NVIDIA Tesla V100 SXM3 32 GB

How It Compares

The NVIDIA Tesla V100 SXM3 32 GB occupies a unique position in the benchmark database, sitting at the 50th percentile among all GPUs tracked. This places it squarely in the middle of the distribution, but its profile is defined more by specialization than by raw ranking. With no nearest rivals listed in the dataset, the comparison must be drawn against the architectural lineage it represents. Its predecessor, Tesla Pascal, established the compute-accelerator template, while its successor, Tesla Turing, shifted focus toward ray tracing and mixed-precision workloads. The V100 SXM3, built on the Volta architecture, sits between these two generations as a pure compute powerhouse.

The data shows that this card is not designed for general-purpose rasterization. Its 50th percentile ranking reflects that reality, as the benchmark suite likely weights gaming and consumer-oriented tasks. In professional compute contexts, however, the V100 SXM3's specifications tell a different story. The 32 GB HBM2 memory configuration with a 4096-bit bus and 981.0 GB/s bandwidth is a massive allocation, exceeding what most contemporary accelerators offered at its release. The FP32 throughput of 16.35 TFLOPS and FP16 capability of 32.71 TFLOPS (2:1 ratio) indicate a device engineered for scientific simulation and AI training, not for frame rendering.

The process node is TSMC's 12 nm, housing 21,100 million transistors on an 815 mm² die. This yields a transistor density of 25.9M per mm², which was competitive for its era but now appears modest compared to later nodes. The chip, designated GV100, operates at a base clock of 1380 MHz and boosts to 1597 MHz. These clocks are moderate, suggesting the design prioritizes sustained throughput over peak frequency. The card is an SXM Module form factor, meaning it is intended for server integration rather than desktop use, and it has no display outputs at all.

Ray Tracing and Feature Set

The V100 SXM3 does not include dedicated ray tracing cores. The FACT PACK lists `rtCores` as null, confirming the absence of hardware-accelerated ray tracing. Instead, the card relies on 640 tensor cores, which are specifically designed for matrix operations used in deep learning and AI inference. This is a Volta-generation feature, and it predates the RTX-focused Turing architecture that followed. The tensor cores operate alongside 5120 shading units, 320 texture mapping units, and 128 raster output units.

The API support reflects the card's compute orientation. DirectX 12 (12_1) is supported, which is the highest feature level for that API, but the lack of display outputs means this is largely theoretical. OpenGL 4.6 and Vulkan 1.4 are also present, providing broad compatibility for compute and headless rendering workloads. The texture rate is 511.0 GTexel/s, and the pixel rate is 204.4 GPixel/s. These figures are respectable for compute tasks but secondary to the FP32 and FP16 throughput numbers.

The absence of ray tracing cores is noteworthy given that the successor, Tesla Turing, would introduce them. The V100 SXM3 is therefore best understood as a pre-ray-tracing accelerator, optimized for dense linear algebra rather than real-time graphics effects. The tensor cores, however, are a distinct advantage for workloads involving neural networks, as they provide a 2:1 FP16 ratio that effectively doubles throughput for mixed-precision training.

Power and Cooling

The Tesla V100 SXM3 has a thermal design power (TDP) of 250 W. This is a moderate figure for a compute accelerator of its capability, especially considering the 32 GB of HBM2 memory and the large die size. The suggested power supply unit is rated at 600 W, which provides ample headroom for the card's peak demands and the rest of the system. The card itself uses no power connectors, as it is an SXM Module designed to draw power through the server motherboard's socket interface.

The cooling solution is not specified in the data, but the SXM form factor implies a system-level cooling approach, typically involving active airflow from server chassis fans or liquid cooling loops. The 250 W TDP is within the range that a capable air cooler can manage, but the module's design assumes integration into a server with proper thermal management. The lack of a slot width specification beyond "SXM Module" indicates this is not a standard PCIe card with a bracket; it is a mezzanine-style module.

The bus interface is PCIe 3.0 x16, which is sufficient for data transfer to and from the host system. Memory bandwidth of 981.0 GB/s is the standout specification here, far exceeding what any PCIe connection can deliver, so the high-bandwidth HBM2 is used for on-device data movement rather than host communication. The production status is end-of-life, and the release date is 2018-03-26, making this a mature product with a well-understood power profile.

FAQ

Q: Does the Tesla V100 SXM3 support hardware ray tracing?

A: No. The FACT PACK lists no ray tracing cores, and the architecture is Volta, which predates the ray tracing capabilities introduced in the Tesla Turing successor.

Q: What is the memory configuration and bandwidth?

A: The card has 32 GB of HBM2 memory with a 4096-bit bus, providing a bandwidth of 981.0 GB/s. This is a high-bandwidth design suited for memory-intensive compute workloads.

Q: What power supply is recommended for this card?

A: The suggested PSU rating is 600 W. The card's TDP is 250 W, and it uses no power connectors because it is an SXM Module that draws power from the server motherboard.

Q: Can this card be used for display output?

A: No. The FACT PACK states "No outputs" for display outputs, making it a compute-only accelerator designed for server environments.

Q: What is the FP16 performance relative to FP32?

A: The FP16 throughput is 32.71 TFLOPS, which is exactly double the FP32 figure of 16.35 TFLOPS. This 2:1 ratio is enabled by the 640 tensor cores.

Q: Is this product still in production?

A: No. The production status is listed as "End-of-life," and the release date is 2018-03-26. Its successor is the Tesla Turing architecture.

Benchmark Performance

The benchmark data for the Tesla V100 SXM3 is sparse, with no entries in the `benchmarks` array and an average benchmark score of zero. The percentile rank of 50.0 indicates that, among all GPUs in the database, this card performs at the median level. However, this ranking must be interpreted with caution, as the database likely includes a wide range of consumer and professional cards, and the V100 SXM3 is not designed for the tasks most benchmarks measure.

The FP32 throughput of 16.35 TFLOPS is the primary compute metric, and it places the card firmly in the high-performance accelerator category for its generation. The FP16 figure of 32.71 TFLOPS doubles this capability, which is directly relevant to AI training and inference workloads that use mixed precision. The texture rate of 511.0 GTexel/s and pixel rate of 204.4 GPixel/s are secondary metrics, reflecting the card's ability to handle texture-heavy compute kernels if needed, but they are not the focus.

The memory bandwidth of 981.0 GB/s is a critical differentiator. Many accelerators from the same era offered less than half of this bandwidth, and the 32 GB capacity allows for larger datasets to reside on-device without host transfers. The bus width of 4096 bits is unusual, confirming that this is a purpose-built HBM2 design rather than a GDDR6-based solution. This bandwidth advantage directly translates to faster iteration times for large matrix operations, which are the core of deep learning.

The lack of nearest rivals in the dataset means there are no direct percentage deltas to cite. However, the architectural comparison to its predecessor, Tesla Pascal, and its successor, Tesla Turing, provides context. Tesla Pascal did not have tensor cores, so the V100 SXM3's 640 tensor cores represent a generational leap for AI workloads. Tesla Turing added ray tracing cores but maintained tensor cores, shifting the balance toward graphics. The V100 SXM3, therefore, holds a middle ground: it lacks the ray tracing of Turing but offers the full compute density of Volta.

The 50th percentile rank suggests that in a mixed benchmark suite, the card does not excel at rasterized gaming or consumer 3D rendering. Its strengths are in FP16 matrix math and high-bandwidth memory access, which are not typically captured in such suites. The zero average benchmark score further indicates that no standardized benchmarks were run or reported for this card, likely because it is a specialized server part.

In summary, the benchmark data shows a card that is neither a top performer in general GPU rankings nor a low-end part. Its 50th percentile placement is a reflection of its niche: it is a compute accelerator that sacrifices display capabilities and ray tracing for raw throughput and memory capacity. The FP16 2:1 ratio is the standout feature, enabling efficient AI training, while the 32 GB HBM2 ensures that large models fit in memory. The absence of nearest rivals and benchmark scores makes quantitative comparisons impossible, but the specification sheet alone indicates a serious compute instrument for its time.