Is Haswell to Skylake actually noticeable?

Side-by-side Plex transcode test: Haswell E3-1230 v3 vs Skylake E3-1230 v5, both 4c/8t, both 32GB DDR3/DDR4 respectively. Same media file (H264 1080p to 720p 4Mbps). Haswell: 4.2x realtime. Skylake: 4.7x realtime. That's 12% difference. For that money you get a Haswell box with 64GB RAM instead of the Skylake with 32GB. RAM matters more for Plex. My specs:

- Haswell: OVHcloud auction, €18/mo, 2x4TB, 32GB
- Skylake: Hetzner auction (https://www.hetzner.com/sb), €22/mo, 2x4TB, 32GB
- Both running Debian 12, ZFS mirror, Docker

IPC doesn't transcode. Quick Sync generation does, but both have it. Save money, buy RAM.

SNAPPED UP a Haswell box from Hostinger last month $14/YR ARE YOU KIDDING ok it was $14/mo but STILL! 32GB RAM! Transcodes 4 streams no sweat! Node shrink is for people with too much money!

Price history tracking:
- Haswell auction servers: $12-18/mo steady since 2024
- Skylake: $20-28/mo and DROPPING
- my prediction: Skylake hits $16 by Q3

Running to get more RAM now

This thread is about CPUs but everyone will ignore that RAM is the actual bottleneck for Plex. For what it's worth, I have observed transcode failures on 16GB systems that disappeared at 32GB with identical CPUs. The transcode buffer lives in RAM and when ZFS ARC also wants that RAM, something gives.

SPF/DKIM/DMARC analogy: you can have the best mail server (CPU) but if your DNS records (RAM) are undersized, delivery fails. Same principle.

Your 12% is within test variance. I have run this test myself across three generations and observed 8-15% swings depending on kernel version and whether spectre/meltdown mitigations are active. The actual answer is not in the silicon.

Peer better

HOST: haswell-test Loss% Snt Last Avg Best Wrst StDev
1. ??? 100.0 100 0.0 0.0 0.0 0.0 0.0

No MTR to share. Both boxes same DC, same AS. Latency irrelevant.

| gen | tdp | passmark | transcode x | $/mo |
|-----|-----|----------|-------------|------|
| haswell | 80W | 6.2E3 | 4.2 | 18 |
| skylake | 80W | 7.1E3 | 4.7 | 22 |
| kabylake | 91W | 7.8E3 | 5.1 | 28 |

Power draw observed identical. 12% performance for 22% cost increase. Peer better on price/perf.

I observed that your test methodology does not control for memory channel population[1]. DDR4-2133 in single-channel configuration exhibits 1.6E1 percent bandwidth reduction versus dual-channel, which may confound your CPU-generation comparison. RFC 2544 throughput testing principles suggest isolating variables.

Additionally, I observed no mention of whether Spectre v2 mitigations were enabled on both hosts. The retpoline overhead is not uniform across generations; Skylake incurs 1.1E1 to 1.5E1 percent penalty in syscall-heavy workloads[2]. Your transcode workload may not be syscall-heavy, but this is not stated.

[1] JEDEC standard JESD79-4, section 4.20.
[2] https://www.kernel.org/doc/html/latest/admin-guide/hw-vuln/spectre.html