AMD EPYC 9684X

The AMD EPYC 9684X is a 96-core, 192-thread server processor built on the Zen 4 architecture and the Genoa-X codename. It is manufactured on a 5 nm process by TSMC, with a base clock of 2.55 GHz and a boost clock of 3.70 GHz. The processor carries a 400 W TDP, uses the AMD Socket SP5, and supports DDR5 memory across a twelve-channel bus with a memory bandwidth of 460.8 GB/s. Its launch MSRP is $14756. The data shows a processor aimed squarely at high-density compute workloads, with benchmark results that place it in the 99th percentile of all CPUs.

Benchmark Performance

The EPYC 9684X delivers dominant multi-threaded performance, as evidenced by its Cinebench R23 multicore score of 103355. This is a massive figure, reflecting the full utilization of its 96 cores and 192 threads. In Cinebench R20 multicore, the score is 43409, and in the older Cinebench R15 multicore test, it reaches 10418. These results indicate that the processor scales exceptionally well under sustained all-core loads, which is typical for server-grade silicon but still impressive given the high core count.

The Geekbench multicore score of 15668 is notably lower than the Cinebench figures, but this is a reflection of the test’s shorter duration and different workload characteristics. PassMark’s multithread score of 121595 reinforces the multi-core strength, while the PassMark integer math score of 844145 and floating point math score of 472174 show balanced arithmetic capability. The data compression score of 2698807 and random string sorting score of 349126 highlight strong performance in data manipulation tasks, which are common in database and analytics workloads.

Compared to its nearest rivals, the EPYC 9684X holds a clear edge in average benchmark score. It is 9.3% ahead of the AMD EPYC 9634, which is a significant margin for a processor in this class. However, it trails the AMD EPYC 9565 by 6.5%, the AMD EPYC 9555P by 8.5%, and the AMD EPYC 9734 by 14.1% in average score. This means that while the 9684X is not the absolute fastest in its immediate peer group, it sits comfortably above the 9634 and within striking distance of the higher-scoring rivals. The average benchmark score of 266914 places it in a competitive position, with the data showing that it is a strong choice for workloads that benefit from its specific cache and core configuration.

How It Compares

AMD EPYC 9565: The 9565 has an average score of 285471, which is 6.5% higher than the 9684X’s average of 266914. This delta indicates that the 9565 edges out the 9684X in overall benchmark performance, likely due to architectural refinements or clock behavior. For users chasing every last point in synthetic tests, the 9565 is ahead, but the 9684X’s massive L3 cache may close the gap in real-world applications that are cache-sensitive.

AMD EPYC 9634: With an average score of 244274, the 9634 is 9.3% behind the 9684X. This is a clear victory for the 9684X, reinforcing its position as the stronger processor in this pairing. The 9634 is a capable chip, but benchmark results indicate that the 9684X offers a measurable performance advantage, making it the better pick for compute-heavy environments.

AMD EPYC 9555P: The 9555P posts an average score of 291664, which is 8.5% higher than the 9684X. This places the 9555P ahead in raw benchmark terms. The 9684X does not close this gap, so users prioritizing synthetic scores may favor the 9555P. However, the 9684X’s unique cache structure could provide benefits in specific workloads that the benchmarks do not fully capture.

AMD EPYC 9734: The 9734 leads this group with an average score of 310619, which is 14.1% ahead of the 9684X. This is the largest delta among the rivals, showing that the 9734 is a more powerful processor on average. The 9684X is not competitive with the 9734 in this metric, so the choice between them depends on whether the 9684X’s other characteristics, such as its L3 cache size, justify the lower benchmark standing.

Power and Thermals

The EPYC 9684X carries a TDP of 400 W, which places it in the high-power tier for server processors. This TDP class implies that it requires a robust cooling solution, typically a high-end air cooler or a liquid cooling system designed for server sockets. The data does not specify exact cooler dimensions or wattage dissipation, but the 400 W TDP signals that thermal management is a primary consideration for system integrators. A passive heatsink with strong airflow or an active liquid loop is necessary to maintain sustained performance, as the 96 cores will generate significant heat under full load. The 5 nm process node helps with efficiency, but the sheer core count and 400 W envelope mean that cooling is not an afterthought. In a dense server chassis, attention to intake and exhaust airflow is critical to avoid thermal throttling, which could reduce the benchmark scores seen here.

Who Should Consider It

The EPYC 9684X is designed for workloads that demand extreme multi-threading. Its Cinebench R23 multicore score of 103355 and PassMark multithread score of 121595 indicate that it excels in rendering, simulation, and scientific computing tasks that can utilize all 192 threads. Data centers running virtualized environments or containerized applications will benefit from the high core count, as will teams performing large-scale data analytics—the PassMark data compression score of 2698807 points to strong performance in compression-heavy pipelines. The single-thread scores are respectable, with Cinebench R23 single-core at 14591 and PassMark single-thread at 2891, but they are not the primary selling point. For gaming, this processor is overkill and not suited to that use case, given its server orientation and lack of integrated graphics. For office productivity, the multi-core power is wasted on typical workloads like word processing or spreadsheets, which are single-thread bound. The intended audience is clearly the server and workstation market, where the 96 cores and high memory bandwidth of 460.8 GB/s drive throughput in databases, engineering simulations, and machine learning training.

Platform and Compatibility

The EPYC 9684X uses the AMD Socket SP5, which is the platform for the Genoa and Genoa-X generation of EPYC processors. It supports DDR5 memory across a twelve-channel bus, with ECC memory enabled for data integrity in mission-critical environments. The memory bandwidth of 460.8 GB/s is substantial, allowing the processor to feed its many cores without starvation. PCIe Gen 5 support is included with 128 lanes available from the CPU, which provides ample connectivity for high-speed storage devices, network adapters, and GPU accelerators. This makes the platform suitable for systems that require extensive I/O, such as large-scale storage servers or AI training nodes. The upgrade path is tied to the SP5 socket, which also accommodates other Zen 4-based EPYC processors, so systems built around this socket can potentially move to different SKUs without a motherboard change. The production status is active, meaning the processor is currently available for new builds. The processor is not multiplier-unlocked, which is standard for server parts, so overclocking is not a consideration; performance is delivered through the stock configuration.

Single-Thread vs Multi-Thread Behavior

The EPYC 9684X shows a clear split between single-thread and multi-thread performance. In Cinebench R23, the single-core score is 14591, while the multicore score is 103355, which indicates a scaling factor of over seven times when moving from one core to all cores. This is expected for a 96-core processor, but the single-thread scores are not negligible. The Geekbench single-core score of 1586 and PassMark single-thread score of 2891 are respectable for a server chip, though they lag behind high-clock desktop parts. The base clock of 2.55 GHz and boost clock of 3.70 GHz are moderate, which limits single-thread performance compared to lower-core-count processors with higher clocks. In real workloads, this means that tasks like compilation, which involve both single-threaded parsing and multi-threaded compilation steps, will see strong overall performance but occasional bottlenecks in the single-threaded phases. Conversely, workloads like video encoding or 3D rendering, which are highly parallel, will saturate the cores and deliver near-peak throughput. The PassMark extended instructions score of 177956 and encryption score of 178453 show that the processor handles SIMD and cryptographic workloads efficiently, though these are also multi-threaded in nature. The single-thread scores are adequate for server tasks like database queries that are not fully parallelized, but the processor’s design philosophy is clearly about maximizing multi-thread throughput. The 1152 MB of shared L3 cache is a defining feature—this large cache helps mitigate the latency of accessing main memory, which is beneficial for multi-threaded workloads that share data across cores, but it does not boost single-thread performance beyond what the clock speeds allow.