AMD EPYC 9684X vs AMD EPYC 9745 Comparison

Head-to-Head Benchmarks

The benchmark data reveals a clear overall victor, but the path to that result is more nuanced than a simple tally of wins. The AMD EPYC 9745 secures 13 of the 17 recorded head-to-head comparisons, while the AMD EPYC 9684X claims 4. However, the magnitude of those victories tells a compelling story about each processor's architectural priorities.

The EPYC 9745 dominates the Cinebench suite with remarkable consistency. Across all six Cinebench tests, from R15 to R23, it posts a uniform 7.5% advantage. This consistency is striking because it spans both single-core and multi-core workloads. The R23 multi-core score of 111093 against 103355, and the R23 single-core score of 15683 against 14591, both show the exact same 7.5% delta. This suggests a fundamental per-thread efficiency advantage rather than a raw core-count effect, since the 128-core 9745 would be expected to show a larger gap in heavily threaded tests if core scaling were the dominant factor.

The Passmark suite reveals where the 9745 truly separates itself. Data compression shows a massive 45.6% lead, with scores of 3929890 versus 2698807. Extended instructions amplify this further, with a 57.6% gap (280477 versus 177956). Floating point math reaches the largest disparity at 61.2%, scoring 761219 against 472174. Integer math follows closely at 45% ahead, posting 1224315 versus 844145. Data encryption shows a 28.6% advantage, and random string sorting lands at 34.3% in favor of the 9745. These are not marginal differences; they represent substantial throughput advantages in compute-heavy server workloads.

The EPYC 9684X, despite losing the overall contest, posts some significant wins of its own. The most dramatic is in prime number finding, where it scores 2020 against the 9745's 979, a 51.5% advantage. Physics calculations show a 30.6% lead, with 24686 against 17122. Single-thread performance in Passmark also favors the 9684X by a slim 2.9%, scoring 2891 versus 2806. These wins point to scenarios where the 9684X's architecture excels at specific mathematical operations, even if it falls behind in aggregate throughput.

The average benchmark score differential reinforces the 9745's overall position. The 9745 averages 425973 across all recorded benchmarks, while the 9684X averages 266914. This places the 9745 at the 100th percentile among all CPUs in the database, with the 9684X at the 99th. The nearest rivals for the 9745 include the AMD Ryzen Threadripper PRO 9995WX at 3.4% behind, the AMD EPYC 9655P at 7.1% behind, and the AMD EPYC 9535 at 12.3% behind. The 9684X, by contrast, sits behind the AMD Ryzen Threadripper 9970X by 4.6%, behind the Intel Xeon 6780E by 4.8%, and behind the AMD EPYC 9565 by 6.5%, while leading the Intel Xeon 6980P by 6.1%.

The multi-thread Passmark score shows a narrower 7.5% gap, matching the Cinebench pattern. This consistency across different benchmark suites suggests the 9745's advantage in general multi-threaded workloads is steady but not overwhelming, while its advantage in specialized compute tasks is where it truly runs away from the 9684X.

FAQ

Q: Which processor has more cores?

A: The AMD EPYC 9745 has 128 cores and 256 threads, while the AMD EPYC 9684X has 96 cores and 192 threads. The 9745 therefore offers 32 additional cores and 64 additional threads.

Q: How do their single-core performances compare?

A: The Cinebench R23 single-core scores favor the 9745 at 15683 versus 14591, a 7.5% lead. However, the Passmark single-thread score favors the 9684X at 2891 versus 2806, a 2.9% advantage. The two tests disagree on which processor has the better single-thread performance.

Q: What is the largest performance gap between these two processors?

A: The largest gap is in floating point math, where the EPYC 9745 leads by 61.2%, scoring 761219 against the 9684X's 472174. The second largest is in extended instructions at 57.6%, followed by prime number finding where the 9684X leads by 51.5% in the opposite direction.

Q: How do the processors compare in memory bandwidth?

A: The EPYC 9745 offers 576.0 GB/s of memory bandwidth, while the EPYC 9684X offers 460.8 GB/s. Both support DDR5 memory with a twelve-channel memory bus.

Q: Are both processors currently in production?

A: Yes, both the AMD EPYC 9745 and the AMD EPYC 9684X have an active production status and are aimed at the server and workstation market segment.

Q: What are the release dates for these processors?

A: The EPYC 9745 was released on 2024-10-09, while the EPYC 9684X was released on 2023-06-12. The 9745 is the newer part by over a year.

The Verdict

The data points to the AMD EPYC 9745 as the superior processor for the majority of workloads. Its 61.2% lead in floating point math and 57.6% lead in extended instructions indicate a major advantage in scientific computing, financial modeling, and other mathematically intensive tasks. The 45.6% lead in data compression and 45% lead in integer math further support its case for database workloads, analytics, and general server-side processing. The consistent 7.5% advantage across all Cinebench tests suggests that even workloads which do not specifically stress the 9745's strongest features will still see a measurable improvement.

The AMD EPYC 9684X, however, should not be dismissed. Its 51.5% lead in prime number finding is exceptional and points to strengths in cryptography and number-theoretic computations. The 30.6% lead in physics calculations indicates advantages in simulation workloads that rely on physics engines. The slight 2.9% single-thread Passmark lead, while small, could matter for lightly threaded applications that are sensitive to per-core performance.

Organizations running mixed enterprise workloads, large-scale virtualization, or data processing pipelines should favor the EPYC 9745. The raw throughput advantages in most measured categories, combined with the higher average benchmark score of 425973 against 266914, make it the safer default choice. Organizations with workloads specifically heavy in prime number operations or physics simulations should evaluate the 9684X, as its specialized wins could translate to meaningful real-world speedups in those narrow domains.

The 9745 also holds the edge in memory bandwidth at 576.0 GB/s versus 460.8 GB/s, which can benefit memory-bound applications beyond what the core count alone would suggest. For most server deployments, the data supports choosing the EPYC 9745.

Specification Differences

The two processors differ in several fundamental specifications. The EPYC 9745 belongs to the EPYC 9005 series with a Zen 5 architecture and the codename Turin, while the EPYC 9684X belongs to the EPYC 9004 series with a Zen 4 architecture and the codename Genoa-X. The 9745 uses a 3 nm process node, while the 9684X uses a 5 nm process node, both fabricated by TSMC.

Core and thread counts differ significantly, with the 9745 offering 128 cores and 256 threads against the 9684X's 96 cores and 192 threads. The base clock of the 9745 is 2.40 GHz, while the 9684X has a higher base clock of 2.55 GHz. Both share the same boost clock of 3.70 GHz and the same TDP of 400 watts.

Cache configurations show notable differences. The 9745 has 80 KB of L1 cache per core and 256 MB of shared L3 cache. The 9684X has 64 KB of L1 cache per core but a much larger 1152 MB of shared L3 cache. Both have 1 MB of L2 cache per core.

Memory bandwidth differs, with the 9745 rated at 576.0 GB/s and the 9684X at 460.8 GB/s, despite both supporting DDR5 over a twelve-channel memory bus. Both support ECC memory and offer PCIe Gen 5 with 128 lanes on the CPU only.

The 9684X has published transistor and die size data: 135,240 million transistors across 12 dies of 72 mm² each. The 9745 does not have transistor count or die size data recorded in the database. The launch MSRP for the 9745 is $12141, while the 9684X launched at $14756.

Architecture Differences

The architectural divide between these processors is substantial. The EPYC 9745 is built on Zen 5 architecture using a 3 nm process, representing AMD's latest generation of server cores. The EPYC 9684X uses the older Zen 4 architecture on a 5 nm process. This generational leap explains much of the performance differential, as Zen 5 delivers higher instructions per clock and improved efficiency across most workload types.

The 9684X's name carries the X suffix, which in AMD's naming scheme indicates the presence of 3D V-Cache technology. The database records its L3 cache at 1152 MB shared, a figure that reflects this cache-heavy design. The 9745's 256 MB of shared L3 cache is far smaller, though its faster process node and newer core architecture compensate in most benchmarks.

The transistor counts tell a story of design philosophy. The 9684X packs 135,240 million transistors across 12 dies of 72 mm² each, a configuration that allows for the massive L3 cache. The 9745's transistor count is not recorded in the database, but its 3 nm process allows for 128 cores in the same socket power envelope of 400 watts.

The L1 cache difference, 80 KB per core on the 9745 versus 64 KB per core on the 9684X, suggests Zen 5's improved per-core cache hierarchy. Both processors use the AMD Socket SP5 and support DDR5 memory with twelve channels, indicating platform continuity between the generations. The 9684X's higher base clock of 2.55 GHz versus 2.40 GHz on the 9745 shows that the older part compensates for its architectural deficit with a slightly higher operating frequency at idle and low load.

The 9745's newer release date of 2024-10-09, compared to the 9684X's 2023-06-12, aligns with the architectural progression from Zen 4 to Zen 5. The database does not record an integrated graphics solution for either processor, consistent with their server-focused design.

Where Each One Wins

The AMD EPYC 9745 wins in the majority of measured scenarios. Its 61.2% lead in floating point math makes it the clear choice for scientific computing, engineering simulations, and any workload that relies heavily on floating-point operations. The 57.6% advantage in extended instructions covers workloads that leverage advanced CPU instruction sets, including vectorized code and multimedia processing. Data compression tasks show a 45.6% lead, making the 9745 preferable for storage systems, database compression, and data transfer pipelines. The 45% advantage in integer math supports its use in general-purpose server workloads, financial calculations, and data processing. Encryption tasks run 28.6% faster on the 9745, which matters for secure communications, VPN gateways, and encrypted storage. Random string sorting benefits from a 34.3% lead, relevant for sorting-heavy data workloads and certain database operations.

The 9745 also wins all six Cinebench tests by 7.5%, covering both single-core and multi-core rendering and 3D modeling scenarios. Its Passmark multi-thread score of 130698 against 121595, a 7.5% lead, confirms its advantage in general multi-threaded throughput. The 9745's higher memory bandwidth of 576.0 GB/s versus 460.8 GB/s gives it an edge in memory-bound applications where data movement is the bottleneck.

The AMD EPYC 9684X wins in specific niches. Its 51.5% lead in prime number finding makes it the superior choice for cryptography, number theory research, and certain hashing algorithms. The 30.6% advantage in physics calculations points to strengths in physics-based simulations, game physics, and computational physics workloads. Its 2.9% lead in Passmark single-thread performance, while small, could benefit applications that are heavily dependent on single-core speed and cannot parallelize effectively.

The 9684X's massive 1152 MB L3 cache likely explains its wins in cache-sensitive workloads. Prime number finding and physics simulations often benefit from large working sets that fit in cache, and the 9684X's enormous L3 capacity gives it a structural advantage in these scenarios despite the 9745's newer architecture. For organizations running workloads that fit within the 9684X's cache footprint and rely on the mathematical operations where it excels, the 9684X remains a viable choice. For everything else, the 9745's data demonstrates a clear and substantial performance advantage.