Running Kimi K3 on MI355X at Better Performance per Dollar Than B300
A benchmark report from Wafer.ai claiming that the Kimi K3 model runs more cost‑effectively on AMD’s MI355X GPUs than on Nvidia’s B300s draws heavy scrutiny for opaque methods, unreplicable setups, and cherry‑picked cloud rental prices. Commenters argue that the performance tables actually show B300s winning on raw throughput, question the realism of quoted GPU hourly rates and total cost of ownership, and see the post as more marketing than rigorous engineering. A large side debate centers on whether releasing only model weights qualifies as “open source,” highlighting unresolved tensions over how traditional software freedom concepts apply to modern AI models.
Article quality and “AI slop” concerns
- Multiple commenters describe the post as “slop” or lazily written, though some note it’s more readable than typical AI-generated content.
- One person used an advanced model to critique the article and felt it exposed serious flaws; others argue that given the harsh prompt, any article would be torn apart.
- Specific complaints include poor editing, unclear prefill section, and weak visual design (e.g., bad text/background contrast).
Benchmark methodology and correctness
- Several commenters question whether the Kimi K3 setup is still coherent after low-level optimizations (e.g., head-count padding) and ask if correctness benchmarks were run.
- A wafer representative states that benchmarks such as tau/gpqa and coherency/thinking tests are required for hosting on OpenRouter.
- Some criticize use of a short 1024-token input context as outdated and unrepresentative.
- Others highlight that the B200 numbers are handicapped by cross-node all-reduce over RoCE, with missing details about better interconnects like Infiniband.
Performance-per-dollar and pricing assumptions
- Many challenge the claimed GPU-hour prices, especially the $2.50/hr MI355X vs $6/hr B300 framing.
- Critics say MI355X pricing is cherry-picked from a specific aggregator, may not reflect real, reliably available capacity, and ignores ownership TCO, power, and discounts.
- Some note that spot/on-demand pricing, scarcity, and commitment discounts make most “GPU price discourse” unreliable.
- There is skepticism that MI355X rentals at those rates could cover hardware capex over realistic lifetimes.
Wafer’s positioning and reputation
- Several accuse the piece of being functionally an AMD/Wafer advertisement with exaggerated or unfair comparisons and unreproducible benchmarks.
- Past issues like a quickly-abandoned token subscription product are cited as evidence of hype-driven behavior.
- Others argue that even if it’s promotional, optimizing inference economics is societally important given large spending on LLM serving.
“Open source” vs “open weights” models
- Long debate over whether releasing only weights qualifies as “open source.”
- One side: weights are the “source” for modifying the model; open licenses on weights + code satisfy software-freedom-like criteria.
- Other side: training data and process are the true “source”; weights are akin to binaries, so “open weights” is the honest term.
- Some distinguish “open science” (full reproducibility, including data and process) from merely open weights.