Spec Card SPEC CARD
NVIDIA V100
Volta flagship · The first Tensor Core GPU: available in SXM2 (300W / NVLink 300GB/s) and PCIe (250W) form factors, with 16/32GB HBM2 of large memory capacity. With FP64 double-precision HPC plus AI training capability, it remains one of the most widely circulated models on the used data center GPU market (both SXM2 and PCIe variants: Tesla V100, SXM2/PCIe dual form factors)
Overview
NVIDIA Tesla V100 is a data center GPU based on the Volta architecture (GV100) and the first GPU product to feature Tensor Cores. Built on TSMC's 12nm FFN process, it integrates 5120 CUDA cores, 640 first-generation Tensor Cores, and 2560 FP64 cores, paired with 16GB or 32GB of HBM2 memory (CoWoS packaging, with ECC) on a 4096-bit memory interface. The V100 comes in two form factors: SXM2 and PCIe. The SXM2 version supports up to 6 NVLink 2.0 links with a total bidirectional bandwidth of 300GB/s, while the PCIe version uses a PCIe Gen3 x16 interface. Combining double-precision compute with AI training capability, the V100 has long served HPC science and deep learning workloads, and remains one of the most widely circulated models on the used data center GPU market.
Core Specs
| Architecture | Volta (GV100) |
| Memory | 16GB / 32GB HBM2 (CoWoS) with ECC, two versions |
| Memory Bandwidth | 900 GB/s (standard 16GB/32GB); V100S 32GB high-clock PCIe version at 1134 GB/s |
| Memory Bus Width | 4096bit |
| Compute | FP32 15.7 / FP64 7.8 / Tensor 125 TFLOPS (SXM2 at Boost clocks); PCIe version per official page: 14 / 7 / 112 |
| CUDA cores | 5120 (80 SM) / Tensor Cores 640 (first generation) / FP64 cores 2560 |
| Power | 300W (SXM2) / 250W (PCIe) |
| Interconnect | NVLink 2.0 300GB/s bidirectional (SXM2, up to 6 links) / PCIe Gen3 x16 32GB/s (no NVLink on PCIe version) |
| Launch Year | 2017 (announced at GTC on 2017-05-10, SXM2 shipped Q3 2017; 32GB version launched in 2018) |
| Process | TSMC 12nm FFN, 21.1B transistors, 815mm² |
| Form Factor | SXM2 module / PCIe dual-slot full-height full-length passive cooling |
| Highlights | First Tensor Core GPU / First data center GPU to break 100 TFLOPS of deep learning compute |
Specs verified against official documentation (
official benchmarks , updated 2026-10-10) · actual price subject to the seller's quote
Use Cases
- Scientific computing and engineering simulation: with its double-precision floating-point capability, it suits traditional HPC workloads such as computational fluid dynamics, molecular dynamics, and climate modeling
- Deep learning model training: first-generation Tensor Cores accelerate Mixed Precision training, while the large HBM2 memory capacity supports larger batch sizes and deeper networks
- Multi-GPU large-scale parallelism: the SXM2 form factor enables high-speed multi-GPU interconnect via NVLink 2.0, ideal for building DGX-class nodes or GPU clusters for distributed training
- Inference and R&D test environments: the 16GB version offers strong value on the used market and is commonly used for inference serving, algorithm validation, and teaching lab platforms
Same-Family Models
| Model | Memory | Power |
|---|
| NVIDIA T4 Compare | 16GB GDDR6 with ECC (enabled by default) | 70W (total board power, default) |
| NVIDIA A100 Compare | Dual versions: 40GB HBM2 / 80GB HBM2e (80GB added at SC20 on 2020-11-16), with ECC | SXM4 400W (same for 40GB/80GB); PCIe version 250W for 40GB / 300W for 80GB |
| NVIDIA H20 Compare | 96GB HBM3 (consistent across Flopper, VipraTech, and Tencent News sources) | 400W TDP (three sources: Flopper, VipraTech spec sheet, and Tencent News' "400W thermal design power"; VipraTech's body text also cites 350W, a conflicting figure that has been discarded) |
| NVIDIA GB300 Compare | 288GB HBM3e (per GPU package; 576GB for the superchip; HGX board-level 2.1TB / rack-level 20TB coexist) | TGP up to 1400W |
| NVIDIA A30 Compare | 24 GB HBM2 | 165 W |
| NVIDIA GeForce RTX 4090 Compare | 24GB GDDR6X | 450W TGP (315W average gaming, 850W recommended system power) |
Video Overview VIDEO
Procurement & Inspection Notes
- Clarify form factor and interconnect requirements: the SXM2 version must be paired with a server that supports an SXM2 carrier board (e.g., DGX-1 or HGX-1 platforms) and cannot be used in a standard chassis; the PCIe version is a full-height, full-length dual-slot passive-cooling card that relies on forced server airflow — before purchasing, verify chassis compatibility and cooling capacity, and budget the system's power headroom. The two form factors have different power ceilings, so plan server power and cooling accordingly
- Verify memory capacity and part numbers: the V100 comes in 16GB and 32GB configurations, plus variants like the V100S (with different bandwidth specs). When purchasing, confirm the exact model via the official part number — for example, PCIe 16GB is 900-2G500-0000, PCIe 32GB is 900-2G500-0010, SXM2 16GB is 900-2G503-0000, and SXM2 32GB is 900-2G503-0010 — to avoid version mix-ups
- Used-unit inspection essentials: high market circulation means wide variation in condition. Run actual stress tests to check for memory ECC errors and core stability, inspect the heatsink and thermal pads, and ask for proof of provenance; for SXM2 units, also confirm carrier board compatibility and NVLink bridge availability to avoid future expansion limits
FAQ
What's the difference between the V100 SXM2 and PCIe versions, and which should I choose?
The SXM2 version is a module requiring a dedicated carrier board and supports NVLink 2.0 multi-GPU interconnect (up to 6 links, 300GB/s total bidirectional bandwidth), suited for multi-GPU parallel training and HPC clusters; the PCIe version is a standard full-height full-length dual-slot passive-cooling card with better installation compatibility, suited for single-card or small-scale deployments. Without server room and carrier board infrastructure, the PCIe version is usually the way to go.
How to choose between the V100 16GB and 32GB versions?
Both share identical core specs; the difference is memory capacity. The 32GB version suits large models, large-batch training, or scientific computing workloads that need a big memory footprint; the 16GB version is cheaper on the used market and covers most mid-scale training and inference needs. Choose based on your actual model's peak memory usage, with some headroom to spare.
What should I watch out for when buying a used V100?
Focus on three things: first, form factor fit — the SXM2 version requires a dedicated server platform; second, hands-on testing — run a Full Load stress test, review memory ECC error logs, and confirm the cooling components are intact; third, verify the part number matches the actual configuration (PCIe/SXM2 and 16GB/32GB), and ensure the power connectors, chassis airflow, and cooling conditions meet the requirements of a passive-cooling card.