⚡ Top decode tok/s/card
Gemma 4 26B on 4× H100 SXM with FP8
1,700 tok/s/card
SHOWCASE
Showcase · what the data tells us
Insights auto-computed from the data corpus, refreshed every build
Every insight here is computed live from the corpus — adding a new case automatically refreshes these numbers.
💰 Lowest $/M tokens
Gemma 4 26B on 4× H100 SXM with FP8
$0.42 /M tokens
h100-sxm5 ×4 · assumes $2.5/h/card + $0.10/kWh + PUE 1.3
🎯 Most-mature software stack
AMD Instinct MI300X
1.50 / 1.00
measured/theoretical ratio across 1 cases. Reaching 150% of the theoretical roofline.
⚠️ Largest software-stack gap
昇腾 910C
0.00 / 1.00
0% of theoretical — large headroom for kernel/op-library tuning. Common pattern for Chinese silicon: stacks like CANN/MUSA/MindIE close this gap year over year.
📊 Most case coverage
NVIDIA H100 SXM5 80GB
3 cases
The data flywheel is spinning — this card has the most independent reproductions logged.
🚀 China stack catch-up opportunity
昇腾 910C
+138 pp headroom
Currently 0.00 vs overseas mean 1.38. Each +0.05 in efficiency ≈ +10% effective hardware throughput.
📈 Biggest gen-over-gen jump
NVIDIA L40S → NVIDIA B200 SXM 180GB
+515 % BF16
366 → 2250 TFLOPS BF16 · 2023 → 2024
🌍 Most portable model
DeepSeek R1
3 distinct hardware
Logged on 3 different accelerators — the most deployment-friendly frontier model in the corpus.
🏗 Largest scale-up domain
Google TPU v5p
8960 cards / domain
ICI · 4800 GB/s. Holds an entire frontier MoE in a single scale-up domain.