Nvidia Unveils RTX 5880 Graphics Card with 14,080 CUDA Cores and 48GB VRAM

Original Article ↗ Hacker News Discussion ↗

Nvidia’s new RTX 5880 workstation GPU, essentially a cut‑down RTX 6000 Ada with 14,080 CUDA cores and 48GB of VRAM, is framed as a tool for design, simulation, and data science but widely seen as an export‑control‑compliant high‑end card for the Chinese market. Commenters weigh its technical merits, pricing, and subdued industrial design while focusing on how U.S. sanctions, shifting “cutlines,” and regulatory intent shape Nvidia’s product strategy and China’s push for semiconductor self‑sufficiency. The thread also highlights how CUDA’s ecosystem, VRAM capacity, and power efficiency underpin Nvidia’s dominance, and why rivals like Intel and AMD struggle to compete even when they could differentiate on memory or price.

Card specs & positioning

  • RTX 5880 Ada: 14,080 CUDA cores, 48GB GDDR6, ~69 TFLOPS FP32, 440 tensor cores, 110 RT cores, 285W TGP.
  • Raw performance seen as between RTX 4080 and 4090, but with double the VRAM of a 4090.
  • Some note footnotes: peak tensor TFLOPS rely on FP8, structured sparsity, and tiled GEMMs, so real performance is unclear without benchmarks.
  • Page markets it for rendering, data science, and simulation; AI/LLM not mentioned but widely assumed to be a major use.

Export controls & “sanctions-compliant” design

  • Many see this as essentially a cut-down RTX 6000 Ada tailored to meet US export rules for China.
  • Debate over US Commerce’s stance:
    • One side says firms are expected to honor the intent (deny China “frontier” AI training) and regulators will keep tightening rules if vendors sit exactly on the cutline.
    • Others argue the law should just ban what’s actually intended; changing rules to punish literal compliance is compared to arbitrary or “banana republic” behavior.
  • Analogies drawn to speed limits and anti–money laundering “structuring” rules to illustrate intent vs letter-of-law conflicts.

China, geopolitics & long-term impact

  • Some argue the export regime won’t meaningfully slow China, which is already driving toward tech self-sufficiency and can proxy imports or develop its own chips over time.
  • Others think central planning, corruption, and political risk will hamper China’s semiconductor catch-up.
  • Thread includes criticism of US foreign policy and “bully” image, versus defenses citing human-rights concerns and US-led global stability.
  • Possible future blowback via dedollarization and trade retaliation is mentioned, but effectiveness is debated.

VRAM capacity, Intel, and alternatives

  • Big subthread: could Intel win AI mindshare by shipping cheaper GPUs with 32–128GB VRAM?
  • Constraints cited: GDDR6 module capacities, pin count and routing complexity, bandwidth requirements, and demand patterns (games typically fine with 12–16GB at 4K).
  • Some say lots of VRAM without matching compute or bandwidth isn’t that valuable; others note huge pent-up demand from hobbyists who want large models and high-res diffusion.

CUDA, ROCm, Gaudi, and ecosystem lock-in

  • NVIDIA’s main moat is said to be CUDA and its mature tooling, docs, and ecosystem, not just raw hardware.
  • AMD’s ROCm is described as lagging enough that users accept NVIDIA’s pricing rather than switch.
  • Intel Gaudi2 is discussed: appears competitive on some LLM training/inference metrics and especially “performance per dollar” in vendor-backed benchmarks, but:
    • Access is mainly via Intel’s cloud; real perf-per-dollar once fully priced and powered is unclear.
    • Perf-per-watt and high-batch / multi-user serving seem worse than A100 in the cited tests.
    • Some see current Gaudi economics as unsustainable loss-leader; fear ecosystem investment could be stranded if Intel discontinues it.

Pricing, segmentation & availability

  • Expectations that RTX 5880 will be very expensive (multi‑thousand USD), targeting workstations and Chinese professional buyers constrained by export rules.
  • Discussion that NVIDIA carefully limits VRAM on consumer cards (e.g., 4080’s 16GB) to protect higher-end SKUs and workstation/AI products.
  • H100/Hopper cards are described as extremely expensive and scarce, largely snapped up by big cloud providers and rented out at very high hourly rates, driving AI compute costs.

Industrial design & form factor

  • Some praise the “understated” black-box, non-RGB aesthetic as professional and non-distracting.
  • Others find it bland compared with flashy gamer cards, but note businesses don’t value RGB and often prefer quiet, functional cooling.
  • Commenters complain about noisy fans on consumer GPUs and praise workstation-style designs focused on acoustics and practicality.

Performance, power & historical context

  • 285W TGP is viewed as impressive for its theoretical performance, especially versus prior-gen workstation cards.
  • Ada is praised as very power-efficient, though AI power bills in general are seen as “ridiculous.”
  • Several reflect on how quickly GPU compute has scaled (from 1 TFLOP being a milestone to ~70 TFLOPS on a single card) and lament that much of this gain feels “wasted” from an end-user perspective.