Infrastructure Benchmarks
Real-world performance measurements from Fugoku infrastructure. No marketing numbers — just raw data from production workloads.
GPU Performance
Measured on bare-metal GPU servers with NVMe storage and 10 Gbps networking.
LLM Training Throughput
Llama 3 70B (INT4)
A100 80GB~28 tok/s
H100 80GB~52 tok/s
Metric: Tokens per second
Inference Latency (p95)
ResNet-50
A100 80GB23ms
H100 80GB18ms
Metric: Latency
Distributed Training Scaling
ResNet-50 across 8 GPUs
A100 80GB95%
H100 80GB97%
Metric: Scaling efficiency
Network Latency
Round-trip latency between Fugoku regions and major internet exchange points.
Lagos → Nairobi
~15ms
Lagos → Accra
~20ms
Lagos → Johannesburg
~35ms
Lagos → Frankfurt
~120ms
Frankfurt → London
~8ms
Frankfurt → New York
~65ms
Storage Performance
NVMe RAID storage benchmarks measured with fio on production hardware.
Sequential Read
NVMe RAID
5.2 GB/s
Sequential Write
NVMe RAID
4.8 GB/s
Random Read (4K)
NVMe RAID
850K IOPS
Random Write (4K)
NVMe RAID
720K IOPS
Need Custom Benchmarks?
We can run workload-specific benchmarks on your behalf. Contact us for custom performance testing.