AMD Radeon HD 8750M

The AMD Radeon HD 8750M is a 28 nm GPU from the Solar System (HD 8700M) generation, built by TSMC around the Mars chip and the GCN 1.0 architecture. The die contains 950 million transistors across 77 mm², for a transistor density of 12.3M / mm². Its compute array consists of 384 shading units, 24 texture mapping units, and 8 ROPs. The GPU runs at a 775 MHz base clock and 825 MHz boost clock, with memory at 900 MHz and 1800 Mbps effective. That memory is 1024 MB of DDR3 on a 128-bit bus, producing 28.80 GB/s of bandwidth. The resulting peak rates are 633.6 GFLOPS FP32, 19.80 GTexel/s texture fill, and 6.600 GPixel/s pixel fill. It was released on 2013-02-25, is now end-of-life, and its recorded predecessor and successor are London and Gem System. In the only recorded benchmark, Geekbench OpenCL, it scores 5946 and places in the 33rd percentile of all GPUs.

How It Compares

Against the NVIDIA Quadro K620M, the HD 8750M produces a 5946 Geekbench OpenCL score, while the K620M’s average score is 5957. The deltaPct is -0.2, meaning the HD 8750M is 0.2 percent below the K620M in the recorded metric. This is the smallest gap on the nearest-rival list, and it puts the two GPUs in effectively the same performance class.

Against the AMD Radeon HD 8730M, the HD 8750M has a deltaPct of -0.4. The HD 8730M’s average score is 5970, so the HD 8750M sits slightly behind a direct family member. The gap is still very small, and the two AMD parts are closer to each other than to the high end of the rival cluster.

Against the AMD FirePro W4100, the HD 8750M again records a deltaPct of -0.4. The FirePro W4100’s average score is 5972, which is only marginally above the HD 8730M. Despite being aimed at a different workstation segment, the FirePro W4100 is measured as nearly identical to the HD 8750M in this OpenCL result.

Against the NVIDIA Quadro K4000M, the HD 8750M has its largest deficit on the nearest-rival list: a deltaPct of -0.7. The K4000M’s average score is 5986. Even at the largest gap in this group, the difference is only 0.7 percent, so the HD 8750M is not separated from any listed rival by a meaningful score margin.

Ray Tracing and Feature Set

No dedicated ray tracing core count is recorded for the HD 8750M. No tensor core count is recorded either. As a result, ray tracing acceleration and tensor-based features cannot be characterized from the supplied data; these fields are absent from the specification.

The API support includes DirectX 12 (11_1), OpenGL 4.6, and Vulkan 1.2.170. The bus interface is PCIe 3.0 x8. The card uses 1024 MB of DDR3 memory over a 128-bit interface, with 28.80 GB/s of bandwidth. No FP16 performance figure is recorded, and no display output information is listed in the data.

The feature set is otherwise defined by the GCN 1.0 architecture: 384 shading units, 24 texture mapping units, and 8 ROPs are the fixed hardware resources available to applications. The absence of RT and tensor core data is significant because it leaves those acceleration paths unsupported by the specification record.

Benchmark Performance

The only benchmark score in the data is Geekbench OpenCL at 5946. Because that is the sole recorded test, the average benchmark score is also 5946. On the full GPU distribution, the HD 8750M sits at the 33rd percentile, meaning most tracked GPUs score higher in this same test.

Against the nearest rivals, the HD 8750M is at the bottom of a very dense cluster. The Quadro K620M averages 5957, the Radeon HD 8730M averages 5970, the FirePro W4100 averages 5972, and the Quadro K4000M averages 5986. The deltaPct values are -0.2, -0.4, -0.4, and -0.7 percent, respectively. The largest listed gap is only 0.7 percent, so the entire nearest-rival group behaves almost identically in the recorded OpenCL metric.

The compute result is consistent with the GPU’s peak specifications. The base and boost clocks of 775 MHz and 825 MHz feed 384 shading units, for 633.6 GFLOPS of FP32 throughput. The 24 TMUs and 8 ROPs yield 19.80 GTexel/s and 6.600 GPixel/s, respectively. The memory path is 128-bit DDR3 with 28.80 GB/s, and with a memory clock of 900 MHz / 1800 Mbps effective, the measured score reflects a part that is positioned in the lower third of the distribution but closely matched with its immediate competitors.

Who Should Consider It

Users should consider the HD 8750M when the target workload is adequately served by an OpenCL score of 5946. The 33rd-percentile ranking indicates that it is not a high-performance part in the broader GPU population. However, the nearest-rival comparison is equally important: the K620M, HD 8730M, FirePro W4100, and Quadro K4000M are all within 0.7 percent of the HD 8750M, so any of these products would deliver essentially the same measured compute performance.

The 1024 MB DDR3 frame buffer and 28.80 GB/s bandwidth place practical limits on memory-heavy work. Workloads that fit within that memory capacity and that tolerate this bandwidth level are reasonable targets. Workloads that need more resident data or higher sustained data movement would likely exceed the HD 8750M’s capabilities.

The FP32 rate of 633.6 GFLOPS gives modest compute headroom for lightly threaded or simple tasks, while the 6.600 GPixel/s pixel fill and 19.80 GTexel/s texture fill define the rasterization ceiling. No resolution-specific or game-specific benchmarks are recorded, so detailed settings recommendations cannot be grounded in the FACT PACK. The available grounds for selection are the OpenCL score, the 33rd-percentile rank, the 1024 MB memory configuration, and the end-of-life production status. A user choosing among the nearest rivals should expect a nearly identical compute result.

Power and Cooling

No TDP figure is recorded for the HD 8750M. No suggested PSU is recorded, and no power connector requirement is recorded. Without these fields, the data cannot quantify the card’s thermal load or power-supply demands.

The specification does record a 28 nm TSMC manufacturing process and a 950-million-transistor die, but those figures do not translate directly into a cooling recommendation. No slot width is recorded, and no physical dimensions are included, so chassis fit and cooler compatibility cannot be assessed from this data.

The production status is end-of-life, and the data leaves power and cooling fields empty. As a result, the thermal envelope, power connector topology, and power supply sizing are not available in the benchmark record; original platform documentation would be needed for those details.