AMD outsells Intel in the datacenter space

Original Article ↗ Hacker News Discussion ↗

Context of the milestone

  • AMD datacenter revenue has, for the first time in the modern reporting era, edged out Intel’s.
  • Earlier eras (e.g., Opteron vs. Xeon) are hard to compare because “server/datacenter” wasn’t consistently broken out in earnings.

Why AMD’s datacenter share took so long

  • Enterprise buyers are conservative: they want long track records, rich management features, security support, and vendor certifications.
  • Early EPYC generations were strong on paper but lacked ecosystem maturity and certifications vs. Intel.
  • Long server refresh cycles (often 5–7 years) and existing contracts slowed migration.
  • Public clouds helped validate AMD by driving volume and forcing software support.

Performance, efficiency, and features

  • Many argue AMD has led in performance-per-watt since early EPYC (Rome and onward), especially versus Intel’s node troubles.
  • Others note Intel’s on-chip accelerators (QAT, AVX-512, AI/crypto/TLS primitives) can yield better efficiency on supported workloads.
  • Counterpoint: key Intel accelerators are often disabled on cheaper SKUs, reducing real-world impact and frustrating smaller buyers.
  • AMD is praised for simpler product segmentation (mostly quantitative differences: cores, clocks, cache).

Intel’s strategic missteps and current products

  • Thread highlights Intel “missing” mobile, bungling phone SoCs, and being late/weak in AI GPUs vs. Nvidia and AMD.
  • Historical anecdotes: killing StrongARM, failed Atom mobile push, over-betting on Itanium, long 10nm/7nm delays.
  • Lunar Lake laptop CPUs are seen as technically strong, but:
    • Fabbed by TSMC, undermining the “made by Intel” story for some.
    • Compete with Apple M-series and Snapdragon X; software compatibility still favors x86.
    • Integrated RAM packaging seen as a one-off; some view Intel’s retreat here as prioritizing its own margins over user value.

ARM, RISC‑V, and alternative ISAs

  • ARM is seen as eroding x86’s moat: Apple, Ampere, AWS Graviton, and Windows-on-ARM momentum.
  • Estimate in thread: ARM already ~20% of server CPU market (uncertain, not rigorously sourced).
  • Some discuss hybrid x86+RISC‑V or ISA-flexible designs; others note real estate and complexity constraints.

Enterprise buying behavior and TCO

  • Large buyers optimize for total cost of ownership, not “brand loyalty,” and will use a mix of vendors.
  • For small buyers paying list prices, Intel’s segmented SKUs and high-end pricing are described as poor value.
  • A few assert that, outside very specific use cases, recommending Intel in the datacenter is now hard to justify.

Market and stock perspectives

  • AMD’s stock volatility is debated:
    • One side: valuation already prices in big datacenter wins and strong AI GPU growth; any shortfall swings price.
    • Another: current price doesn’t reflect rising MI300 revenue; market may be over-fixated on Nvidia’s roadmap and hyperscalers’ in-house chips.
  • Being fabless is framed as both a limit on margins and a relief from massive fab capex and risk.

FPGAs and tooling

  • Side discussion: midrange FPGAs are either old-but-cheap or new-but-very-expensive with costly toolchains.
  • AMD/Xilinx tools (Vivado/Vitis) are praised vs. rivals; alternatives (Lattice, Microchip, Efinix) seen as rougher, especially on Linux.
  • Intel is spinning off Altera; low-cost segments are perceived as underserved.

Historical nostalgia

  • Several reminisce about 386/486-era AMD vs. Intel, NetWare servers, token ring, and early web/VoIP setups, contrasting simple, efficient stacks then with today’s more complex software.