When we redesigned our VPS hardware fleet in 2024, we ran a structured benchmark series across the EPYC and Xeon options that mattered. The TL;DR: for VPS workloads (single-thread heavy, mixed read/write, latency-sensitive), AMD EPYC 9004/9005 series outperformed Intel Xeon 4th/5th Gen on every metric except one (very heavy AVX-512 workloads, where Intel still wins). This post walks through the methodology and the numbers, so you can verify the choice for your own workloads.
What we tested
Five hardware configurations:
- AMD EPYC 9554 (64-core "Genoa", 3.1 GHz base / 3.75 GHz boost)
- AMD EPYC 9654 (96-core "Genoa", 2.4 GHz base / 3.7 GHz boost)
- AMD EPYC 9755 (128-core "Turin", 2.7 GHz base / 4.1 GHz boost)
- Intel Xeon Platinum 8480+ (56-core "Sapphire Rapids", 2.0 GHz base / 3.8 GHz boost)
- Intel Xeon Platinum 8592+ (64-core "Emerald Rapids", 1.9 GHz base / 3.9 GHz boost)
Each was tested as the host of a 4-vCPU / 8 GB / 80 GB NVMe Linux VPS, with the test workload running as a guest. We chose this configuration because it represents the median of our customer fleet — testing single-thread on a 96-core monster doesn't tell you much about what an actual customer experiences.
Single-thread performance
The single most-cited benchmark in VPS comparisons is single-core performance, and for good reason: most production workloads (Node.js event loops, Python web servers, single-threaded job runners) are bottlenecked on per-core speed.
Geekbench 6 single-core scores, average of 10 runs:
| CPU | Score | vs Best |
|---|---|---|
| EPYC 9755 (Turin) | 2,812 | +0% |
| Xeon 8592+ (Emerald) | 2,415 | −14% |
| EPYC 9654 (Genoa) | 2,388 | −15% |
| EPYC 9554 (Genoa) | 2,290 | −18% |
| Xeon 8480+ (Sapphire) | 2,047 | −27% |
EPYC 9755 ("Turin") leads. The 4.1 GHz boost combined with the larger L3 cache and improved branch predictor gives it about 14% advantage over the closest Intel competitor. For our default Cloud VPS tier, we use Genoa-X (EPYC 9554) which trades absolute peak frequency for larger cache — better for cache-sensitive workloads like databases.
Multi-thread performance and core scaling
For workloads that genuinely scale across cores (compilation, video encoding, parallelizable jobs), the picture is different — but EPYC still wins on absolute throughput because of higher core counts.
Linux kernel compile time (allyesconfig, full build):
| CPU | Time | Speedup vs Xeon 8480+ |
|---|---|---|
| EPYC 9755 (128c) | 92 sec | 2.7× |
| EPYC 9654 (96c) | 108 sec | 2.3× |
| EPYC 9554 (64c) | 148 sec | 1.7× |
| Xeon 8592+ (64c) | 198 sec | 1.3× |
| Xeon 8480+ (56c) | 248 sec | baseline |
The EPYC 9755 is 2.7× faster than Sapphire Rapids on this workload. Some of that is core count, but normalized per-core, EPYC still wins by 18-22% — primarily due to the more efficient L3 cache hierarchy and higher memory bandwidth.
Database performance: the workload that matters most
For our customer base, database performance is the workload that actually drives infrastructure decisions. We benchmarked PostgreSQL 16 with pgbench (TPC-B variant, 100 GB dataset, scaled to 1000 clients).
Transactions per second (higher is better):
| CPU | TPS | P95 latency |
|---|---|---|
| EPYC 9554 (64c, 3D V-Cache) | 32,400 | 4.2 ms |
| EPYC 9755 (128c) | 29,800 | 4.6 ms |
| EPYC 9654 (96c) | 27,500 | 5.0 ms |
| Xeon 8592+ (64c) | 21,300 | 6.4 ms |
| Xeon 8480+ (56c) | 18,700 | 7.3 ms |
EPYC 9554 with 3D V-Cache wins this benchmark even over the higher-frequency 9755, because PostgreSQL's transaction processing is highly cache-sensitive. The extra 1.1 GB of L3 cache from the 3D V-Cache stacking lets the working set fit, which dramatically reduces L3 miss rates.
This is why our standard Cloud VPS uses 9554 with 3D V-Cache rather than the higher-frequency 9755. For database-heavy workloads (which is most production SaaS), cache wins over frequency.
Network and I/O
Both AMD and Intel deliver similar peak network throughput when configured similarly — at 40 Gbps line rate, the bottleneck isn't the CPU. However, on packet-rate-sensitive workloads (load balancers, proxies, real-time streaming), there's a difference.
iperf3 with small packets (64-byte payloads, single-thread RX):
| CPU | Packets/sec |
|---|---|
| EPYC 9755 | 2.84M |
| EPYC 9554 | 2.51M |
| Xeon 8592+ | 2.39M |
| Xeon 8480+ | 1.98M |
EPYC's advantage in packet rate comes from improved memory bandwidth and the way it handles cross-NUMA-node packet processing — particularly relevant for high-throughput proxies and CDN edge nodes.
Where Intel still wins: AVX-512
Intel still has a clear advantage in heavy AVX-512 workloads — specifically, video encoding using x265, certain machine learning inference paths, and some scientific simulations.
x265 4K HEVC encoding (single 30-second clip, slow preset):
| CPU | Encoding time |
|---|---|
| Xeon 8480+ (AVX-512) | 184 sec |
| Xeon 8592+ (AVX-512) | 176 sec |
| EPYC 9755 (AVX-512 via emulation) | 219 sec |
| EPYC 9554 | 241 sec |
EPYC 9004/9005 do support AVX-512 via "double-pumping" two AVX2 units, which works but isn't quite as fast as Intel's native implementation. For workloads that are heavily AVX-512-optimized and where encoding throughput is the primary metric, Intel still wins. We acknowledge this and recommend customers with such workloads choose our Dedicated Server tier with Xeon configuration available on request.
Power efficiency and TCO
Power efficiency matters because it directly affects what we can charge. EPYC's lower thermal design power (TDP) per performance unit means we can pack more compute into the same datacenter rack with the same cooling envelope.
Performance per watt (Geekbench 6 multi-core / Watt at full load):
| CPU | Score/Watt |
|---|---|
| EPYC 9554 (360W TDP) | 140 |
| EPYC 9755 (500W TDP) | 132 |
| EPYC 9654 (360W TDP) | 128 |
| Xeon 8592+ (350W TDP) | 97 |
| Xeon 8480+ (350W TDP) | 83 |
EPYC's 30-50% lead in performance per watt is what allows us to offer competitive pricing. Cooling and power are 35-40% of our datacenter operating costs, and EPYC's efficiency advantage flows directly into customer pricing.
Why we standardized on EPYC
For our customer mix — predominantly web applications, databases, and general-purpose compute — EPYC wins on the metrics that matter:
- 14% advantage in single-thread (where most apps spend their time)
- 50%+ advantage in transactions per second on cache-sensitive workloads
- 30-50% better performance per watt (which we pass through to pricing)
- Higher absolute core counts for parallel workloads
The exception is heavy AVX-512 workloads, where we recommend Intel-based dedicated servers (available on request). For everything else — and that's >95% of customer demand — EPYC is the better choice in 2026.