AMDFP8
369tok/s/request
GLM 5.3 · 8× MI350X · AgentX
~44% faster than comparable
NVIDIA Blackwell
THE SELF-IMPROVING AGENTIC COMPILER
Ultrafast inference.
By agentic compilers.
RSI agent swarms that discover high-performance just-in-time engines. Verified model quality.
01 / PERFORMANCEBest-in-class speed.
Verified model quality
Best-in-class speed.
Quality preserved.
Verified model qualityNVIDIANVFP4
477tok/s/request · p50
GLM 5.3 · 8× B200 · Megakernels
384 tok/s/request at p90
2–3× faster TTFT & end-to-end latency*
Benchmark notes & comparisons
AMD: GLM 5.3 FP8 on 8× MI350X, measured on AgentX, a coding-agent benchmark. The ~44% comparison is against the best comparable NVIDIA Blackwell result.
NVIDIA: GLM 5.3 NVFP4 on 8× B200; 477 tokens/sec/request at p50 and 384 at p90. Asari reports leading performance among 8-GPU NVIDIA and AMD configurations. *The 2–3× TTFT and end-to-end latency comparison is supplied by Asari; the comparison setup is available on request. FP8 and NVFP4 figures use different precisions and are not a like-for-like comparison.
Explore AgentX02 / THE AI STACK
Built to invent.
And reinvent.
Full-stack recursive self-improvement.
No human input to the discovery loop.
AUTONOMOUS DISCOVERY01
New engines.
In days.
Rust engines built from scratch.
Explore → test → verify → improve.
FULL-STACK RSI02
Every experiment.
A smarter system.
One integrated stack.
Knowledge that compounds.
STATEFUL SANDBOXES03
More experiments.
Same hardware.
Virtualized GPU clusters.
Flexible scheduling. Denser packing.
MANY CHIPS. ONE STACK.
NVIDIAAMDTPU v5 / v7Trainium Exploring partnership
03 / IN THE REAL WORLD