SXM Meaning in NVIDIA H200 | Generated by AI
Question: What does SXM mean in NVIDIA H200?
Answer:
SXM = SXM form factor (Socket eXchange Module). In NVIDIA GPU products, it refers to the server module packaging format, not the GPU architecture.
For the NVIDIA H200 Tensor Core GPU, you usually see two physical versions:
-
H200 SXM
- Designed for large AI servers
- Uses NVIDIA’s SXM connector
- Connects through NVIDIA NVLink directly between GPUs
- Much higher bandwidth and power envelope
-
H200 PCIe
- Standard PCIe card (like a normal GPU expansion card)
- Easier to put into existing servers
- Lower inter-GPU communication bandwidth
The “SXM” module looks more like a dense accelerator board that plugs into a custom NVIDIA baseboard:
HGX H200 system
CPU
|
PCIe / NVLink
|
+----------------+
| H200 SXM |
| GPU module |
+----------------+
|
NVLink fabric
|
+----------------+
| H200 SXM |
| GPU module |
+----------------+
For AI training, SXM is the serious version because:
-
Higher power:
- H200 SXM: ~700W
- H200 PCIe: ~350W
-
More memory bandwidth:
- H200 SXM: ~4.8 TB/s HBM3e bandwidth
- PCIe version is more limited
-
Better GPU-to-GPU scaling:
- SXM uses NVLink/NVSwitch
- PCIe relies more on PCIe lanes
So when NVIDIA engineers say:
“SM120”
they mean the GPU compute architecture / streaming multiprocessor generation (Blackwell SM version).
When they say:
“H200 SXM”
they mean:
H200 = GPU product
SXM = physical module/package
Similar examples:
A100 SXM4
A100 PCIe
H100 SXM5
H100 PCIe
H200 SXM
H200 PCIe
For LLM training clusters (GPT-scale), almost all serious systems use SXM + NVLink + NVSwitch, because communication between GPUs becomes the bottleneck.