and the higher shading unit count of 640 provides a solid foundation for parallel workloads. The higher memory clock of 1000 MHz also benefits memory-bound compute tasks. The AMD part sits at the 40th percentile of all GPUs, which is slightly above the NVIDIA part's 38th percentile, indicating a marginally better overall standing in the database.

The NVIDIA GeForce GTX 670M has its own advantages, primarily in texture-heavy rendering scenarios. Its texture rate of 33.49 GTexel/s is 34% higher than the AMD part, driven by 56 TMUs. This gives it an edge in games or applications that are fill-rate limited by texture sampling. The wider 192-bit memory bus and higher bandwidth of 72.00 GB/s also provide an advantage in bandwidth-sensitive scenarios, even though the memory clock is lower. The NVIDIA part has 24 ROPs versus 16, which could help in certain pixel-heavy workloads despite the lower pixel rate. The 75 W TDP and MXM form factor make it a known quantity for system integrators designing mobile workstations or high-performance laptops.

For gaming, the choice depends on the bottleneck. If a game is shader-bound or compute-heavy, the AMD Radeon HD 8850M's 640 shading units and higher clocks will likely provide better performance. If a game is texture-bound, the NVIDIA GeForce GTX 670M's higher texture fill rate could prevail. The AMD part also has a notable feature advantage: Vulkan 1.2.170 support, which the NVIDIA part lacks entirely. This makes the AMD GPU more future-proof for modern titles that rely on Vulkan. Both support DirectX 12, though the AMD part supports the 11_1 feature level versus 11_0 on the NVIDIA part.

The AMD Radeon HD 8850M is the better choice for users prioritizing compute performance and modern API compatibility. The NVIDIA GeForce GTX 670M is the better choice for users who need maximum texture throughput or who operate within the constraints of an MXM-based system design. The data shows a single decisive benchmark win for AMD, but the architectural differences suggest that NVIDIA retains strengths in specific rendering paths. The 14.3% OpenCL delta is the headline number, but the texture rate and memory bandwidth advantages for NVIDIA should not be overlooked in application-specific evaluations.