AMD Radeon HD 8730M vs NVIDIA GeForce GTX 670M Comparison

The database pits two end-of-life mobile GPUs against each other: NVIDIA's GeForce GTX 670M, a 40 nm Fermi part from the GeForce 600M generation, and AMD's Radeon HD 8730M, a 28 nm GCN part from the HD 8700M family. Both are TSMC-built, both carry DirectX 12 support on paper, and both sit in the lower third of the all-GPU rankings. The recorded data shows a clear but modest NVIDIA edge in compute, while the underlying specifications reveal a far more interesting story about how each chip reaches its numbers.

FAQ

Q: Which GPU is faster in the recorded benchmarks?

A: The GTX 670M wins the only head-to-head test in the database, Geekbench OpenCL, with a score of 6513 against 5955 for the HD 8730M. That is a 9.4 percent advantage for the NVIDIA card.

Q: How does each GPU rank against all GPUs in the database?

A: The GTX 670M sits at the 38th percentile, the HD 8730M at the 34th percentile. Both are below the median, but the NVIDIA card is the stronger of the two in relative standing.

Q: Which GPU has the better memory subsystem?

A: The GTX 670M, by a wide margin. It pairs GDDR5 memory on a 192-bit bus for 72.00 GB/s of bandwidth, versus DDR3 on a 128-bit bus delivering 28.80 GB/s for the AMD part. The AMD card does carry more memory capacity, 2 GB against 1536 MB.

Q: Which GPU has the newer architecture and manufacturing process?

A: The HD 8730M. It uses GCN 1.0 on a 28 nm process, while the GTX 670M uses Fermi 2.0 on a 40 nm process. The AMD chip is also roughly a year newer, released in March 2013 versus March 2012.

Q: Are both GPUs still in production?

A: No. Both are flagged as end-of-life. The GTX 670M belongs to the GeForce 600M line between the GeForce 500M predecessor and the GeForce 700M successor, while the HD 8730M's predecessor is listed as London and its successor as Gem System.

Q: Which GPU has the higher theoretical compute throughput?

A: The GTX 670M, at 803.7 GFLOPS FP32 versus 537.6 GFLOPS FP32 for the HD 8730M.

The Verdict

The data points to the GTX 670M for anyone choosing strictly on performance. It wins the only recorded benchmark, leads on every throughput metric in the database (FP32, pixel rate, texture rate), and more than doubles the memory bandwidth of its rival. Its 38th percentile ranking versus the AMD card's 34th confirms the same ordering against the wider GPU landscape.

The HD 8730M's case rests entirely on efficiency and platform features, not speed. Its GCN 1.0 architecture on 28 nm packs 950 million transistors into a 77 mm² die at 12.3M transistors per mm², roughly double the density of the GTX 670M's 1,950 million transistors spread across 332 mm² at 5.9M per mm². It also supports Vulkan 1.2.170 and DirectX 12 at the 11_1 feature level, one tier above the NVIDIA card's 11_0, and offers more video memory. For a modern API-dependent workload that favors GCN, or where the smaller die and the absence of the NVIDIA card's 75 W TDP matter, the AMD chip is the more forward-looking design. But in raw measured speed, the GTX 670M takes it.

Head-to-Head Benchmarks

There is a single head-to-head result in the database, and it goes to NVIDIA. In Geekbench OpenCL, the GTX 670M scores 6513 and the HD 8730M scores 5955, a 9.4 percent win for the GeForce. That gap is real but not enormous: it is the kind of margin that shows up as a settings notch or two in older titles rather than a generational difference.

Context from each card's nearest rivals sharpens the picture. The GTX 670M's score of 6513 puts it within a hair of the GT 555M at 6493 (0.3 percent ahead), the Quadro M5000M at 6481 (0.5 percent ahead), and the Radeon Vega 10 Mobile at 6476 (0.6 percent ahead), while the Intel UHD Graphics P750 actually edges it at 6554 (0.6 percent ahead). The HD 8730M's 5955 is essentially tied with the Quadro K620M at 5957, within touching distance of the HD 8750M at 5970 (0.3 percent behind), the UHD Graphics 730 at 5929 (0.4 percent ahead), and the Quadro K4000 at 5982 (0.5 percent behind). In other words, both cards compete in tightly packed clusters, and the NVIDIA card's cluster sits measurably higher.

The theoretical metrics reinforce the OpenCL result. The GTX 670M delivers 803.7 GFLOPS FP32 against 537.6, a roughly one-third advantage in raw shader compute. Pixel throughput is 8.372 GPixel/s against 5.600 GPixel/s, and texture throughput is 33.49 GTexel/s against 16.80 GTexel/s, exactly double in the fill-rate comparison. Bandwidth is the biggest gap of all: 72.00 GB/s versus 28.80 GB/s.

Specification Differences

The two cards differ on nearly every measurable axis. The GTX 670M was released on 2012-03-21, the HD 8730M on 2013-03-31, so the AMD part is about a year newer. The NVIDIA chip (GF114) is built on 40 nm with 1,950 million transistors and a 332 mm² die; the AMD chip (Mars) is 28 nm with 950 million transistors and a 77 mm² die.

Memory is split in both directions. The GTX 670M has 1536 MB of GDDR5 on a 192-bit bus at 3 Gbps effective, yielding 72.00 GB/s. The HD 8730M has 2 GB of DDR3 on a 128-bit bus at 1800 Mbps effective, yielding 28.80 GB/s. NVIDIA wins speed; AMD wins capacity.

Shader resources also split. The HD 8730M actually has more shading units, 384 versus 336, but the GTX 670M has far more TMUs (56 versus 24) and ROPs (24 versus 8), which is why its fill rates and compute output dominate. Clocks tell the same story of different philosophies: the HD 8730M runs a defined 650 MHz base and 700 MHz boost, while the GTX 670M lists no core clock in the database, only its memory clock.

The interfaces differ too. The GTX 670M is an MXM-B (3.0) module with portable-device-dependent display outputs and a 75 W TDP; the HD 8730M connects over PCIe 3.0 x8 with no TDP recorded. On APIs, both support DirectX 12 and OpenGL 4.6, but the AMD card adds Vulkan 1.2.170 and the higher 11_1 feature level versus 11_0.

Architecture Differences

These are two fundamentally different design philosophies separated by a process-node jump. The GTX 670M is Fermi 2.0 on TSMC's 40 nm node, carrying 1,950 million transistors across a large 332 mm² die at 5.9M transistors per mm². The HD 8730M is GCN 1.0 on TSMC's 28 nm node, with 950 million transistors in a 77 mm² die at 12.3M transistors per mm². The AMD chip is less than a quarter of the die area while delivering roughly two-thirds of the compute output, a direct illustration of what the node transition bought.

Fermi 2.0 and GCN 1.0 also represent opposite API strategies. GCN was built with explicit low-level APIs in mind, and the database reflects that: the HD 8730M lists Vulkan 1.2.170 support and DirectX 12 at feature level 11_1. The GTX 670M tops out at feature level 11_0 within DirectX 12 and has no Vulkan entry at all. Neither card has RT cores or tensor cores; hardware ray tracing and AI acceleration simply were not features of either era.

The resource mix differs notably. GCN 1.0 gives the AMD card 384 shading units on a narrow backend of 24 TMUs and 8 ROPs. Fermi's configuration in the GTX 670M is the inverse pattern: fewer shaders (336) but a heavy 56-TMU, 24-ROP back end fed by GDDR5. That structure explains why the NVIDIA card doubles texture rate and triples pixel rate despite the AMD card's shader-count advantage.

Where Each One Wins

The GTX 670M wins every performance-oriented category in the database. It takes the sole head-to-head benchmark by 9.4 percent, leads FP32 compute 803.7 to 537.6 GFLOPS, doubles texture rate at 33.49 versus 16.80 GTexel/s, and wins pixel rate 8.372 to 5.600 GPixel/s. Its 72.00 GB/s of GDDR5 bandwidth makes it the clear pick for bandwidth-sensitive workloads and higher resolutions or detail settings, where the AMD card's DDR3 would be the first bottleneck. Anyone choosing strictly on the recorded numbers takes the GeForce.

The HD 8730M wins on modern platform support and efficiency of design. Its Vulkan 1.2.170 support and higher DirectX feature level give it API headroom the Fermi card lacks. Its 28 nm, 77 mm² die delivers its performance from a fraction of the silicon, and with no recorded TDP versus the GTX 670M's 75 W, it is the architecturally leaner part. The extra 2 GB of memory, while slow, can matter for capacity-bound tasks that fit neither card's bandwidth profile well. It also sits in a tight cluster with the HD 8750M and the Quadro K620M, scoring essentially identically, so it holds its ground within its own class. The split is straightforward: measured speed goes to NVIDIA, efficiency and API modernity go to AMD.