BREAKING
Gemma 4 runs 16-way on one DGX Spark
0
x
parallel runs
0
tok/s
per instance
0
tok/s
aggregate
0
B
total params
0
B
active params
0
K
context
Throughput scales with parallel runs
1 session
68
4 parallel
246
16 demo
300
8 parallel
403
The DGX Spark setup at a glance
DGX Spark
hardware
●
128GB unified memory
●
273GB/s bandwidth
●
About $3,999 to $4,699
Caveats
limits
●
FP4 kernel driver issues
●
Efficiency drops on huge prompts
●
Bandwidth questioned
Open weights, ready to run
AI NEWS BLITZ
Google just showed its open Gemma 4 model running sixteen instances at once on a single desktop machine.