NVIDIA A10M

NVIDIA A10M is a server-oriented Ampere GPU built on the GA102 chip, fabricated on Samsung's 8 nm process with 28,300 million transistors on a 628 mm² die. It sits in the 97th percentile of all GPUs in the benchmark database, achieving a Geekbench OpenCL score of 135230. The card is now end-of-life, but its benchmark position remains relevant for comparison against contemporary workstation and high-end consumer parts.

Power and Cooling

The A10M carries a 150 W TDP, a remarkably low figure given its transistor count and compute capabilities. This power envelope is achieved through the 8 nm process and a conservative clock profile: a base of 975 MHz and a boost of 1635 MHz. The board is designed as a single-slot solution, measuring 267 mm in length and 112 mm in height, which allows for dense server installations. Power is delivered through a single 8-pin EPS connector, and NVIDIA recommends a 450 W power supply for systems hosting this card. Because the card has no display outputs, all power is directed toward compute workloads rather than video output. The low TDP also simplifies cooling in multi-GPU chassis, though the single-slot form factor limits the size of the heatsink and fan assembly. The combination of a modest 150 W envelope and a single-slot design makes the A10M suitable for rack-mounted servers where space and thermal headroom are constrained.

Memory Subsystem

The A10M is equipped with 20 GB of GDDR6 memory on a 320-bit bus, yielding a bandwidth of 500.2 GB/s. The memory operates at 1563 MHz, which translates to 12.5 Gbps effective data rate. This capacity and bandwidth are well matched to server workloads such as AI inference, rendering, and data analytics, where large datasets must reside on the GPU. At high resolutions, the 20 GB frame buffer reduces the need to spill to system memory, while the 500.2 GB/s bandwidth supports high texture throughput and large compute shader workloads. The pixel rate of 130.8 GPixel/s and texture rate of 366.2 GTexel/s further indicate the card’s ability to handle high-resolution rasterization tasks, though its primary role is compute. The 320-bit bus width is narrower than some high-end consumer parts, but the combination of capacity and bandwidth is balanced for the intended server market.

Ray Tracing and Feature Set

The A10M integrates 56 RT cores and 224 tensor cores, reflecting its Ampere architecture’s focus on ray tracing and AI acceleration. The RT cores enable hardware-accelerated ray tracing, while the tensor cores provide substantial throughput for deep learning and neural network operations. The card supports DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4, covering the major graphics and compute APIs used in professional and server environments. The FP32 and FP16 compute rates are both 23.44 TFLOPS, with a 1:1 ratio, indicating that the card does not rely on rate-boosted FP16 paths but rather delivers consistent performance across precisions. This makes the A10M a versatile accelerator for mixed-precision workloads. The presence of 224 tensor cores further enhances its suitability for inference tasks, though the card lacks dedicated video outputs, reinforcing its compute-only role.

How It Compares

The A10M’s closest rival in the benchmark database is the NVIDIA RTX 4000 Ada Generation, which scores 135218, a delta of 0% from the A10M’s 135230. This effectively places the two cards in a statistical tie, with the A10M holding a negligible 12-point lead. Both cards target similar professional compute segments, and the performance parity suggests that users can expect nearly identical OpenCL results from these two generations.

Next is the AMD Radeon RX 9070 GRE, with an average score of 134417, placing it 0.6% behind the A10M. This is a small but measurable gap, indicating that the A10M holds a slight edge over this AMD part in the Geekbench OpenCL workload. The RX 9070 GRE is a consumer-oriented GPU, yet it comes close to the A10M’s server-grade performance, which speaks to the A10M’s efficiency.

The AMD Radeon PRO W6800 scores 133588, which is 1.2% lower than the A10M. This professional workstation card trails the A10M by a clear margin, though the difference is still modest. The A10M’s higher score may be attributed to its larger memory bus and higher shading unit count, though the benchmark result alone does not reveal the cause.

Finally, the NVIDIA GeForce RTX 3090 Ti, a high-end consumer card, averages 131911, which is 2.5% behind the A10M. This is the largest gap among the nearest rivals, showing that the A10M outperforms even a flagship gaming GPU in this specific OpenCL test. The RTX 3090 Ti has a higher TDP and larger memory bus, but the A10M’s server-tuned drivers and compute-oriented design likely contribute to its advantage.

Benchmark Performance

The A10M’s Geekbench OpenCL score of 135230 places it in the 97th percentile of all GPUs in the database, indicating that it outperforms the vast majority of tested graphics cards. The score is derived from a single benchmark run, with an average benchmark score identical to the reported figure, suggesting consistency across runs. When compared to its nearest rivals, the A10M leads by margins ranging from 0% to 2.5%. The tight clustering of scores—135230, 135218, 134417, 133588, and 131911—shows that the A10M sits at the top of a performance tier where differences are often within a few percentage points.

The 0% delta against the RTX 4000 Ada Generation means the two cards are functionally equivalent in this workload. The 0.6% lead over the RX 9070 GRE is marginal and unlikely to be perceptible in real-world applications, but it does give the A10M a statistical advantage. Against the PRO W6800, the 1.2% gap is slightly more pronounced, and against the RTX 3090 Ti, the 2.5% difference is the most significant. These results indicate that the A10M, despite being an end-of-life server part, still holds its own against newer and more power-hungry GPUs in compute-heavy tasks.

The benchmark performance aligns with the card’s specifications: 7168 shading units, 224 TMUs, and 80 ROPs, combined with a 500.2 GB/s memory bandwidth, deliver strong raw compute throughput. The FP32 performance of 23.44 TFLOPS is consistent with the observed OpenCL score, as OpenCL workloads often rely on FP32 arithmetic. The 1:1 FP16 ratio further ensures that mixed-precision tasks do not suffer a penalty, making the A10M a balanced choice for server deployments where diverse workloads are common. Overall, the A10M’s benchmark standing reflects its position as a capable, efficient compute accelerator, albeit one that is now succeeded by newer architectures.