Kimi K3 on MI355X Beats B300 on Price‑Performance

Kimi K3 on MI355X shatters the B300 price‑performance myth

The Wafer.ai blog reports that running Kimi K3 on the MI355X delivers better performance per dollar than the B300. The headline alone forces engineers to question the default assumption that higher‑end cards always win the cost‑efficiency battle.

The post, titled Running Kimi K3 on MI355X at Better Performance per Dollar Than B300, does not list raw benchmark numbers, but the claim rests on a direct dollar‑for‑dollar comparison. Kimi K3 is a large language model; MI355X is a newer inference accelerator; B300 is an established competitor. The author’s confidence suggests that the MI355X’s silicon efficiencies outweigh the B300’s raw throughput. For teams budgeting cloud credits, the implication is immediate: a switch to MI355X‑based instances could stretch every dollar further.

Apple’s M1 Ultra vs Nvidia: a lesson in selective graphs

Apple’s recent launch of the M1 Ultra sparked a wave of performance claims. Macworld notes that Apple paired a 64‑core GPU chart with a quote that the chip “delivers faster performance than even the highest‑end PC GPU available while using 200 fewer watts of power.” The chart cut off the Nvidia RTX 3090 at roughly 320 watts, far below its 350‑watt peak.

When the graph is extended to the RTX 3090’s full 350‑watt envelope, Nvidia wins by a large margin. The RTX 3090 Ti, now on the market, pushes power draw to 450 watts and promises even higher raw performance. Apple’s comparison, while technically accurate on power efficiency, misleads on sheer compute power. The episode underscores how selective data presentation can distort the very cost‑performance calculus that Kimi K3’s claim seeks to clarify.

Cloud GPU pricing tables: transparency meets complexity

Sergey Karayev and Charles Frye maintain a community‑driven cloud GPU pricing table, updated on October 30 2023. The repository aggregates raw TFLOPs, benchmark scores, and hourly rates across major providers. It distinguishes between server‑based and serverless offerings, the latter defined by the original AWS serverless criteria: no server management, flexible scaling, high availability, and no idle capacity.

The table does not list every possible configuration, but it gives engineers a baseline for comparing on‑premise cards like the MI355X against rented instances of the B300 or Nvidia’s RTX line. The inclusion of benchmark data from AIME and Lambda Labs adds a performance dimension to pure price tags. When a claim such as Kimi K3 on MI355X beating B300 on dollars‑spent surfaces, the pricing table becomes the yardstick to verify whether the claim holds across cloud markets.

The shifting economics of AI acceleration

The convergence of three narratives—Kimi K3’s price‑performance edge, Apple’s selective GPU bragging, and the open‑source pricing tables—highlights a market in flux. Companies can no longer rely on headline‑grabbing performance numbers; they must factor power draw, hourly cost, and workload characteristics into a single decision matrix.

Engineers who prioritize raw throughput still gravitate toward Nvidia’s flagship cards, especially as the RTX 3090 Ti demonstrates that raw power continues to climb despite steep energy costs. Those who need balanced efficiency find Apple’s M1 Ultra compelling for mixed workloads, but they must accept that it lags behind Nvidia on pure performance. The Kimi K3 claim suggests that newer, less‑heralded accelerators can tip the balance for specific models, especially when the model’s architecture aligns with the chip’s strengths.

What to watch next

The next data point to track is the release of independent benchmarks for Kimi K3 on the MI355X, ideally measured against the B300 in identical cloud environments. Simultaneously, watch for updates to the cloud GPU pricing table as providers adjust rates in response to the latest silicon drops. Finally, monitor Apple’s next‑generation silicon announcements; if Apple continues to cherry‑pick comparison points, the industry will need clearer standards for reporting performance versus power versus cost.

Updates

• 2026-08-04 — The Best Cordless Vacuums (2026): My Brand-New Top Pick (source)