AMD EPYC 9475F vs AMD EPYC 9684X Comparison

Head-to-Head Benchmarks

The head-to-head results tell a clear story of two opposing design philosophies. The AMD EPYC 9684X wins eight of the eleven shared benchmarks, often by substantial margins. The most lopsided victory belongs to the 9684X in data encryption, where it scores 178,453 against the 9475F's 116,648, a 34.6% advantage. Data compression also heavily favors the 9684X, with a score of 2,698,807 compared to 2,156,305, a 20.1 percent gap. Integer math shows a similar trend: 844,145 versus 605,696, putting the 9684X ahead by 28.2 percent. Physics performance is another strong point for the 9684X, delivering 24,686 against 16,443, a 33.4 percent lead.

The 9475F does not go down without a fight. Its most commanding victory is in single-threaded performance, where it scores 3,779 versus 2,891, a 30.7 percent advantage. This is the defining metric for the 9475F. It also edges out the 9684X in the multithread benchmark, scoring 122,476 against 121,595, though the 0.7 percent margin is narrow. The extended instructions test is closer still: the 9684X scores 177,956 against the 9475F's 173,169, a 2.7 percent gap. The 9684X also wins floating point math (472,174 vs 406,524, a 13.9 percent lead), prime number finding (2,020 versus 1,507, a 25.4 percent lead), and random string sorting (349,126 vs 253,936, a 27.3 percent lead).

The benchmark averages confirm the split. The 9475F's average benchmark score is 350,933, while the 9684X sits at 266,914. Interestingly, despite fewer wins, the 9475F has a higher average, indicating its losses are not as severe as its single-thread dominance might suggest. The 9475F also ranks in the 100th percentile among all CPUs, while the 9684X sits at the 99th. The nearest rivals for the 9475F include the AMD EPYC 9754 (3.7 percent lower average score), Intel Xeon 6960P (3.9 percent lower), and AMD EPYC 9655 (6 percent lower). The 9684X's rivals include the AMD Ryzen Threadripper 9970X (4.6 percent higher), Intel Xeon 6780E (4.8 percent higher), and Intel Xeon 6980P (6.1 percent lower).

Where Each One Wins

The use-case split is driven by core count and clock speed. The 9684X, with 96 cores and 192 threads, is built for parallel throughput. Its wins in data compression, encryption, integer math, floating point math, physics, and string sorting all point to workloads that scale with core count and cache capacity. Any database, analytics, or scientific computing task that can saturate many threads will see the 9684X pull ahead by the margins shown above. The 33.4 percent lead in physics is particularly telling for simulation and rendering workloads.

The 9475F, with 48 cores and 96 threads, wins where single-core speed matters most. Its 30.7 percent lead in single-thread performance makes it the choice for latency-sensitive applications, legacy software that relies on one or two threads, or mixed workloads where a few fast threads dominate. The multithread win, though small, suggests that the 9475F's high clock speed can compensate for half the core count in some synthetic loads. For tasks like sequential database queries, high-frequency trading, or real-time analytics where each thread matters, the 9475F's speed advantage is decisive.

The extended instructions test, where the 9684X wins by only 2.7 percent, indicates that modern instruction sets are the great equalizer. Both processors handle them well, and the difference here is negligible.

Architecture Differences

The two CPUs represent different generations of AMD server silicon. The 9475F is from the EPYC 9005 series, built on the Zen 5 architecture and codenamed Turin. It uses a 4 nm process from TSMC, with 66,520 million transistors across a die size of 8x 70.6 mm². The 9684X is from the EPYC 9004 series, featuring the Zen 4 architecture and codenamed Genoa-X. It relies on a 5 nm process, also from TSMC, with 135,240 million transistors spread across 12x 72 mm². The 9684X packs more than twice the transistors, though on a less advanced node.

Cache is where the 9684X makes its biggest architectural statement. It has 1,152 MB of shared L3 cache, compared to the 9475F's 256 MB. This four-to-one ratio in L3 cache is the defining feature of the 9684X. The 9475F counters with a larger L1 cache of 80 KB per core versus 64 KB per core. Both have 1 MB of L2 per core. The 9684X's massive L3 cache is the likely reason for its wins in data-heavy benchmarks like compression and encryption.

Clock speeds tell the opposite story. The 9475F has a base clock of 3.65 GHz and a boost clock of 4.80 GHz. The 9684X runs at 2.55 GHz base and 3.70 GHz boost. The 9475F's higher clocks drive its single-thread advantage. Both CPUs consume 400 watts TDP, use AMD Socket SP5, and support DDR5 memory across a twelve-channel bus. The 9475F has a memory bandwidth of 576.0 GB/s, while the 9684X achieves 460.8 GB/s. Both support ECC memory and use PCIe Gen 5 with 128 lanes. Neither has integrated graphics. The 9475F was released on October 9, 2024, while the 9684X came earlier on June 12, 2023.

FAQ

Q: Which CPU wins in single-threaded performance?

A: The AMD EPYC 9475F scores 3,779 in the single-thread test, beating the 9684X's 2,891, a 30.7 percent advantage.

Q: Does the EPYC 9684X have more cache?

A: Yes, the 9684X has 1,152 MB of shared L3 cache, while the 9475F has 256 MB. The 9475F has more L1 cache per core: 80 KB versus 64 KB.

Q: What are the core counts?

A: The 9475F has 48 cores and 96 threads. The 9684X has 96 cores and 192 threads.

Q: How do the clock speeds compare?

A: The 9475F runs at a base clock of 3.65 GHz and boost clock of 4.80 GHz. The 9684X runs at 2.55 GHz base and 3.70 GHz boost.

Q: Which CPU has higher memory bandwidth?

A: The 9475F has a memory bandwidth of 576.0 GB/s. The 9684X has 460.8 GB/s. Both use twelve-channel DDR5.

Q: What are the average benchmark scores?

A: The 9475F has an average benchmark score of 350,933, while the 9684X has 266,914.

Specification Differences

| Specification | AMD EPYC 9475F | AMD EPYC 9684X |

|----------------|----------------|----------------|

| Series | EPYC 9005 series | EPYC 9004 series |

| Architecture | Zen 5 | Zen 4 |

| Codename | Turin | Genoa-X |

| Process Node | 4 nm | 5 nm |

| Transistors | 66,520 million | 135,240 million |

| Die Size | 8x 70.6 mm² | 12x 72 mm² |

| Cores | 48 | 96 |

| Threads | 96 | 192 |

| Base Clock | 3.65 GHz | 2.55 GHz |

| Boost Clock | 4.80 GHz | 3.70 GHz |

| L1 Cache | 80 KB (per core) | 64 KB (per core) |

| L3 Cache | 256 MB (shared) | 1,152 MB (shared) |

| Memory Bandwidth | 576.0 GB/s | 460.8 GB/s |

| Release Date | 2024-10-09 | 2023-06-12 |

| Launch MSRP | $7592 | $14756 |

The 9684X's transistor count is more than double that of the 9475F, and its L3 cache is over four times larger. The 9475F counters with a newer node, higher clocks, and faster memory bandwidth.

The Verdict

The data suggests a clear performance split: the EPYC 9684X is the heavy lifter for parallel, cache-hungry workloads, while the EPYC 9475F is the speed champion for single-threaded and latency-critical tasks. The 9684X's wins in encryption, compression, integer math, and physics make it the obvious choice for large-scale server workloads, database processing, and scientific computing that can utilize all 96 cores and 192 threads. Its 1,152 MB L3 cache keeps massive datasets close to the cores, which explains its 28.2 percent lead in integer math and 33.4 percent lead in physics.

The 9475F, on the other hand, is the pick for environments where clock speed is king. Its 30.7 percent single-thread lead and 0.7 percent multithread win, despite half the cores, indicate that the Zen 5 architecture with its 4 nm process and 4.80 GHz boost clock delivers exceptional per-thread efficiency. It also offers higher memory bandwidth, 576.0 GB/s versus 460.8 GB/s, which helps in memory-bound single-thread tasks. The 9475F's launch MSRP of $7592 is lower than the 9684X's $14756, but decisions should rest on the benchmark deltas above.

The 9684X is the stronger overall performer across the majority of metrics. The 9475F is the stronger choice for legacy, single-threaded, or latency-sensitive applications. The data does not point to a universal winner; it points to a clear fork in the road.