AMD EPYC 9384X

The AMD EPYC 9384X is a 32-core, 64-thread server processor built on the Zen 4 architecture with the Genoa-X codename. It operates at a base clock of 3.10 GHz and a boost clock of 3.90 GHz, with a 320 W TDP. The chip is manufactured on a 5 nm process at TSMC, containing 90,160 million transistors across eight 72 mm² dies. It features a substantial L3 cache of 768 MB shared across all cores, supplementing 64 KB of L1 and 1 MB of L2 per core. Memory support includes DDR5 with a twelve-channel bus delivering 460.8 GB/s of bandwidth, along with ECC capability. The processor provides Gen 5 PCIe with 128 CPU-only lanes and fits the AMD Socket SP5. Its average benchmark score of 120,427 places it in the 98th percentile of all CPUs, indicating top-tier performance within the database.

How It Compares

Against the AMD EPYC 7642, the 9384X holds a narrow edge. The rival's average score is 124,006, which translates to a delta of -2.9% relative to the 9384X. This means the 9384X outperforms the EPYC 7642 by approximately 2.9% in aggregate benchmarks, a modest but consistent lead that favors the newer architecture and the massive L3 cache.

The AMD Ryzen Threadripper PRO 5975WX presents a similar challenge, with an average score of 124,171 and a delta of -3%. The 9384X is 3% ahead of this workstation-oriented rival, a gap that is slightly larger than against the EPYC 7642 but still within the same competitive band. The Threadripper's single-thread strength is notable, but the 9384X's server-focused design and cache hierarchy give it the overall advantage.

The AMD EPYC 9354, another Genoa part, scores 126,810 on average, a delta of -5% compared to the 9384X. This 5% deficit for the rival shows the 9384X's advantage is not merely architectural but also tied to its specific cache configuration, as both share the Zen 4 core design. The 9384X's 768 MB of L3 clearly contributes to its superior aggregate performance.

The most significant gap appears against the AMD EPYC 9275F, which averages 133,174, a delta of -9.6%. Here, the 9384X trails by 9.6%, indicating that the 9275F holds a distinct performance lead. This rival likely benefits from higher clock speeds or core efficiency, though the data shows the 9384X remains competitive in specific workloads despite this overall deficit.

Power and Thermals

The 9384X carries a TDP of 320 W, classifying it as a high-power server processor. This TDP figure implies that a robust cooling solution is necessary, likely a high-end air cooler or liquid cooling system designed for socket SP5. The thermal design point is typical for flagship EPYC parts, where sustained multi-core loads generate substantial heat. Given the 5 nm process and the dense 768 MB L3 cache, the power envelope is directed toward both compute and cache maintenance. The 320 W TDP does not indicate overclocking headroom, as the multiplier is locked, so cooling requirements are strictly for stock operation. For server deployments, this means chassis airflow and heatsink capacity must meet the demands of 32 cores running at full load. The data does not provide specific thermal figures, but the TDP alone suggests a cooling tier that prioritizes high static pressure and large surface area.

Single-Thread vs Multi-Thread Behavior

The 9384X exhibits a clear split between single-thread and multi-thread performance, as shown by the Cinebench scores. In Cinebench R23, the single-core score is 8,359, while the multi-core score reaches 59,215. This represents a multi-core to single-core ratio of roughly 7.1, which is consistent with a 32-core processor where scaling is strong but not perfectly linear due to thermal and memory contention. The single-thread score of 8,359 is competitive for a server chip, indicating that lightly threaded tasks like database queries or latency-sensitive operations will perform well, though not at the level of top desktop parts. Meanwhile, the multi-core score of 59,215 shows that heavily threaded workloads, such as scientific simulations or video encoding, can utilize nearly all cores effectively.

In Cinebench R20, the single-core score is 3,510 and multi-core is 24,870, yielding a similar ratio. The R15 results, with 842 single-core and 5,968 multi-core, follow the same pattern. This consistency across versions suggests the architecture scales predictably. PassMark data reinforces this: single-thread score is 3,015, while multithread score is 69,665, a ratio of 23.1. The much higher ratio in PassMark reflects its specific test composition, where some multi-threaded subtests (like integer math and floating-point math) scale almost perfectly across 64 threads. For real workloads, this means the 9384X excels in parallel tasks that can saturate all cores, while maintaining respectable responsiveness in serial tasks. The data implies that users should match workloads to the core count to maximize throughput.

FAQ

Q: What is the average benchmark score of the AMD EPYC 9384X?

A: The average benchmark score is 120,427, placing it in the 98th percentile of all CPUs in the database.

Q: How does the 9384X compare to the AMD EPYC 9354 in terms of average score?

A: The EPYC 9354 has an average score of 126,810, which is 5% lower than the 9384X's score, meaning the 9384X leads by 5%.

Q: What is the total L3 cache size on the 9384X?

A: The processor has 768 MB of shared L3 cache, which is a defining feature of the Genoa-X design.

Q: What memory bandwidth does the 9384X support?

A: It supports DDR5 memory with a twelve-channel bus, providing a memory bandwidth of 460.8 GB/s.

Q: Does the 9384X have an unlocked multiplier for overclocking?

A: No, the multiplier is locked, so the processor runs at its specified base and boost clocks without manual overclocking.

Q: How many PCIe lanes does the 9384X provide?

A: It offers 128 PCIe Gen 5 lanes, but these are CPU-only, meaning dedicated to the processor rather than shared with chipset functions.

Benchmark Performance

The Cinebench R23 multi-core score of 59,215 is a strong indicator of parallel throughput. When compared to the nearest rivals, the 9384X's position is nuanced. The EPYC 7642, with a delta of -2.9%, scores roughly 2.9% lower on average, but the R23 multi-core result suggests the 9384X's advantage may be larger in cache-sensitive workloads. The Threadripper PRO 5975WX, at -3%, trails similarly, but its lower core count likely means the 9384X wins decisively in multi-threaded tasks. The EPYC 9354, at -5%, shows a more substantial gap, despite being a 32-core part. This indicates the 768 MB L3 cache provides a measurable benefit in Cinebench's rendering engine, which often uses large datasets.

The single-core Cinebench R23 score of 8,359 is also telling. While the 9384X trails the EPYC 9275F by 9.6% in average score, the single-core result is competitive, suggesting the deficit arises from multi-thread efficiency rather than raw single-thread speed. The PassMark multithread score of 69,665 further emphasizes this, as it represents a 23.1x improvement over the single-thread score of 3,015. This scalability is exceptional, though the PassMark find prime numbers test scores only 596, indicating that integer-heavy prime enumeration does not scale as well, possibly due to memory latency. The floating-point math score of 174,630 and integer math score of 297,833 both contribute significantly to the multithread total, while data encryption at 72,631 and data compression at 1,119,983 show strong performance in I/O and security workloads.

Compared to the EPYC 9275F's 9.6% lead in average score, the 9384X's R23 multi-core score of 59,215 suggests it closes the gap in certain applications. The exact deltaPct values from the rivals indicate that the 9384X is 2.9% faster than the EPYC 7642, 3% faster than the Threadripper PRO 5975WX, and 5% faster than the EPYC 9354. These percentages are based on average scores across all benchmarks, so individual tests may vary. The 9384X's strength lies in its massive cache, which boosts performance in workloads that fit within 768 MB, while its 32 cores provide a solid baseline for scaling. The data shows a processor that is consistently ahead of most peers but not the absolute leader, as the 9275F demonstrates.

Who Should Consider It

The 9384X is suited for workloads that benefit from both high core counts and large cache capacity. For gaming, this processor is not typical, but if used in a gaming server or a high-end workstation, the single-thread score of 3,015 in PassMark and 8,359 in Cinebench R23 ensures adequate performance for game logic and physics, though the locked multiplier limits enthusiast tuning. The multi-thread scores, particularly the Cinebench R23 result of 59,215, make it a candidate for content creation tasks like 3D rendering and video editing, where every core contributes to reducing render times. The PassMark data compression score of 1,119,983 indicates strong performance in file archiving and database compression, making it ideal for data-heavy server roles.

For office and enterprise applications, the 9384X's 64 threads and 768 MB L3 cache handle virtualization and large-scale data processing efficiently. The twelve-channel memory bus with 460.8 GB/s bandwidth ensures that memory-bound tasks, such as large in-memory databases, have ample throughput. The 98th percentile ranking underscores its top-tier status, but the 9.6% deficit to the EPYC 9275F suggests that workloads requiring the absolute highest multi-thread performance might consider that rival. However, the 9384X's cache advantage could tip the balance for specific datasets. Ultimately, the data supports recommending this processor for server and workstation environments where parallel compute, large cache residency, and high memory bandwidth are priorities, but not for users needing the fastest single-thread response or who rely on overclocking.