I have two Sparks. One is MSI Edge Expert. The other one is the Gigabyte one. MSI Edge Expert shows 121.7 GiB available while Gigagbyte has 119.7 GiB
Is this normal? What do you guys get on NVIDIA boxes?
I want to get the Qwen 3.5 279B model working. And I think that one needs all the Vita one can get.
Also, Gigabyte, I had to resize the NVMe partition, out of box it had only made 625 GiB available for some reason. I had to grow the partition to be able to use the whole space.
So much weird bugs in these boxes.
I have checked after every update and no change so far. Probably needs a firmware update from them, I even contacted the black hole of their customer support a week age and no response. So basically can’t recommend gigabyte atom dgx, msi edge expert seems to be the optimal model so far. From what I understand seeing here on forums asus might be similar to gigabyte.
I can confirm it seems to be the same case with the HP ZGX Nano G1n:
zgx-spark:~$ sudo dmidecode -t system | grep -A3 "System Information" ; free -g
System Information
Manufacturer: HP
Product Name: HP ZGX Nano G1n AI Station
Version: A.7
total used free shared buff/cache available
Mem: 119 94 0 1 26 24
Swap: 15 0 15
vs the FE DGX Spark:
dgx-spark:~$ sudo dmidecode -t system | grep -A3 "System Information"; free -g
System Information
Manufacturer: NVIDIA
Product Name: NVIDIA_DGX_Spark
Version: A.7
total used free shared buff/cache available
Mem: 121 100 2 1 22 21
Swap: 15 0 15
(check the totalcolum, and ignore the used one, as I have some models running).
Both have been updated regularly, and no change in the total memory listed in the HP ZGX even after power-cycling.
It hasn’t impacted my workflows, but it’s still a little annoying.
I have one spark and one gigabyte atom. that missing 2GB, I think is the reason that I can’t run 397B int4-autoround with vllm. But I can run the model of gptq-int4 on sglang, so far my speed is 14 t/s, I just can’t increase the performance with triton backend. the quality of 397B is much better compare with 122B on a single spark.
sudo dmidecode -t system | grep -A3 “System Information” ; free -g Can someone with an Asus GX 10 run this command? I’m thinking of buying this model but I’d like to know its memory specifications. Here are the Dell and Lenovo options.
sudo dmidecode -t system | grep -A3 "System Information" ; free -g
System Information
Manufacturer: LENOVO
Product Name: 30KL0004FC
Version: ThinkStation PGX
total usado libre compartido búf/caché disponible
Mem: 121 117 1 0 3 4
Inter: 15 3 12
sudo dmidecode -t system | grep -A3 "System Information" ; free -g
System Information
Manufacturer: Dell Inc.
Product Name: Dell Pro Max with GB10 FCM1253
Version:
total usado libre compartido búf/caché disponible
Mem: 121 46 40 0 35 74
Inter: 15 0 15
Please understand that the interface is in Korean. :)
On my Asus GX 10, it shows as 121. What’s interesting is that I’m fairly certain it used to be 119 in the past. It’s possible that a recent BIOS update might have had an effect on this.
System Information
Manufacturer: ASUSTeK COMPUTER INC.
Product Name: GX10
Version: 5.36_GX10DGX
total used free shared buff/cache available
Mem: 121 114 0 0 7 7
Swap: 15 0 15
The 2 GiB difference between OEMs is firmware carveout — each manufacturer reserves a different amount for platform use.
The baseline thread has reproducible measurement tools for cold standalone runs — memory, temperature, power, and clocks — to compare how your unit behaves against other GB10 systems. Note the numbers reflect controlled baseline conditions, not peak inference throughput: GB10 Hardware Baseline — First Direct Measurements and Findings - #15 by parallelArchitect
I think you could update the firmware. I have one gigabyte, it used to be 119GB for so long and finally we get an update, it is now 121GB, be the same with DGX Spark.
Yeah, I know. I looked at gigabyte website and very confusing page with slapped instructions, no explanation, no change log. Just sloppy work. If someone updated gigabyte please share proven path.