Intel Core Ultra 9 290HX Plus vs Intel Xeon 638 Comparison

The Intel Xeon 638 and Intel Core Ultra 9 290HX Plus occupy opposite ends of the computing spectrum, yet both land in the 95th percentile of all CPUs. The Xeon is a 16-core, 32-thread Granite Rapids workstation part aimed at sustained throughput, while the Ultra 9 is a 24-core, 24-thread Arrow Lake-HX mobile chip designed for high single-thread performance in a 55 W envelope. The benchmark data shows a split decision: the Xeon dominates in older Cinebench single-core tests and several PassMark integer workloads, while the Ultra 9 sweeps the modern Cinebench multi-core tests and floating-point math. Neither chip is a universal victor; the choice hinges entirely on workload type and platform constraints.

Head-to-Head Benchmarks

The most dramatic result in this comparison is the Cinebench R23 single-core test, where the Xeon 638 scores 6663 against the Ultra 9’s 2356, a 182.8% advantage. This is not a small margin; the Xeon is nearly three times faster in that specific legacy benchmark. The same pattern appears in Cinebench R15 single-core, with the Xeon winning 671 to 340 (97.4% ahead). These are enormous wins for the Xeon, suggesting its Granite Rapids cores retain exceptional per-thread performance in older instruction paths or that the Ultra 9’s boost behavior is heavily constrained in this test.

However, the Ultra 9 strikes back in modern multi-core workloads. In Cinebench R15 multi-core, the Ultra 9 scores 5981 versus the Xeon’s 4757, a 20.5% lead. In Cinebench R20 multi-core, the Ultra 9 wins 21198 to 19824 (6.5% ahead). Yet the Xeon flips the script in Cinebench R23 multi-core, taking 47202 against 39684, an 18.9% margin. This inconsistency across Cinebench versions points to scaling differences: the Xeon’s 32 threads help it in R23’s longer render, while the Ultra 9’s higher boost clocks win the shorter R15 and R20 runs.

PassMark results further muddy the waters. The Ultra 9 leads decisively in floating-point math (201773 vs 144757, 28.3% ahead) and data encryption (50008 vs 36030, 28% ahead). It also wins find prime numbers (519 vs 381, 26.6% ahead) and single-thread performance (4951 vs 3670, 25.9% ahead). The Xeon counters with wins in integer math (184884 vs 164839, 12.2% ahead), extended instructions (56498 vs 51290, 10.2% ahead), and data compression (725818 vs 658724, 10.2% ahead). Physics is a Xeon blowout: 4704 vs 3387, a 38.9% lead.

The overall win count is 10 for the Ultra 9 and 7 for the Xeon, but that raw tally hides the magnitude of the Xeon’s single-core Cinebench victories. The Ultra 9’s wins tend to be in the 6–28% range, while the Xeon’s two biggest wins are 97.4% and 182.8%. For anyone running legacy single-threaded Cinebench, the Xeon is the obvious choice. For mixed PassMark workloads, the Ultra 9’s balanced performance makes it more versatile.

Architecture Differences

The two CPUs are built on fundamentally different nodes and foundries. The Xeon 638 uses Intel’s 5 nm process and a 598 mm² die, while the Ultra 9 uses TSMC’s 3 nm node with a 243 mm² die and 17,800 million transistors. The Xeon’s architecture is Granite Rapids, specifically the Xeon 600 series (Granite Rapids-WS), with 16 cores and 32 threads. The Ultra 9 is Arrow Lake-HX Refresh from the Core Ultra Series 2, with 24 cores and 24 threads—no hyperthreading.

Cache hierarchies diverge sharply. The Xeon provides 112 KB L1 per core, 2 MB L2 per core, and a massive 72 MB shared L3. The Ultra 9 counters with 192 KB L1 per core, 3 MB L2 per core, but only 36 MB shared L3. The Xeon’s triple-level cache totals far more capacity, which likely explains its data compression win (725818 vs 658724). The Ultra 9’s larger per-core L1 and L2 help with latency-sensitive single-thread tasks, but its shared L3 is half the Xeon’s.

Memory subsystems are equally distinct. The Xeon uses quad-channel DDR5 with 204.8 GB/s bandwidth and 80 PCIe Gen 5 lanes. The Ultra 9 is dual-channel DDR5 at 102.4 GB/s and only 20 PCIe Gen 5 lanes. Both support ECC memory, but the Xeon’s quad-channel bandwidth is double the Ultra 9’s, which matters for server-style workloads. The Ultra 9 includes integrated Arc Xe-LPG Graphics 64EU; the Xeon has no integrated graphics.

Socket and power are the final major splits. The Xeon is an Intel Socket 4710 part with a 180 W TDP, while the Ultra 9 is BGA 2114 with a 55 W TDP. The Xeon is unlocked, as is the Ultra 9, but the 125 W TDP difference means the Xeon requires serious cooling and power delivery. The Ultra 9’s 3 nm node and lower TDP make it viable for mobile platforms. The Xeon’s 5 nm node is larger and less efficient, reflected in its higher die size and power draw.

Where Each One Wins

The Xeon 638 wins where raw cache capacity and thread count matter. Its 72 MB L3 and 32 threads deliver a 10.2% win in data compression and a 12.2% win in integer math. The physics result (4704 vs 3387, 38.9% ahead) suggests the Xeon’s memory bandwidth and cache hierarchy excel in simulation-style workloads. The Cinebench R23 multi-core win (47202 vs 39684) and the massive single-core Cinebench wins (671 vs 340 in R15, 6663 vs 2356 in R23) indicate that the Xeon’s core architecture is exceptionally strong in sustained rendering and legacy single-thread tests.

The Ultra 9 290HX Plus wins in floating-point math (201773 vs 144757, 28.3% ahead), which is critical for scientific computing, 3D rendering, and AI inference. Its data encryption lead (50008 vs 36030, 28% ahead) points to strong cryptographic throughput, likely from newer instruction extensions. The single-thread PassMark win (4951 vs 3670, 25.9% ahead) confirms that the Ultra 9 is the better general-purpose single-thread performer despite its poor Cinebench R23 single-core showing. The find prime numbers win (519 vs 381, 26.6% ahead) further supports the Ultra 9’s integer arithmetic efficiency.

The Ultra 9 also wins the multi-core PassMark test (59439 vs 55651, 6.4% ahead) and random string sorting (80327 vs 74318, 7.5% ahead). These are broad, real-world mixed workloads where the Ultra 9’s higher boost clock (5.50 vs 4.80) and newer node compensate for fewer threads. The Ultra 9’s wins in Cinebench R15 and R20 multi-core are smaller but consistent, suggesting it handles short bursts of multi-threaded work better than the Xeon.

FAQ

Q: Which CPU is faster in single-core performance?

A: It depends on the benchmark. The Xeon 638 wins Cinebench R15 single-core (671 vs 340, 97.4% ahead) and R23 single-core (6663 vs 2356, 182.8% ahead), but the Ultra 9 wins PassMark single-thread (4951 vs 3670, 25.9% ahead). The Xeon’s Cinebench wins are massive, but the Ultra 9’s PassMark win is more representative of general desktop tasks.

Q: Does the Ultra 9 have more cores than the Xeon?

A: Yes, but without hyperthreading. The Ultra 9 has 24 cores and 24 threads, while the Xeon has 16 cores and 32 threads. The Xeon has more threads overall, which helps in heavily parallel workloads like Cinebench R23 multi-core (47202 vs 39684, 18.9% ahead).

Q: Which CPU has better memory bandwidth?

A: The Xeon 638, with quad-channel DDR5 delivering 204.8 GB/s versus the Ultra 9’s dual-channel 102.4 GB/s. This exactly doubles the bandwidth, which benefits the Xeon’s data compression score (725818 vs 658724) and physics result (4704 vs 3387).

Q: Are both CPUs unlocked for overclocking?

A: Yes, both the Xeon 638 and the Ultra 9 290HX Plus have unlocked multipliers. However, the Xeon’s 180 W TDP versus the Ultra 9’s 55 W TDP means overclocking headroom is very different in practice.

Q: Which CPU has the larger L3 cache?

A: The Xeon 638, with 72 MB shared L3 versus the Ultra 9’s 36 MB. The Xeon also has more total cache when accounting for L1 and L2, though the Ultra 9 has larger per-core L1 (192 KB vs 112 KB) and L2 (3 MB vs 2 MB) caches.

Q: Can the Ultra 9 be used without a discrete GPU?

A: Yes, it includes integrated Arc Xe-LPG Graphics 64EU, whereas the Xeon 638 has no integrated graphics. This makes the Ultra 9 viable for compact or mobile systems without a dedicated graphics card.

The Verdict

The data points to two distinct buyers. The Intel Xeon 638 is the pick for anyone running legacy Cinebench benchmarks or workloads that depend on massive L3 cache and quad-channel memory bandwidth. Its 182.8% lead in Cinebench R23 single-core and 38.9% lead in physics are not outliers; they indicate a core design that excels in sustained, cache-heavy tasks. The 72 MB L3 and 32 threads make it a server/workstation workhorse, and the 204.8 GB/s bandwidth is non-negotiable for data compression (725818 score) and integer math (184884 score). Its launch MSRP is $899.

The Intel Core Ultra 9 290HX Plus is the better all-rounder for modern, mixed workloads. Its 28.3% win in floating-point math and 28% win in data encryption are directly applicable to AI, scientific, and security workloads. The PassMark single-thread win (4951 vs 3670) shows it handles everyday apps better. The 24 cores with 24 threads and 5.50 GHz boost clock deliver a 6.4% multi-thread PassMark win despite the thread disadvantage. The 55 W TDP and integrated graphics make it the only option for mobile or power-constrained builds.

Choose the Xeon 638 if your software is single-threaded legacy code, particularly Cinebench R23, or if you need maximum cache and memory bandwidth for server-style processing. Choose the Ultra 9 290HX Plus for general productivity, floating-point-heavy computation, encryption, and any deployment where power or space is limited. The Xeon’s Cinebench single-core dominance is impressive but narrow; the Ultra 9’s wins span more test categories and align better with current software trends.

Specification Differences

| Specification | Intel Xeon 638 | Intel Core Ultra 9 290HX Plus |

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

| Cores | 16 | 24 |

| Threads | 32 | 24 |

| Base Clock | 3.20 GHz | 2.70 GHz |

| Boost Clock | 4.80 GHz | 5.50 GHz |

| TDP | 180 W | 55 W |

| Socket | Intel Socket 4710 | Intel BGA 2114 |

| Architecture | Granite Rapids | Arrow Lake-HX Refresh |

| Process Node | 5 nm | 3 nm |

| Foundry | Intel | TSMC |

| Die Size | 598 mm² | 243 mm² |

| L1 Cache | 112 KB (per core) | 192 KB (per core) |

| L2 Cache | 2 MB (per core) | 3 MB (per core) |

| L3 Cache | 72 MB (shared) | 36 MB (shared) |

| Memory Bus | Quad-channel | Dual-channel |

| Memory Bandwidth | 204.8 GB/s | 102.4 GB/s |

| PCIe | Gen 5, 80 Lanes | Gen 5, 20 Lanes |

| Integrated Graphics | N/A | Arc Xe-LPG Graphics 64EU |

| Market Segment | Server/Workstation | Mobile |

| Launch MSRP | $899 | N/A |