"Nvidia's DGX chips are in high demand in industry, though, and it's more likely that Micro Center's one-Spark limit is to prevent businesses scooping them up by the rack-load to run AI applications in their data centers. The DGX Spark contains an Nvidia GB10 Grace Blackwell chip, 128GB of unified system memory, a ConnectX-7 smart NIC for connecting two Spark's in parallel, and up to 4TB of storage in a package just 150mm (about 6 inches) square."
"It consumes 240W of electrical power and delivers 1 petaflop of performance at FP4 precision - that's one million billion floating point operations with four-bit precision per second. In comparison, Nvidia said, its original DGX-1 supercomputer based on its Pascal chip architecture and launched in 2016 delivered 170 teraflops (170,000 billion operations per second) at FP16 precision, but cost $129,000 and consumed 3,200W. It also weighed 60kg, compared to the Spark's 1.2kg or 2.65 pounds."
Nvidia's DGX Spark meets strong industrial demand for compact, efficient AI compute. Retailers impose one-unit limits to deter businesses acquiring racks of Sparks for data-center AI use. The Spark integrates an Nvidia GB10 Grace Blackwell chip, 128GB of unified system memory, a ConnectX-7 smart NIC for linking two Sparks, and up to 4TB of storage in a 150mm-square enclosure. The device consumes 240W and delivers 1 petaflop at FP4 precision. The original DGX-1 delivered 170 teraflops at FP16, cost $129,000, consumed 3,200W, and weighed 60kg versus the Spark's 1.2kg.
Read at Computerworld
Unable to calculate read time
Collection
[
|
...
]