400G
Line Rate
<10µs
P99 Latency
0%
Packet Loss
100%
Isolation
Throughput
Bandwidth Performance
DPU throughput varies by packet size and processing complexity. Small packets stress the packet rate, while large packets stress bandwidth.
📊 Throughput by Packet Size
Full 400 Gbps line rate achieved with jumbo frames (MTU 9000)
Latency
End-to-End Latency Breakdown
Every microsecond counts in AI workloads. Understanding where latency comes from helps optimize the entire path.
⏱️ Latency Components (Total: 8.5µs)
Wire
Parse
Policy
QoS
TX
Wire: 1µs
Parse: 1.5µs
Policy: 3µs
QoS: 2µs
Transmit: 1µs
Benchmarks
Real-World Performance
Lab benchmarks vs production reality. The numbers that matter for actual AI training workloads.
| Metric | Spec | Measured | Status |
|---|---|---|---|
| Max Throughput | 400 Gbps | 398 Gbps | ✓ PASS |
| Packet Rate | 595 Mpps | 520 Mpps | ⚠ 87% |
| P99 Latency | <10µs | 8.5µs | ✓ PASS |
| Jitter | <1µs | 0.8µs | ✓ PASS |
| Packet Loss | 0% | 0.001% | ⚠ Under Load |
| Isolation Breach | 0 | 0 | ✓ SECURE |
Live Metrics
System Utilization
Real-time view of DPU resource utilization during typical AI training workload.
📈 Resource Utilization
45%
ARM CPU
8x A78 Cores
78%
Accelerators
Crypto + RegEx
62%
Memory
16GB DDR5
Traffic Analysis
Latency Heatmap
24-hour view of latency distribution across all tenant traffic. Each cell represents 2 hours × one latency bucket.
🗓️ Latency Distribution Over Time
00:00
04:00
08:00
12:00
16:00
20:00
24:00
<5µs
>15µs
Comparison
DPU vs Traditional
Side-by-side comparison of DPU-based isolation versus traditional software approaches.
🔷 BlueField-3 DPU
Throughput
400 Gbps
Latency (P99)
8.5µs
CPU Overhead
0%
Isolation
Hardware
💻 OVS + DPDK
Throughput
100 Gbps
Latency (P99)
150µs
CPU Overhead
8 cores
Isolation
Software