NVIDIA A100 SXM4
Expert rating8.6 / 10
- GPU Brand:
- NVIDIA
- VRAM:
- 80GB HBM2e
- Memory Bandwidth:
- 2039 GB/s
- TDP:
- 400W
Consumercarts is reader-supported. Learn more
Disclaimer: The price and specification of the given product may differ from the actual product. Please confirm on the retailer's website before buying.
Manufacturer: NVIDIA Last updated: 2026-09-14
NVIDIA A100 SXM4 Review
The NVIDIA A100 SXM4 (80GB) is the full-power, baseboard-mounted version of NVIDIA's Ampere datacenter GPU, and it was the dominant GPU behind large-scale AI training clusters from 2020 through the H100's rollout in 2023.
It shares its 6,912 CUDA cores with the A100 PCIe card but runs at a higher 400W power envelope with faster HBM2e (2039 GB/s vs 1935 GB/s) and a 600 GB/s NVLink fabric linking up to 8 GPUs on one HGX baseboard, which is what made it the backbone of most pre-Hopper large-model training infrastructure.
Overview
The SXM4 form factor mounts the A100 directly onto NVIDIA's HGX baseboard alongside up to 7 other GPUs, sharing power and a dense NVLink switch fabric rather than plugging into an individual PCIe slot.
Performance & Use Case
It remains in active service for multi-GPU AI training and inference clusters, HPC simulation, and data analytics workloads at organizations that built out Ampere-generation infrastructure and haven't yet migrated to Hopper or Blackwell.
Comparison to Alternatives
Against the NVIDIA A100 PCIe (nvidia-a100-pcie), the SXM4 module offers about 5% more memory bandwidth and a 600 GB/s NVLink fabric across up to 8 GPUs, versus a 600 GB/s two-GPU bridge on the PCIe card. Against the NVIDIA H100 SXM5 (nvidia-h100-sxm5), the H100 delivers roughly 3x the AI training throughput on the newer Hopper architecture with a faster NVLink fabric. Against the NVIDIA A800 (nvidia-a800), the export-restricted A800 SXM4 has identical compute specs but a NVLink bandwidth capped at 400 GB/s instead of 600 GB/s.
Our verdict
Reviewed by Burooj Alam. Rated by us, not by users.
Expert rating8.6 / 10
The GPU that trained a huge share of today's deployed AI models - still a capable multi-GPU training platform, though clearly surpassed by Hopper and Blackwell for new builds.
Best for
- Multi-GPU AI training clusters (existing HGX/DGX A100 systems)
- Large-scale AI training and inference at proven, mature scale
- HPC and scientific computing requiring high inter-GPU bandwidth
Not for
- Gaming or any graphics workload - no display output, no RT cores
- Single-GPU deployments - only sold as part of a multi-GPU HGX baseboard
- New cutting-edge LLM training builds - Hopper/Blackwell generations are substantially faster per GPU
Specifications
What these specs meanPlatform
| GPU Brand | NVIDIA |
|---|---|
| Board Partner | NVIDIA Reference - HGX A100 4-GPU or 8-GPU baseboard module, not a standalone card |
Identity and type
| Series | Ampere A-Series |
|---|---|
| Tier | Datacenter / AI Accelerator |
Performance
| VRAM | 80 GB |
|---|---|
| Memory Type | HBM2E |
| Boost Clock | 1410 MHz |
| Ray Tracing Support | No |
Physical & Power
| TDP | 400 W |
|---|---|
| Slot Width | N/A (SXM module, not a PCIe card) |
Silicon
| Architecture | Ampere (GA100) |
|---|---|
| Process Node | TSMC 7N |
| Shader Units | 6912 |
Memory
| Memory Bus Width | 5120 bit |
|---|---|
| Memory Bandwidth | 2039 GB/s |
Host interface
| PCIe Generation | N/A - SXM4 socket connects to the HGX baseboard, not a PCIe slot |
|---|
Market and workstation
| Launch Date | 2020-11 (80GB SXM4 variant; 40GB variant launched 2020-05) |
|---|
FAQs
What are the key specifications of the NVIDIA A100 SXM4 (80GB)?
Key specs for the NVIDIA A100 SXM4: VRAM 80 GB, Memory Bandwidth 2039 GB/s, Boost Clock 1410 MHz, TDP 400 W, PCIe Interface N/A - SXM4 socket connects to the HGX baseboard, not a PCIe slot, Memory Type HBM2E.
Who is the NVIDIA A100 SXM4 (80GB) best suited to?
Best suited to: Multi-GPU AI training clusters (existing HGX/DGX A100 systems); Large-scale AI training and inference at proven, mature scale; HPC and scientific computing requiring high inter-GPU bandwidth.
What are the drawbacks of the NVIDIA A100 SXM4 (80GB)?
Not the best fit for: Gaming or any graphics workload - no display output, no RT cores; Single-GPU deployments - only sold as part of a multi-GPU HGX baseboard; New cutting-edge LLM training builds - Hopper/Blackwell generations are substantially faster per GPU.