What happened
The InferenceX comparison page presents a head-to-head inference benchmark for Muse Spark Open 816B on GB300 NVL72, based on NVIDIA Blackwell, and MI355X, based on AMD CDNA 4. The displayed selection is FP4 precision and an “agentic-traces” scenario. The page says AgentX replays real coding-agent sessions rather than fixed-length prompts. In those sessions, context grows across turns and most of each request is served from cache rather than recomputed.
That setup makes the comparison a systems test as well as a chip-throughput test. The page identifies KV transfer between nodes, prefix-aware routing, and cache capacity as factors that move the performance curve alongside raw chip throughput. It contrasts this with fixed-sequence workloads, which it describes as a clean baseline for kernel and silicon performance; the two scenarios answer different questions about the same hardware.
At a target of 172 tokens per second per user, the page reports 34,845 tokens per second per chip for GB300 and 17,249 for MI355X. The displayed costs are $0.018 and $0.024 per million tokens, respectively. At 271 tokens per second per user, GB300 reaches 18,202 tokens per second per chip and MI355X 12,141; cost is $0.035 for GB300 and $0.034 for MI355X, making MI355X about 3% cheaper at that point. At 370 tokens per second per user, the reported throughput is 9,915 for GB300 and 4,927 for MI355X, with costs of $0.065 and $0.085 per million tokens.
The page describes its values as interpolated from real benchmark data and notes that the figures shown reflect the default agentic-traces and FP4 selection for this URL. Its chart and table update when the user changes sequence, precision, or model. The page labels the run date “Latest” and does not give a publication date in the supplied material. It also provides a performance-per-dollar view and controls for changing target interactivity, making the comparison a configurable benchmark rather than a single fixed ranking.
What the archive adds
This is a new hardware-comparison fact on a known Muse Spark Open serving story. NVIDIA details Muse Spark Open deployment on GB300 NVL72 reported deployment and throughput claims for GB300; this page adds a direct comparison with MI355X across workload-specific cost and throughput points. The Fireworks coverage and the model coverage concern serving routes and the model, not this chip comparison.
Trend signal
The comparison supports agent economics moves to completed work only indirectly: it measures cost and performance under agent traces, but does not report cost per verified completed task. Its useful signal is that inference economics depend on interactive operating point and serving behavior, not merely a per-token price or peak chip throughput.
Why now
The benchmark page is marked “Latest” and exposes selectable targets and configurations, but the supplied material does not establish a publication date or explain a specific timing trigger. The comparison is therefore best read as a current configurable reference point, not as evidence of a newly announced hardware or product change.
Why it matters
Teams choosing inference infrastructure need to satisfy a user-facing response target while keeping cost within budget. In this comparison, GB300 has markedly higher chip throughput at all three displayed agentic operating points, but it is not always cheaper: at 271 tokens per second per user, MI355X is about 3% lower in per-token cost. That difference matters if a fleet is tuned around a particular concurrency and latency envelope rather than maximum throughput.
The benchmark also highlights why coding-agent workloads can diverge from fixed prompts. Growing context and cache reuse change what the system must serve; routing and KV movement can influence results alongside accelerator specifications. A team can use the page to identify candidate operating points, then should reproduce its own workload, precision, model, and serving configuration before making a procurement or deployment decision. The supplied results are evidence about this selection, not a universal ranking of the two platforms.
Limits and uncertainty
The supplied page content does not identify the benchmark run’s calendar date, provide confidence intervals, or establish how representative the replayed sessions are of any particular production workload. The displayed results are specific to Muse Spark Open 816B, FP4, and the selected agentic-traces scenario; the page says that changing model, precision, or sequence changes the table and chart. Its fixed-sequence and agentic scenarios are explicitly different tests. The material also does not establish end-to-end application latency, full deployment cost, hardware availability, or performance for other models and serving stacks. The results should not be treated as an independent validation beyond the benchmark data and configurations shown.
Reader takeaway
GB300 delivers more chip throughput in the three displayed Muse Spark Open agentic operating points, but the cost comparison is not identical at every target. Compare at the interactivity level your users require, and treat cache behavior and serving architecture as part of the result—not as details that can be inferred from chip specifications alone.