DDR5/PCIe5 is 15-20% faster for LLM pre-training, but DDR4 wins on cost per dollar
Any-Winter-4079 · reddit · 2026-10-02
Benchmarking two Vast.ai workstations (EPYC 7352 DDR4/PCIe4 vs 9975WX DDR5/PCIe5), the author found DDR5 gives 15-20% faster LLM pre-training at equal GPU count. But with current RAM prices, the cost of 256GB DDR5-6400 buys an extra RTX PRO 6000, making a dual-GPU DDR4 setup 50% higher throughput per dollar. Also covers motherboard availability, BIOS compatibility with newer GPUs, DIMM/channel trade-offs, and why inference may favor DDR5 more.
More from Infra
- Benchmark Backs Chip Startup Tendrils Compute in Round That Could Top $1B Valuation — Sethwinterroth · 2026-10-02
- Modal Runtime ships multi-node clusters, VM sandboxes and sticky sessions — graceisford · 2026-10-02
- How to run local T2V/I2V/V2V video AI on an RTX 5070 with only 12GB VRAM — AirSea2062 · 2026-10-02
- Lightmatter CEO explains how spare lasers and optical switching route around failures in AI systems — BenBajarin · 2026-10-02
- Nebius Outperforms CoreWeave by 8X Despite 5X Less Revenue: What's Going On — Beth_Kindig · 2026-10-02
- Pedro Domingos: data centers lower electricity rates by paying the same overhead as residents — pmddomingos · 2026-10-02