Lovelace (2024) Cluster - GVM Service L40S/6000Ada

New GPU cluster features with Ada Lovelace (2024) architecture NVIDIA cards model L40S and RTX 6000Ada has been established, with a total of 208 GPU cards. The GPU-enabled Virtual Machine (GVM) service is similar to the Azure and AWS Cloud Instance services, except that UBDA provide a more prominent direct pass-through GPU card, which means that these GPU cards are NOT virtualized but 100% reserved and owned by the dedicated GVM users.

In 2025, UBDA has migrated the Lovelace (2024) cluster to the latest NVMe that backends read/write speeds of up 7,400/6,400 MB/s over dual 40 Gbps Ethernet networks. Additionally, each GVM can be embedded with direct pass-through local NVMe to further elevate the I/O speeds, enabling fast data caching for crucial modelling tasks. Users can thus process data from the GVMs’ local RAM to local NVMe (all within a single host server) without going through the Ethernet network, the fastest I/O among other VM cloud topologies.

Upon request, UBDA will provision GPU-enabled VMs for PolyU users (staff/students/external-guests), the maximum number of GPU cards in each VM is 8 at the moment. Users can RDP/SSH to their remote instance using their PolyU NetID and Net Password. They are configured as Linux Superuser and can therefore sudo install their own software/libraries, and highly customized their applications in GVM environment.

The following are illustrative configurations for selection:

Package	No. of vCPU Cores	No. of GPU Cards	vRAM Size (GiB)	OS-vDisk Size (GiB)	NFS-HDD Mount Drive Quota	No. of Local 3.84TiB NVMe	GPU Card Model
1	8	1	96	512	512 GiB to 100 TiB	-	NVIDIA L40S 48GB or RTX-6000Ada 48GB
2	16	2	192	512		-
3	56	2	384	1024		-
4	56	4	384	2048		1
5	112	6	448	2048		1
6	112	8	480	4096		1
7	32	-	192	512		-
8	56	-	384	1024		-
9	96	-	448	2048		1
10	112	-	480	4096		1

Total No. of GPU Card (NVIDIA Ada Lovelace Architecture)	208
Total No. of CPU Core (AMD EPYC 7004/9004 Series Processors)	4,632
Total No. of RAM (TiB, DDR5)	58.5
Total No. GPU Server	29
Total NFS-HDD Size (TiB, raw)	6,720
Total NFS-NVMe Size (TiB, raw)	1,475

Usually, the GVM bookings has to be renewed every 4-6 weeks, and based on the demand, UBDA reserves the right to re-allocate resources to others as needed. However, users can still raise a request for a longer usage term, such as 6-months project, by filling out a special request form or by email. Note that each PI group should claim NO more than 12 GPUs or 168 vCPU cores in total in their research for the fair use of other PolyU users.

For more information, please contact with our scientific officer Dillian Wong (dillian.wong@polyu.edu.hk)