AMD Radeon HD 8690M vs NVIDIA GeForce GTX 670MX Comparison

The AMD Radeon HD 8690M and the NVIDIA GeForce GTX 670MX are two end-of-life mobile GPUs that land in nearly the same place in the benchmark database despite being built on radically different silicon. Both were fabbed at TSMC on a 28 nm process, and both sit in the lower third of the historical GPU field, with the HD 8690M at the 36th percentile against all GPUs in the database and the GTX 670MX at the 33rd. Yet the recorded data shows they arrive at that position through completely different means: one is a small, efficiency-focused chip, the other a cut-down version of a much larger desktop-class design. The only direct head-to-head measurement available is the Geekbench OpenCL test, and there the margin between them is almost imperceptible.

Head-to-Head Benchmarks

The database contains a single direct comparison point between these two GPUs, and it is extraordinarily close.

• Geekbench OpenCL: AMD Radeon HD 8690M scores 6,137 points; NVIDIA GeForce GTX 670MX scores 6,125 points. The winner is the HD 8690M, by a delta of 0.2 percent.

That 12-point gap is functionally a tie in compute workloads. A 0.2 percent difference is well within the noise band of a benchmark like OpenCL, and neither GPU can claim a meaningful advantage there. What makes the result interesting is the hardware behind it. The GTX 670MX delivers 1,153.9 GFLOPS of FP32 compute against the HD 8690M's 624.0 GFLOPS, meaning NVIDIA's part has nearly double the theoretical compute throughput, yet it converts that advantage into essentially nothing in the recorded OpenCL run. The HD 8690M's higher clocks (925 MHz base, 975 MHz boost, versus a flat 601 MHz base and boost on the GTX 670MX) and its GCN 1.0 architecture evidently extract far more real-world compute from far less hardware.

Context from each card's rival list reinforces how tightly packed this performance tier is. The HD 8690M sits within 1.2 percent of the AMD Radeon HD 6950 (6,210 average score, delta of -1.2 percent), one percent of the NVIDIA RTX A400 (6,078) and NVIDIA GeForce MX230 (6,077), and just 0.3 percent from the Intel Iris Pro Graphics 6200 (6,117). The GTX 670MX's neighborhood is equally dense: it is 0.2 percent from the Intel Iris Pro Graphics P6300 (5,712), 0.2 percent from the NVIDIA GeForce GTX 550 Ti (5,731), 0.5 percent from the AMD Radeon HD 8790M (5,691), and 2.1 percent from the NVIDIA Quadro M500M (5,604). These are GPUs separated by rounding errors, not generations.

The GTX 670MX also has a recorded Geekbench Vulkan score of 5,316, a test for which no HD 8690M result exists in the database, so no cross-comparison can be drawn from it.

FAQ

Q: Which GPU wins the only direct benchmark between them?

A: The AMD Radeon HD 8690M wins Geekbench OpenCL with 6,137 points to the GTX 670MX's 6,125, a margin of just 0.2 percent. It is the only head-to-head test recorded in the database, and the AMD card holds one win to NVIDIA's zero.

Q: Do they support the same graphics APIs?

A: Almost. Both report DirectX 12 support and OpenGL 4.6. The HD 8690M lists DirectX 12 (11_1) feature level and Vulkan 1.2.170, while the GTX 670MX lists DirectX 12 (11_0) and Vulkan 1.2.175. The AMD part therefore claims a slightly newer DirectX feature level, while the NVIDIA part claims a slightly newer Vulkan version.

Q: Which has more memory?

A: The GTX 670MX, with 3 GB of GDDR5 on a 192-bit bus delivering 67.20 GB/s of bandwidth. The HD 8690M has 2 GB of GDDR5 on a 64-bit bus with 32.00 GB/s of bandwidth, less than half the NVIDIA card's throughput.

Q: Which GPU is physically smaller?

A: The HD 8690M, by a wide margin. Its die measures 56 mm² with 690 million transistors, against 294 mm² and 3,540 million transistors for the GTX 670MX's GK104 chip.

Q: How much power does each draw?

A: The GTX 670MX has a recorded TDP of 75 W and requires no power connectors. The database records no TDP for the HD 8690M, so no power comparison can be made from the available data.

Q: When were they released, and are they still in production?

A: The GTX 670MX launched on September 30, 2012, and the HD 8690M followed on February 28, 2013. Both are listed as end-of-life.

Where Each One Wins

The case for the HD 8690M rests on measured efficiency rather than raw capability. It matches a GPU with triple its shading units (320 versus 960), four times its TMUs (20 versus 80), and three times its ROPs (8 versus 24) in the only benchmark the two share. It also edges the GTX 670MX in percentile ranking, 36 versus 33, and in average benchmark score, 6,137 versus 5,721. Its texture rate of 19.50 GTexel/s and pixel rate of 7.800 GPixel/s are lower on paper, but the OpenCL data shows those theoretical rates do not translate into a compute deficit. For any workload resembling the recorded test, general-purpose GPU compute being the clearest example, the two are interchangeable in the data.

The case for the GTX 670MX is a paper-specs case. Its 1,153.9 GFLOPS of FP32 compute, its 48.08 GTexel/s texture rate, and its 12.02 GPixel/s pixel rate all more than double the HD 8690M's figures. Its memory subsystem, 3 GB across a 192-bit bus at 67.20 GB/s, offers substantially more capacity and bandwidth than the HD 8690M's 2 GB, 64-bit, 32.00 GB/s configuration. The database's single shared benchmark cannot confirm whether those advantages materialize in graphics workloads, since the head-to-head contains only an OpenCL result, but for texture-heavy or memory-bound scenarios the specification gap is the only evidence available, and it all points one way.

Specification Differences

The two GPUs differ on nearly every recorded specification:

• Chip and architecture: Sun chip with GCN 1.0 architecture (HD 8690M) versus GK104 chip with Kepler architecture (GTX 670MX).
• Generation: Solar System (HD 8600M) versus GeForce 600M.
• Transistors: 690 million versus 3,540 million.
• Die size: 56 mm² versus 294 mm², though transistor density is nearly identical at 12.3M per mm² versus 12.0M per mm².
• Clocks: 925 MHz base and 975 MHz boost on the HD 8690M, versus 601 MHz base and boost on the GTX 670MX. Memory clocks are 1000 MHz (4 Gbps effective) versus 700 MHz (2.8 Gbps effective).
• Memory: 2 GB on a 64-bit bus at 32.00 GB/s versus 3 GB on a 192-bit bus at 67.20 GB/s.
• Shader resources: 320 shading units, 20 TMUs, 8 ROPs versus 960 shading units, 80 TMUs, 24 ROPs.
• Throughput: 7.800 GPixel/s and 19.50 GTexel/s versus 12.02 GPixel/s and 48.08 GTexel/s.
• Compute: 624.0 GFLOPS FP32 versus 1,153.9 GFLOPS FP32.
• Bus interface: PCIe 3.0 x8 versus PCIe 3.0 x16.
• TDP and connectors: null versus 75 W with no power connectors.
• Display outputs: not recorded versus "Portable Device Dependent".
• APIs: DirectX 12 (11_1) and Vulkan 1.2.170 versus DirectX 12 (11_0) and Vulkan 1.2.175.
• Lineage: predecessor London and successor Gem System versus predecessor GeForce 500M and successor GeForce 700M.

Both use TSMC 28 nm fabrication, GDDR5 memory, OpenGL 4.6, and both are end-of-life. Neither has RT cores or tensor cores recorded.

Architecture Differences

Architecturally these are two philosophies separated by a common process node. The HD 8690M is a Sun chip from AMD's GCN 1.0 generation, part of the Solar System family that the HD 8600M line belonged to. GCN 1.0 was AMD's compute-oriented design, and the database results bear that out: a 56 mm² die with 690 million transistors produces an OpenCL score that matches a chip five times its area.

The GTX 670MX is a GK104 part, meaning it shares its silicon lineage with NVIDIA's Kepler flagship design of that era, cut down for the GeForce 600M mobile generation. Its 294 mm² die carries 3,540 million transistors, and the near-identical transistor densities of the two chips (12.0M versus 12.3M per mm²) confirm that the size difference reflects scope, not fabrication quality. Kepler runs cooler-clocked and wider: 960 shading units at 601 MHz against AMD's 320 units at up to 975 MHz. NVIDIA's wider, slower approach yields the higher theoretical compute figure, while AMD's narrower, faster approach wins the one measured test they share.

The interface difference also matters at the platform level. The HD 8690M connects over PCIe 3.0 x8, half the lanes of the GTX 670MX's PCIe 3.0 x16, a constraint consistent with a chip positioned lower in its family.

The Verdict

The data supports a simple split. If the workload resembles the recorded OpenCL test, general-purpose GPU compute, there is nothing to choose between them: 6,137 versus 6,125 is a 0.2 percent margin, and the HD 8690M technically holds the win. It achieves that result from a die one-fifth the size and with roughly half the theoretical FP32 throughput, which makes it the efficiency story of this pairing even though no TDP figure is recorded for it.

If the workload leans on graphics resources, textures, pixels, and memory bandwidth, the recorded specifications favor the GTX 670MX across the board: triple the shading units, quadruple the TMUs, triple the ROPs, double the pixel rate, more than double the texture rate, and double the memory bandwidth with a larger 3 GB frame buffer. The database cannot verify those advantages in a graphics benchmark, since no such head-to-head exists, but every spec-level signal points the same way.

Both are end-of-life parts separated by roughly five months of launch timing, and both sit within a few points of integrated graphics like Intel's Iris Pro 6200 and P6300 in their respective tiers. For a modern buyer, neither is competitive; for the record, the HD 8690M wins the only measured contest, and the GTX 670MX wins every contest the database has not yet measured.