Intel Xeon 678X vs Intel Xeon 6980P Comparison

The recorded data for the Intel Xeon 6980P and the Intel Xeon 678X tells a story of two Granite Rapids parts aimed at different buyers, and the split is stranger than the spec sheet suggests. On paper, the 6980P is the monster: 128 cores, 256 threads, twelve-channel DDR5 and 614.4 GB/s of memory bandwidth. Yet the benchmark log shows the 678X, with only 48 cores and a 300W TDP, winning 10 of 16 head-to-head tests against the 6980P's 6. How does a part with barely a third of the cores win the majority of recorded tests? The answer lies in clock speeds, and it reshapes the entire comparison.

Where Each One Wins

The database splits cleanly along two axes: thread count versus per-thread speed.

The Xeon 678X wins everything that rewards frequency and responsiveness. Its boost clock of 4.90 GHz, against 3.90 GHz on the 6980P, carries it to victory in every Cinebench test, single-threaded and multi-threaded alike, plus PassMark single thread, physics, find-prime-numbers, and the overall PassMark multithread rating. That is a remarkable result: a 48-core workstation part outscoring a 128-core flagship in Cinebench R23 multi-core, 83775 to 63175, a 24.6 percent gap in the 678X's favor.

The Xeon 6980P wins the throughput workloads, and by enormous margins. Integer math leads by 57.4 percent, extended instructions by 51.9 percent, encryption by 49.8 percent, string sorting by 46.7 percent, data compression by 39.8 percent and floating point math by 38.6 percent. These are sustained, parallel, memory-hungry computations, and that is exactly where 128 cores, 504 MB of shared L3 and twelve memory channels pay off. The pattern suggests the 6980P is built for scale-out server workloads: virtualization, databases, scientific computing, anything that fills every thread. The 678X, carrying the Xeon 600 Granite Rapids-WS designation, is built for workstation software that still leans on a smaller number of fast threads.

Architecture Differences

Both chips share the Granite Rapids architecture, Intel's 5 nm process, and identical per-core cache: 112 KB of L1 and 2 MB of L2 per core. From there they diverge sharply.

The 6980P belongs to the Granite Rapids-AP tier and uses three dies of 598 mm² each, feeding 128 cores and 256 threads with 504 MB of shared L3. Its memory subsystem is the standout: twelve-channel DDR5 delivering 614.4 GB/s of bandwidth. It sits in Socket 7529, carries a 500W TDP, and provides 96 lanes of PCIe Gen 5 from the CPU.

The 678X is a Granite Rapids-WS part with two 598 mm² dies, 48 cores, 96 threads and 192 MB of shared L3. Memory drops to eight-channel DDR5 at 409.6 GB/s of bandwidth, but PCIe expands to 128 Gen 5 CPU lanes, an unusual inversion that matters for GPU-heavy workstation builds. It uses Socket 4710 and a 300W TDP. Crucially, the 678X has an unlocked multiplier; the 6980P does not. Base and boost clocks also differ by design philosophy: 2.00/3.90 GHz on the big part, 2.40/4.90 GHz on the small one. Neither has integrated graphics, both support ECC, and both sit in the 99th percentile versus all CPUs in the database.

Head-to-Head Benchmarks

Start with the most surprising line in the table. In Cinebench R23 multi-core, the Xeon 678X scores 83775 against 63175 for the 6980P, a 24.6 percent win for the part with fewer than half the cores. The same 24.6 percent margin repeats across Cinebench R15 and R20 multi-core, and again in the PassMark multithread rating (98559 versus 74324). This consistency points to one cause: the 678X sustains much higher clocks per core, and rendering simply does not scale to 256 threads efficiently enough for the 6980P to compensate.

Single-threaded results confirm it. The 678X scores 3758 in PassMark single thread versus 1681 for the 6980P, a 55.3 percent lead, and mirrors it in Cinebench single-core tests at roughly 24.7 percent. Physics amplifies the trend: 7809 versus 3350, a 57.1 percent win for the 678X. Find-prime-numbers, a test where fewer, faster threads dominate, goes to the 678X by 45 percent (1010 versus 555).

Now flip to raw computation and the picture inverts. Integer math: 637476 versus 405075, the 6980P ahead by 57.4 percent. Extended instructions: 214794 versus 141431, up 51.9 percent. Encryption: 125246 versus 83598, up 49.8 percent. String sorting: 240792 versus 164102, up 46.7 percent. Compression: 2364519 versus 1690896, up 39.8 percent. Floating point: 501720 versus 362070, up 38.6 percent. Every one of these is a parallel integer or vector workload, and every one belongs to the 128-core part.

Context from the rivals sharpens the frame. The 6980P's average score of 251516 places it 3 percent ahead of the AMD EPYC 9634 and 5.6 percent ahead of the Intel Xeon 6747P, but 5.8 percent behind the AMD EPYC 9684X and 10.1 percent behind the AMD Ryzen Threadripper 9970X. The 678X, averaging 193477, sits in a tighter cluster: essentially even with the AMD EPYC 9335 (down 0.4 percent) and the Intel Xeon 6741P (down 0.7 percent), ahead of the Intel Xeon 6740E by 3.1 percent and the AMD EPYC 8534P by 4.5 percent.

FAQ

Q: The Xeon 6980P has far more cores, so why does it lose most benchmarks?

A: Its boost clock tops out at 3.90 GHz versus 4.90 GHz on the 678X. Workloads that do not scale across 256 threads, such as rendering and single-threaded tests, favor the faster-clocked part, which wins 10 of 16 recorded tests.

Q: Which chip is better for number-crunching workloads like encryption and integer math?

A: The 6980P, decisively. It leads integer math by 57.4 percent, extended instructions by 51.9 percent, encryption by 49.8 percent and floating point math by 38.6 percent, while offering 614.4 GB/s of memory bandwidth versus 409.6 GB/s.

Q: Which platform offers more PCIe lanes?

A: The 678X, with 128 Gen 5 CPU lanes against 96 on the 6980P, despite being the smaller chip.

Q: Can either processor be overclocked?

A: Only the 678X has an unlocked multiplier. The 6980P is locked.

Q: Are the two chips socket-compatible?

A: No. The 6980P uses Intel Socket 7529, while the 678X uses Intel Socket 4710.

Q: How do they compare against AMD rivals in the database?

A: The 6980P trades blows, leading the EPYC 9634 by 3 percent but trailing the EPYC 9684X by 5.8 percent and the Threadripper 9970X by 10.1 percent. The 678X is essentially tied with the EPYC 9335 at 0.4 percent behind.

Specification Differences

• Cores/threads: 128/256 (6980P) versus 48/96 (678X)
• Clocks: 2.00 GHz base and 3.90 GHz boost versus 2.40 GHz base and 4.90 GHz boost
• TDP: 500W versus 300W
• Socket: Intel Socket 7529 versus Intel Socket 4710
• Generation tier: Xeon 6 Granite Rapids-AP versus Xeon 600 Granite Rapids-WS
• Dies: 3x 598 mm² versus 2x 598 mm²
• L3 cache: 504 MB shared versus 192 MB shared
• Memory bus: twelve-channel versus eight-channel DDR5
• Memory bandwidth: 614.4 GB/s versus 409.6 GB/s
• PCIe: Gen 5 with 96 CPU lanes versus Gen 5 with 128 CPU lanes
• Multiplier: locked versus unlocked
• Launch MSRP: $12460 for the 6980P, $3749 for the 678X
• Release timing: the 6980P entered the database earlier, the 678X later

Shared traits include the Granite Rapids architecture, Intel 5 nm fabrication, identical per-core L1 and L2 cache, DDR5 support, ECC capability, no integrated graphics, and 99th-percentile standing against all CPUs in the database.

The Verdict

The data points to a clear division of labor. Choose the Xeon 6980P when the workload is genuinely parallel and throughput-bound: encryption, compression, integer and floating point math, large in-memory databases fed by twelve channels of DDR5. The 57.4 percent integer-math lead over the 678X is the kind of margin that compounds across a server fleet, and the 504 MB of L3 exists to keep those 128 cores fed.

Choose the Xeon 678X when software behavior matters more than core count. It wins every Cinebench render test by 24.6 percent, every single-threaded test by a wide margin, and the overall PassMark multithread rating, all while drawing a 300W TDP instead of 500W. Its 128 PCIe Gen 5 lanes and unlocked multiplier make it the flexible workstation choice, and its average score lands within a point of the AMD EPYC 9335 and Intel Xeon 6741P, indicating a tightly contested segment.

The uncomfortable question the data raises is aimed at the 6980P: a flagship that loses most general-purpose tests to a sibling with a third of its cores. But the six tests it does win are the heaviest sustained computations in the suite, and there the margin is consistently near 40 to 57 percent. For buyers matching software to silicon rather than chasing headline core counts, that distinction is the whole decision.