KernelForge
Speedup vs SGLang
Attention decode · NVIDIA H200
Best measured result · 26 Oct 2026
MEASURED RESULTS
SGLang = 1.00×. Five Muse Spark workloads on NVIDIA H200, with speedup and 95% intervals for each complete task.
Attention decode
dilate_attention_decode
Attention during generation and verification.
vs SGLang
Correctness passed
Expert computation
dilate_moe_w4a16
Selected expert networks with 4-bit weights.
vs SGLang
Correctness passed
Decode layer
dilate_decode_layer
A complete transformer layer during generation.
vs SGLang
Correctness passed
BF16 projections
dilate_bf16_proj
Matrix projections and expert selection.
vs SGLang
Correctness passed
Prefill layer
dilate_prefill_layer
A complete layer while processing a prompt.
vs SGLang
Correctness passed
COMPUTE CAPABILITIES
Dilate’s research spans the core operations of modern AI models, from attention and expert networks to complete transformer layers. Faster compute creates room for more ambitious AI.
Tokens
Projections
Attention
Experts
Output
ATTENTION
Faster generation workloads
Attention compute for generation and verification, with our strongest measured workload result reaching 1.92× the SGLang baseline.
EXPERT NETWORKS
More speed for expert models
Compute for mixture-of-experts models, including workloads with 4-bit weights. Our measured expert workload reaches 1.85× the SGLang baseline.
TRANSFORMER LAYERS
Acceleration across the layer
Research covering complete decode and prefill layers, alongside BF16 projections. Measured improvements span all five workloads.
WHAT DILATE OFFERS
MODEL ACCELERATION
Faster compute for your models
GPU kernel expertise across attention, expert networks and transformer layers, focused on the workloads that drive model performance.
HARDWARE EFFICIENCY
More from the GPUs you have
Performance research that helps model teams explore more useful computation from available hardware, with gains grounded in measured workloads.
RESEARCH PARTNERSHIPS
Build with Dilate
Collaborate with Dilate through the Loopseed research programme to explore faster compute for your models and AI systems.
LOOPSEED RESEARCH
Fish is the model we are testing on as we bring our compute research into model execution. Through the Loopseed research programme, Dilate is advancing the foundations of faster AI.
Research for faster AI
Explore what Dilate’s compute research could enable for your models.
Discuss your model