AMD EPYC 4464P vs Intel Core Ultra 9 275HX Comparison

The Intel Core Ultra 9 275HX and the AMD EPYC 4464P land within touching distance of each other in the database's aggregate rankings, with average benchmark scores of 67469 and 64823 respectively, yet they arrive there by completely different routes. The 275HX is a 24-core mobile flagship built on Arrow Lake-HX silicon; the EPYC 4464P is a 12-core, 24-thread server part on AMD's Zen 4 Raphael die. The recorded data shows the Intel chip winning 13 of 17 head-to-head tests, dominated by wide margins in compute-heavy workloads, while the EPYC fights back in a handful of Cinebench tests, including a dramatic single-thread victory in Cinebench R23. Both sit in the top few percent of all CPUs measured, at the 94th and 93rd percentiles respectively, making this a comparison between two elite parts with sharply different strengths.

FAQ

Q: Which CPU is faster overall?

A: The Intel Core Ultra 9 275HX leads on aggregate, with an average benchmark score of 67469 versus 64823 for the EPYC 4464P, and it wins 13 of the 17 recorded head-to-head tests. Its largest win is in PassMark floating point math, where it scores 191186 against 93090, a gap of 105.4 percent.

Q: Which CPU is better for single-threaded workloads?

A: It depends heavily on the test. The 275HX wins PassMark single-thread by 13.7 percent (4713 vs 4146) and Cinebench R20 single-core by 17.8 percent (2809 vs 2384). But the EPYC 4464P wins Cinebench R23 single-core by a massive 61.2 percent (5677 vs 2204) and Cinebench R15 single-core by 41.6 percent (572 vs 334).

Q: Do these CPUs support ECC memory?

A: Only the AMD EPYC 4464P. Its server/workstation heritage includes ECC memory support; the 275HX does not record ECC support. Both support DDR5 on a dual-channel bus, though the 275HX offers higher memory bandwidth at 102.4 GB/s versus 83.2 GB/s.

Q: Are they made on the same process node?

A: No. Both use TSMC as foundry, but the 275HX is built on a 3 nm process while the EPYC 4464P uses a 5 nm process. The Intel part packs 17,800 million transistors on a 243 mm² die; the AMD part has 13,140 million transistors across two 71 mm² dies.

Q: Can either CPU be overclocked?

A: Yes, the Core Ultra 9 275HX has an unlocked multiplier. The EPYC 4464P does not.

Q: How do they compare to their own nearest rivals?

A: The 275HX sits essentially level with the Intel Xeon w5-3525 (within 0.3 percent), the AMD EPYC 4484PX (0.5 percent) and the AMD Ryzen Threadripper PRO 5955WX (0.6 percent). The EPYC 4464P is similarly bracketed by the Intel Core Ultra 7 265 (0.3 percent), AMD EPYC 9124 (0.4 percent) and Intel Core Ultra 7 265F (0.6 percent).

Architecture Differences

These two chips could hardly be more architecturally distinct despite similar aggregate performance. The Core Ultra 9 275HX belongs to Intel's Core Ultra Series 2, using the Arrow Lake architecture on the Arrow Lake-HX codename, manufactured by TSMC on a 3 nm node. It offers 24 cores and 24 threads, notably with no hyper-threading-style doubling: 24 cores map to 24 threads. Its cache hierarchy is wide and shallow relative to the AMD part, with 192 KB of L1 per core, 3 MB of L2 per core, and a shared 36 MB of L3.

The EPYC 4464P is a different animal. It uses AMD's Zen 4 architecture under the Raphael codename, also fabbed by TSMC but on a 5 nm process spread across two 71 mm² chiplets. It has 12 cores and 24 threads, meaning SMT is active and each core handles two threads. Its per-core caches are smaller, 64 KB of L1 and 1 MB of L2 per core, but the shared L3 is far larger at 64 MB. That near-doubling of L3 capacity is the defining architectural contrast here and helps explain the AMD part's strong showings in some rendering tests.

Platform features diverge too. The EPYC 4464P provides 28 PCIe Gen 5 lanes from the CPU against 20 on the Intel side, plus ECC memory support, both of which matter for its server/workstation market segment. The 275HX counters with higher memory bandwidth (102.4 GB/s vs 83.2 GB/s), integrated Arc Xe-LPG Graphics with 64 EUs versus AMD's basic Radeon Graphics, and an unlocked multiplier. Sockets are incompatible by design: Intel BGA 2114, a soldered mobile interface, versus AMD's Socket AM5.

Head-to-Head Benchmarks

The Intel chip's win column is broad and deep. The single most lopsided result in the entire dataset is PassMark floating point math: 191186 for the 275HX against 93090 for the EPYC, a 105.4 percent advantage. Cinebench R15 multi-core goes to Intel by 38.7 percent (5619.5 vs 4053), data encryption by 31.5 percent (47112 vs 35816), prime number finding by 30.6 percent (448 vs 343), and extended instructions by 19.5 percent (47016 vs 39359). Cinebench R20 multi-core lands at 19899 vs 16890 for a 17.8 percent Intel win, PassMark multithread at 55759 vs 47514 for 17.4 percent, and physics at 3338 vs 2873 for 16.2 percent. Smaller but consistent Intel wins include single-thread (13.7 percent), random string sorting (5.9 percent) and data compression (5.9 percent).

The AMD fightback is concentrated but real. Cinebench R23 single-core is the standout: 5677 for the EPYC versus 2204 for the 275HX, a 61.2 percent margin that is the largest single-test gap in either direction. Cinebench R15 single-core also goes to AMD by 41.6 percent (572 vs 334). More surprising is Cinebench R23 multi-core, where the 12-core EPYC beats the 24-core Intel part, 40215 versus 35589, an 11.5 percent AMD win despite having half the physical cores. Integer math narrowly favors AMD as well, 160410 versus 155218, a 3.2 percent edge.

The pattern is clear. The 275HX dominates sustained, vector-heavy and integer-heavy parallel workloads; the EPYC 4464P excels in certain rendering paths, where its large 64 MB L3 cache and SMT threads evidently pay off, and in the single-threaded Cinebench tests its higher 3.70 GHz base clock appears to matter.

Specification Differences

The two differ on nearly every specification line:

• Cores/threads: 24C/24T (275HX) vs 12C/24T (EPYC 4464P)
• Base clock: 2.70 GHz vs 3.70 GHz; boost clock is identical at 5.40 GHz
• TDP: 55 W vs 65 W
• Architecture: Arrow Lake (Arrow Lake-HX) vs Zen 4 (Raphael)
• Process node: 3 nm vs 5 nm, both TSMC
• Transistors/die: 17,800 million on 243 mm² vs 13,140 million on 2x 71 mm²
• Cache: 192 KB L1 and 3 MB L2 per core with 36 MB shared L3 vs 64 KB L1 and 1 MB L2 per core with 64 MB shared L3
• Memory bandwidth: 102.4 GB/s vs 83.2 GB/s, both DDR5 dual-channel
• ECC: false vs true
• PCIe: Gen 5 with 20 CPU lanes vs Gen 5 with 28 CPU lanes
• Graphics: Arc Xe-LPG 64EU vs Radeon Graphics
• Segment/socket: Mobile, BGA 2114 vs Server/Workstation, AM5
• Release: January 2025 vs May 2024; launch MSRP of $429 recorded for the EPYC
• Multiplier: unlocked vs locked

Where Each One Wins

The data assigns the 275HX to anyone whose workload scales across many threads and leans on numerical compute. Floating point math, encryption, extended instruction throughput, physics simulation and general multithreading all favor it, most by double-digit percentages. Its higher memory bandwidth and stronger integrated graphics reinforce the case for it as a self-contained mobile workstation. Its percentile placement, 94th against all CPUs in the database, confirms top-tier standing, and its rival list (Xeon w5-3525, EPYC 4484PX, Threadripper PRO 5955WX) shows it trading blows with desktop workstation silicon.

The EPYC 4464P wins where rendering and per-thread responsiveness dominate. Cinebench R23 is its showcase, winning both single-core (by 61.2 percent) and multi-core (by 11.5 percent), and it also takes R15 single-core and integer math. Add ECC memory, eight additional PCIe Gen 5 lanes and AM5 platform flexibility, and the recorded data points toward it as the choice for stability-focused deployments and rendering-centric tasks.

The Verdict

For raw throughput across the widest range of measured workloads, the Intel Core Ultra 9 275HX is the stronger CPU: 13 wins from 17 tests, a higher average score, and crushing margins in floating point and encryption. Choose it for compute-bound, heavily parallel work.

Choose the AMD EPYC 4464P when single-threaded Cinebench-class performance, ECC memory, expanded PCIe lanes and a serviceable AM5 platform outweigh sheer multi-core output. The database shows both CPUs separated by a single percentile point, so the deciding factor is workload profile, not overall class.