BREAKING
DGX Spark: 64 Users on 38 Watts
0
concurrent users
0
+
tokens/sec
0
W
power draw
Why Batching Lifts Throughput
1
Single user: memory-bound
↓
2
Add continuous batching
↓
3
GPU utilization rises
↓
4
Aggregate throughput climbs
0
Arm CPU cores
0
GB
unified memory
0
PFLOP
FP4 AI perf
Dev Kit, Not Production Rig
Strengths
●
High aggregate throughput
●
Low 38W power draw
●
Desktop form factor, ~$4,000
Caveats
●
Development kit, not production
●
ARM64 ecosystem still maturing
●
Compatibility needs extra work
Multi-User AI Off the Rack
AI NEWS BLITZ
A desktop NVIDIA DGX Spark just served 64 users at once on only 38 watts.