Yotta NM1 Navi Mumbai
- Uptime Institute Tier IV certified
- Dual utility feeds from separate substations
CDNA 3 · OAM
The MI300X carries 192 GB of memory per GPU, 2.4 times an H100. Large models fit on fewer GPUs, and PyTorch and vLLM run on AMD’s ROCm stack.

Shipping Lead time, October 2026: Quoted per order · All GPU lead times
Good fit for

The hardware
8-GPU OAM platform, 1.5 TB of GPU memory. 192 GB HBM3 per GPU, up to 750 W per GPU, 8u typical of rack space.
Hosting
An 8-GPU MI300X server is in the same power class as an HGX H100 system. We place it against its full rated draw.
How managed hosting worksFor memory-bound inference, often yes: 192 GB per GPU versus 80 GB. Check that your stack runs on ROCm before you buy.
Yes, through ROCm. vLLM and most major inference frameworks support MI300X.
1.5 TB of HBM3 across eight GPUs.

Rack, power, cooling, network and remote hands for servers you own.
Learn more →
Lender-ready quotes and asset records for your bank or NBFC.
Learn more →
kW, amps and tons of cooling for every GPU class.
Learn more →
Tested, valued and sold, or traded up for newer GPUs.
Learn more →