AMD EPYC 9655 vs AMD EPYC 9745 Comparison

The AMD EPYC 9745 and AMD EPYC 9655 are both Zen 5 Turin processors on the same SP5 socket, yet they are engineered for different ends of the server spectrum. The 9745 is the dense-core specialist, while the 9655 is the high-frequency generalist. The benchmark data shows a clear division of labor: the 9745 dominates in throughput-heavy, data-parallel workloads, while the 9655 wins decisively in single-threaded and lightly-threaded tasks. Across 17 head-to-head benchmarks, the 9655 takes 11 wins while the 9745 secures 6, but the margins of victory reveal a more nuanced story than the raw win count suggests.

The Verdict

From the data, the AMD EPYC 9655 is the right choice for workloads where per-core speed and lower latency matter more than raw core count. It wins every Cinebench test — multicore and singlecore alike — by a consistent 16.3% margin over the 9745. In PassMark’s multithread test, the 9655 scores 156085 versus 130698, a 16.3% lead, and it crushes the 9745 in physics simulation (25958 vs 17122, a 34% advantage) and prime number finding (1599 vs 979, a 38.8% lead). For database queries, web serving, or any latency-sensitive enterprise application, the 9655’s higher boost clock and larger L3 cache make it the data-backed pick.

The AMD EPYC 9745 is the choice for scale-out, throughput-oriented environments that can use every one of its 128 cores. It wins data compression (3929890 vs 3271019, a 20.1% lead), extended instructions (280477 vs 203302, a 38% lead), and floating-point math (761219 vs 662949, a 14.8% lead). The 9745 also edges ahead in integer math (7.5%) and random string sorting (6.4%). If the software stack scales linearly with core count, the 9745’s 128 cores versus 96 cores provide a clear capacity advantage, even if the per-core speed is lower.

FAQ

Q: Which processor has more cores and threads?

A: The AMD EPYC 9745 has 128 cores and 256 threads. The AMD EPYC 9655 has 96 cores and 192 threads. The 9745 offers a 33% core-count advantage.

Q: How do their single-threaded scores compare?

A: The EPYC 9655 is markedly faster in single-threaded tests. In Cinebench R23 singlecore, the 9655 scores 18730 versus the 9745’s 15683, a 16.3% difference. PassMark single-thread results show a 27% lead for the 9655 (3845 vs 2806).

Q: Which one wins in multi-core rendering benchmarks?

A: The EPYC 9655 wins all three Cinebench multicore tests. In Cinebench R23 multicore, it scores 132672 against the 9745’s 111093, a 16.3% advantage. This is counterintuitive given the 9745’s extra cores, but the data is consistent across R15, R20, and R23.

Q: Are there workloads where the 9745 is significantly ahead?

A: Yes. The 9745 has a 38% lead in PassMark’s extended instructions test (280477 vs 203302) and a 20.1% lead in data compression (3929890 vs 3271019). It also wins in floating-point math by 14.8% and integer math by 7.5%.

Q: What are the architectural differences that explain these results?

A: Both use Zen 5 architecture and the Turin codename, but the 9745 is built on a 3 nm process while the 9655 uses a 4 nm process. The 9745 uses the "Zen 5c" variant, while the 9655 uses the full "Zen 5" variant. The 9655 has a higher base clock (2.60 GHz vs 2.40 GHz) and boost clock (4.50 GHz vs 3.70 GHz).

Q: How does the L3 cache differ between the two?

A: The EPYC 9655 has 384 MB of shared L3 cache, while the EPYC 9745 has 256 MB. The L1 cache (80 KB per core) and L2 cache (1 MB per core) are identical.

Architecture Differences

The two processors share the same Zen 5 architecture and Turin codename, but they are fabricated differently. The EPYC 9745 uses a 3 nm process node from TSMC, while the EPYC 9655 uses a 4 nm process. This node difference allows the 9745 to pack more cores into the same power envelope, but it comes with a clock speed penalty. The 9745’s generation is listed as "EPYC (Zen 5c (Turin))", indicating it uses the dense "c" variant of Zen 5, whereas the 9655 is listed as "EPYC (Zen 5 (Turin))" and uses the standard Zen 5 core design.

The cache hierarchy differs in total L3 capacity. The 9745 has 256 MB of shared L3, while the 9655 has 384 MB — a 50% larger pool for the 9655. Per-core L1 and L2 caches are identical (80 KB and 1 MB per core, respectively). The 9655’s larger L3 cache is likely a significant factor in its single-threaded and physics test wins, as more data can be held closer to the cores.

The transistor count and die size are only listed for the 9655: 99,780 million transistors spread across 12 dies of 70.6 mm² each. No such data is provided for the 9745. Both processors use the same AMD Socket SP5 platform and are built by TSMC.

Specification Differences

The core and thread counts are the most obvious difference: 128 cores / 256 threads for the 9745 versus 96 cores / 192 threads for the 9655. Clock speeds also differ substantially. The 9745 has a base clock of 2.40 GHz and a boost clock of 3.70 GHz. The 9655 runs at a base clock of 2.60 GHz and boosts to 4.50 GHz. The 9655’s boost clock is 0.80 GHz higher, which directly explains its single-threaded dominance.

Both chips have a TDP of 400 watts and support DDR5 memory over a twelve-channel bus with a memory bandwidth of 576.0 GB/s. They both support ECC memory and offer PCIe Gen 5 with 128 lanes (CPU only). The L3 cache differs as noted: 256 MB for the 9745 versus 384 MB for the 9655. The process node differs (3 nm vs 4 nm), and the 9655 has a documented transistor count and die size while the 9745 does not. The launch MSRP for the 9745 is $12141; for the 9655 it is $11852.

Head-to-Head Benchmarks

The Cinebench suite shows a uniform pattern: the 9655 wins every test by exactly 16.3%. In Cinebench R15, the 9655 scores 13373 multicore and 1887 singlecore, versus the 9745’s 11198 and 1580. In R20, the 9655 posts 55722 multicore and 7866 singlecore against the 9745’s 46659 and 6586. In R23, the 9655 hits 132672 multicore and 18730 singlecore, while the 9745 manages 111093 and 15683. The consistent 16.3% delta across all six Cinebench tests suggests the 9655’s higher clock speed provides a uniform advantage in this rendering workload, regardless of thread count.

PassMark results are split. The 9655 wins the multithread test (156085 vs 130698, -16.3% for the 9745), physics (25958 vs 17122, -34%), prime number finding (1599 vs 979, -38.8%), and single-thread (3845 vs 2806, -27%). The 9745 counters with wins in data compression (3929890 vs 3271019, +20.1%), data encryption (229447 vs 210541, +9%), extended instructions (280477 vs 203302, +38%), floating-point math (761219 vs 662949, +14.8%), integer math (1224315 vs 1139221, +7.5%), and random string sorting (468975 vs 440608, +6.4%).

The largest single delta is in extended instructions, where the 9745 is 38% ahead. The largest win for the 9655 is in prime number finding, where it leads by 38.8%. The physics test is another big win for the 9655 at 34% ahead. These results indicate that the 9655 excels in workloads that are latency-bound or have limited parallelism, while the 9745 shines in highly parallel, data-intensive operations.

Where Each One Wins

The AMD EPYC 9655 wins in every Cinebench test, all three of which are multicore and singlecore rendering benchmarks. This means for 3D rendering, video encoding, and other tasks that rely on high per-core IPC and clock speed, the 9655 is the superior choice. Its wins in PassMark physics (34% ahead) and prime number finding (38.8% ahead) suggest it is better suited for scientific simulations, financial modeling, and any workload with significant sequential dependencies. The single-thread win of 27% reinforces its strength in database transactions, web serving, and real-time analytics where individual thread latency is critical.

The AMD EPYC 9745 wins in data compression by 20.1%, indicating it is the better fit for data warehousing, backup systems, and log processing where large volumes of data must be compressed or decompressed. Its 38% lead in extended instructions points to strengths in cryptography, signal processing, and other SIMD-heavy workloads. The wins in floating-point math (14.8%) and integer math (7.5%) make it suitable for high-performance computing, scientific computing, and machine learning inference where raw throughput across many cores is paramount. The 9% lead in data encryption and 6.4% lead in random string sorting further support its use in big data and search infrastructure.

The overall benchmark counts favor the 9655 (11 wins vs 6), but the 9745’s victories come in categories that are often more representative of large-scale server workloads. The 9655’s wins are more evenly distributed across rendering and general compute, while the 9745’s wins are concentrated in data manipulation and math-heavy tasks. For a mixed workload server, the 9655 offers a more balanced profile; for a dedicated throughput engine, the 9745’s extra cores and specialized strengths are the data-backed choice.