I’ve noticed the RTX Spark delivers a memory bandwidth of 300 GB/s, while the DGX Spark memory bandwidth at 275 GB/s. This makes me wonder if the two products share identical memory configurations, what causes the 25 GB/s bandwidth discrepancy, and whether the DGX Spark is capable of reaching the 300 GB/s mark.
Edit: Misunderstood/misread a previous post and had a couple of things mixed up. Removing false info!
... and whether the DGX Spark is capable of reaching the 300 GB/s mark.
Probably not out of the box on our units.
It looks like the RTX Spark is using Micron memory (based on the report here Hands On With NVIDIA's First RTX Spark Laptops & PCs Ft. ASUS, Dell, HP, Microsoft, MSI, Lenovo), so I’m thinking it may be using 1-gama LPDDR5X Micron announced in June 2025 with ~20% or so more bandwith vs the previous node.
Maybe they’ll do a hardware refresh and the newer DGX units will get a bandwith uptick as a result?
Otherwise, the options are limited (memory overclocks, reballing, etc).
The 600GB/s (not 300GB/s) bandwith is for the C2C link (the interconnect between CPU and GPU chiplets), and that’d be separate from the memory bandwidth.
Whoops. Thanks for the catch! I had misread/misremembered the article I was thinking of.
But the general point was that since the memory is on the MediaTek CPU die, the GPU is accessing it over the C2C. And since the 600GB/s is the bidirectional number, they were chopping it in half for the best case unidirectional and calling that memory bandwidth. I couldn’t remember the logic on why they had supposedly done that, because, well, I invented it somehow ;).
No worries.
Until someone like the folks from ServeTheHome or Wendell from Level 1 Techs get their hands on one and do an actual deep-dive on the hardware, it’s all speculation based on the ‘crumbs’ in the marketing… and we all know how well that turned out for our DGX Sparks :)