AgentX Benchmark GPU: MI355X vs NVIDIA Agentic Coding (2026)

AgentX's early AMD MI355X results beat NVIDIA's best submissions on two separate leaderboard runs, then NVIDIA closed most of that gap through software alone on August 21, 2026, days before the benchmark's own publication date, when upstream vLLM optimizations from Inferact and NVIDIA flipped the cost-efficiency numbers on GLM 5.3. Both of those facts matter more than either vendor's headline number, and neither chip in that fight is what most teams can actually rent for a coding-agent fleet today.

TL;DR: MI355X vs NVIDIA on the AgentX Agentic Coding Benchmark

• Spheron's on-demand catalog listed NVIDIA H100, H200, B200, and B300 plus AMD MI300X across 5+ aggregated providers as of 17 Sep 2026; MI355X, the chip AgentX actually tested, isn't in it. Compare live GPU pricing.

What AgentX Measures That SWE-bench and GAIA Don't

SWE-bench, GAIA, Terminal-Bench, and OSWorld all answer the same question: did the agent get the task right? They score correctness, not cost or speed, which is why the harnesses we've covered for benchmarking agents on GPU cloud provision rollout workers against a fixed pass rate and move on. AgentX asks a different question entirely: assuming the model is already good enough, what does it cost in tokens per second and tokens per dollar to serve that model under the traffic pattern a real coding agent actually generates?

That distinction is not cosmetic. A fixed-length benchmark feeds a GPU a clean 2K-token prompt and measures how fast it comes back with 500 tokens. A coding agent session looks nothing like that: it opens a repo, reads a few hundred thousand tokens of context, makes a small edit, re-reads the updated state, and repeats that cycle dozens of times, with the KV cache growing the whole way. SemiAnalysis's own framing is blunt about why this matters: "In the past most measured performance based on fixed sequence length prefill and decode workloads, but this is an inaccurate way to measure workloads." A GPU that looks fast on a synthetic benchmark can still choke on cache pressure the moment it has to hold a million tokens of session history, and nothing in SWE-bench or GAIA's scoring would ever surface that failure.

Inside the AgentX Benchmark GPU Test: Methodology, Traces, and Hardware

AgentX is built from 393 anonymized Claude Code session traces, replayed through SemiAnalysis's AIPerf load generator rather than synthesized from scratch. A median AgentX session runs about 43 turns and pushes roughly 1 million tokens of context traffic over its life, with a median turn carrying 142,000 input tokens against just 444 output tokens, which is the whole shape of agentic serving in one ratio: almost all of the compute cost sits in re-reading context, not in generating the response.

The input sequence length distribution gets long fast. On DeepSeek V4 inside AgentX, the p50 input length is 88,000 tokens, p90 climbs to 272,000, p95 to 404,000, and p99 reaches 675,000 tokens. That tail is what makes AgentX a cache-pressure test as much as a throughput test. Under 384 concurrent traces on an NVIDIA B300 node, AgentX measured a 91% HBM cache hit rate against a 1.36% DRAM offload rate, holding roughly a 43-million-token working set in memory at once. The hardware AgentX benchmarks spans AMD MI355X against NVIDIA's current generation: B200, B300, GB300 NVL72, and H200.

The benchmark hasn't stayed an academic exercise. SemiAnalysis reports it has already driven more than 70 upstream pull requests across vLLM, SGLang, TensorRT-LLM, ATOM, AITER, Dynamo, LMCache, and Mooncake, as engineering teams on both sides tune their serving stacks against the actual traffic pattern instead of a synthetic one. In SemiAnalysis's own words, "AgentX made those costs large enough to drive generic upstream changes across vLLM, SGLang, TensorRT-LLM, ATOM, AITER, Dynamo, and LMCache." For background on one of the kernel libraries behind that AMD-side tuning work, see our breakdown of HipKittens as a shipped AITER backend.

The Results: MI355X vs NVIDIA on Agentic Coding Sessions

TileRT's 469 tok/s/user Leaderboard Run

The number that matters more than the headline, though, is how that throughput holds up as context grows. As input length scales from 1,000 to 1,000,000 tokens on the same TileRT/MI355X run, single-user generation speed drops from 648 tok/s down to 425 tok/s, a GPU retaining roughly two-thirds of its short-context speed at the far end of a coding agent's realistic context window. That decay curve is exactly the kind of number a fixed-length benchmark would never produce, because it never asks the GPU to hold a million tokens of session state in the first place.

MoRI UMBP's Tokens-Per-Dollar TCO Claim

AMD's second AgentX claim comes from a different angle: not raw speed, but total cost of ownership measured in tokens served per dollar spent. Using MoRI for request disaggregation and MoRI UMBP as a unified KV cache pool spanning the cluster, AMD reports MI355X delivering 69 million total tokens per $1 of TCO on AgentX running DeepSeek-V4-Pro-0813. The NVIDIA comparison point, B200 running Dynamo with SGLang, delivered 46 million tokens per $1, putting MI355X about 1.5x ahead on this specific tokens-per-dollar metric.

That advantage is real but has an expiration date attached, which is the whole story of the next section.

Where NVIDIA Pulled Ahead After August 21

The MoRI UMBP tokens-per-dollar matching held until August 21, 2026, when vLLM optimizations contributed by Inferact and NVIDIA closed the gap and then reversed it. At 150 tok/s/user of interactivity on GLM 5.3, the same workload AMD had been winning, NVIDIA's updated stack showed roughly 5x better cost efficiency than AMD's prior position. A software release note moved the leaderboard more than either vendor's silicon did in the weeks before it.

That flip is also not absolute. SemiAnalysis's Kimi K3 results, a 2.8 trillion-parameter model, show AMD's MI355X running its ATOM inference engine beating NVIDIA's GB300 NVL72 running vLLM specifically in the 40-to-60-second end-to-end latency band: "On the part of the curve between 40 and 60 seconds of end to end latency, MI355X ATOM beats even GB300 NVL72 vLLM on performance per dollar." Outside that window, vLLM on NVIDIA hardware remains the more relevant comparison for most teams running an upstream open-source stack rather than AMD's custom ATOM engine, and the same writeup notes NVIDIA's Hopper generation specifically struggles on this workload because of untuned kernel shapes, not a hardware limit. For a kernel-level look at that gap on currently rentable hardware, see our MI355X vs H200 attention kernel benchmark, which measures the FlashAttention-3 comparison AgentX doesn't isolate on its own.

AMD MI300X vs NVIDIA H100: The Established Pairing Behind This Fight

Every page currently ranking for AMD MI300X vs NVIDIA H100 is a GEMM throughput comparison, a memory-and-cost teardown, or a training benchmark. None of them mention AgentX, because AgentX doesn't test MI300X or H100 at all. It tests MI355X against B200, B300, GB300 NVL72, and H200, the generation above the pairing most teams can actually rent on a general-purpose GPU cloud right now.

That gap is the real reason this pairing still matters. AMD MI300X has 192GB of HBM3 per GPU against 80GB on NVIDIA H100 SXM5, so a 70B FP16 model fits on one MI300X but not on one H100, the same memory-headroom advantage that lets MI355X hold a million-token KV cache longer than you'd expect. If AgentX's finding generalizes, that it's memory capacity and kernel tuning rather than raw FLOPs that decide agentic serving cost, then MI300X's extra HBM is the one piece of that story a team can test on hardware available today, even without access to MI355X.

The caveat runs the other way too. AgentX's Kimi K3 writeup found NVIDIA's own Hopper generation, the generation H100 belongs to, underperforming not from a hardware ceiling but from kernel shapes nobody had tuned yet for this traffic pattern. H100 and H200 have years more upstream vLLM and SGLang optimization behind them than MI355X does, and that maturity gap is a large part of why NVIDIA's cost-efficiency numbers improved so fast after August 21. Renting H100 today buys you that maturity; renting MI300X buys you memory headroom at a real cost-per-token tradeoff in the middle of the batch-size range. For the training side of that same pairing, including a case where a 1,024-GPU AMD cluster trained a frontier model without CUDA at all, see our guide to training LLMs on AMD GPU with ROCm.

Cost Efficiency: Tokens Per Dollar vs Dollars Per Hour

AgentX's tokens-per-dollar metric and a rental catalog's dollars-per-hour rate are measuring two different things, and conflating them is how a cheaper hourly rate turns into a more expensive fleet. A GPU renting for less per hour but serving fewer tokens per second at your actual context length can cost more per million tokens than a pricier GPU that holds its throughput under cache pressure. That's exactly what the TileRT context-scaling numbers show: a GPU's short-context speed tells you little about what it'll cost once a session's context grows toward AgentX's measured p99 of 675,000 input tokens.

Billing mode adds a second layer most pricing comparisons skip. Our pricing breakdown of MI300X and MI355X on-demand versus spot rates found that a lot of the lowest advertised AMD rates on public trackers are preemptible, not on-demand, which matters for a coding-agent fleet that can't tolerate a mid-session eviction. On Spheron, AMD MI300X rents on-demand at $3.59/hr per GPU, against $2.64/hr for NVIDIA H100 SXM5 and $9.75/hr for B200, as of 17 Sep 2026. None of those figures include MI355X, which isn't in the catalog, so a direct dollars-per-hour comparison against AgentX's own hardware isn't possible on Spheron today, only the dollars-per-hour comparison against the established MI300X/H100/B200 pairing is.

Pricing fluctuates based on GPU availability. Spheron rates above are live as of 17 Sep 2026; other providers reflect their most recent published rates and may have changed. Check current MI300X availability → for live rates.

What This Means for Which GPU You Rent for a Coding-Agent Fleet

If you're sizing a coding-agent fleet this quarter, you're choosing among the GPUs AgentX didn't test, not the ones it did. That changes the decision from "which chip won the benchmark" to "which of the rentable options behaves most like the one that did."

A few practical reads from AgentX's findings, applied to what's actually on a rental catalog today:

• Long-context, memory-bound sessions favor more HBM per GPU. If your agent regularly carries six-figure token contexts, MI300X's 192GB headroom is the closest rentable proxy to MI355X's demonstrated cache-hit advantage, at a real cost-per-token tradeoff against H100 in the middle of the batch-size range.
• Teams on a stock vLLM or SGLang stack should default to NVIDIA. AgentX's own Kimi K3 result says the upstream-stack comparison that matters for most teams is vLLM on NVIDIA hardware, and NVIDIA's post-August 21 optimizations only widen that lead on standard tooling.
• Teams with kernel engineering budget can look at AMD's custom engines. TileRT and ATOM only post their leading numbers when paired with AMD-specific tuning most teams won't replicate themselves; treat those results as a ceiling, not a default.
• Check the system-level picture before committing to either vendor. AgentX is one benchmark measuring one workload shape; cross-reference it against broader numbers like our MLPerf Inference v6.0 results breakdown and our ROCm vs CUDA cost and gap analysis before locking in a fleet.

The honest summary is that AgentX settled a fight between two chips most teams can't rent yet, and the chips they can rent today, H100, H200, B200, B300, and MI300X, inherited a software-maturity gap rather than a hardware verdict. That gap narrows every time a vLLM release lands, which is the one thing both AMD's and NVIDIA's AgentX numbers agree on.

Sizing a coding-agent fleet around AgentX's findings means testing on the hardware you can actually rent this week, not the hardware that topped a leaderboard. Spheron's on-demand catalog, aggregated across 5+ providers, is billed per minute; check the live pricing page for which of H100, H200, B200, B300, and MI300X currently have on-demand capacity, since tiers shift with marketplace supply.

Compare GPU rental rates on Spheron →