NVIDIA Tesla V100 DGXS 16 GB

Benchmark Performance

The NVIDIA Tesla V100 DGXS 16 GB occupies a specific and narrow performance tier. With an average benchmark score of zero and a percentile rank of exactly 50 against all GPUs, the data positions this card as a midpoint reference point rather than a top-tier performer or a laggard. This percentile placement is unusual—it suggests the card's compute-oriented design does not translate into the gaming or general-purpose workloads that typically populate benchmark databases.

The FP32 throughput of 16.20 TFLOPS is the headline compute figure. This is a raw, theoretical peak, but it tells a clear story: the V100 is engineered for dense mathematical workloads, not rasterization efficiency. The FP16 rate of 32.40 TFLOPS (2:1 ratio) doubles that throughput, confirming that mixed-precision compute is the primary use case. The pixel rate of 202.5 GPixel/s and texture rate of 506.2 GTexel/s are modest by modern standards, reflecting that the 128 ROPs and 320 TMUs are secondary to the 5,120 shading units and 640 tensor cores.

Because the nearestRivals array is empty, no direct percentage deltas can be calculated against specific competitor cards. The data indicates that the V100's performance profile is defined by its compute ceiling, not by frame rates. Benchmark results indicate that in any scenario where FP32 or FP16 math is the bottleneck, this card will outperform typical consumer GPUs by a wide margin—but the absence of rival data means those margins are not quantified here. The 50th percentile ranking suggests that in mixed workload databases, half of all GPUs score better and half score worse, which is a surprisingly middling position for a compute accelerator.

Ray Tracing and Feature Set

The Tesla V100 DGXS 16 GB has no dedicated RT cores. The FACT PACK lists "rtCores": null, meaning hardware-accelerated ray tracing is entirely absent from this architecture. This is a critical limitation for any modern workload that relies on real-time ray-traced effects. The card does support DirectX 12 (12_1), OpenGL 4.6, and Vulkan 1.4, so it can run API-level ray tracing via compute shaders, but without dedicated hardware, performance in such scenarios will be severely constrained.

The tensor cores are the defining feature here. With 640 tensor cores on die, the V100 delivers its FP16 advantage through these specialized units, which are designed for matrix multiplication and deep learning inference/training. The 2:1 FP16 ratio (32.40 TFLOPS vs 16.20 TFLOPS) is directly attributable to these tensor cores. This makes the card a legitimate option for AI research, neural network training, and scientific computing that leverages mixed-precision arithmetic. However, for gaming or content creation, these tensor cores are largely dormant—they do not accelerate rasterization or traditional rendering pipelines.

The API support is comprehensive for its era, but the lack of RT cores means the feature set is compute-first. Vulkan 1.4 support is notably current, which is surprising for an end-of-life product, but this does not compensate for missing hardware ray tracing. The card also has no display outputs, confirming it is not intended for direct visual output—it is a compute accelerator that must be paired with a separate display adapter.

Memory Subsystem

The memory configuration is where the V100 shows its data-center pedigree. It packs 16 GB of HBM2 memory on a 4096-bit bus, yielding a bandwidth of 897.0 GB/s. This is an enormous memory pipeline, designed to feed the compute cores without starvation. The 4096-bit bus width is four to eight times wider than typical consumer GPUs, and the HBM2 type provides high density with low power consumption per bit.

For high-resolution workloads—think 4K or multi-GPU rendering, large dataset processing, or scientific simulation—this memory subsystem is exceptional. The 897.0 GB/s bandwidth means that data-intensive operations can stream through the card at rates that far exceed GDDR6-based alternatives. In compute tasks where the working set exceeds 8 GB, the 16 GB capacity prevents out-of-memory failures, though it is not the highest capacity available even at its launch.

The memory clock runs at 876 MHz base with 1752 Mbps effective data rate. While this clock speed is modest, the sheer bus width compensates: the effective bandwidth of 897.0 GB/s is the product of that 4096-bit interface, not high clock speeds. For gaming at 4K, this memory bandwidth is more than sufficient, but the card's compute-oriented shading units and lack of RT cores mean that memory headroom does not translate into frame rate advantages. The data shows a memory system built for throughput, not latency, which favors batched operations over interactive rendering.

Who Should Consider It

This card is not for gamers. The 50th percentile ranking and zero benchmark scores in gaming-oriented databases make that clear. Without RT cores and with a dual-slot, no-output design, it cannot serve as a primary display adapter. The data indicates that anyone seeking high-refresh-rate 1080p or 1440p gaming should look elsewhere—the V100's strengths lie entirely in compute.

The ideal user is a researcher, data scientist, or engineer running FP32 or FP16 workloads that fit within 16 GB. The 16.20 TFLOPS FP32 and 32.40 TFLOPS FP16 rates are substantial for deep learning training, where tensor cores accelerate matrix operations. For scientific simulations, climate modeling, or financial risk analysis that uses mixed-precision arithmetic, this card offers a legitimate compute platform. The 250 W TDP and 600 W suggested PSU mean it can slot into a standard workstation, provided the system has a PCIe 3.0 x16 slot and sufficient power delivery—though the "None" power connectors field indicates it draws power directly from the slot, which is unusual for a dual-slot card.

If your workload is rendering, video editing, or any GPU-accelerated application that benefits from ray tracing, this card is unsuitable. The absence of RT cores is a hard stop for modern real-time ray-traced effects. The 12 nm process and 21,100 million transistors on an 815 mm² die from TSMC indicate a large, power-hungry chip, but the 250 W TDP is manageable in a workstation context.

How It Compares

The nearestRivals array is empty, so no direct comparison to specific competitor cards can be made from the FACT PACK data. This absence is itself informative: the V100 does not sit neatly alongside consumer GPUs in benchmark databases. Its percentile rank of 50 places it at the median of all GPUs, which is a surprising position for a card with 16.20 TFLOPS FP32 compute. This suggests that the benchmark database weights gaming performance heavily, and the V100's lack of display outputs and RT cores drags its average score to zero, while its raw compute power keeps its percentile above the bottom half.

Without rival data, the practical comparison is against the card's own architecture. The predecessor "Tesla Pascal" and successor "Tesla Turing" bracket its generation, but no scores or deltas are provided for either. The data indicates that the V100 is a bridge between these two architectures—Volta introduced tensor cores that Turing would refine, but the V100 lacks the RT cores that Turing would add. In the absence of quantified rival deltas, the assessment must rely on the card's internal specifications: it is a compute-first accelerator that trades rendering features for raw math throughput, and its 50th percentile ranking reflects that trade-off in a database that likely favors rasterization performance.

FAQ

Q: Does the NVIDIA Tesla V100 DGXS 16 GB support hardware ray tracing?

A: No. The FACT PACK lists "rtCores": null, meaning there are no dedicated ray tracing cores on this card. It supports DirectX 12 (12_1), OpenGL 4.6, and Vulkan 1.4, but any ray tracing would run via compute shaders without hardware acceleration.

Q: What is the memory bandwidth of this card?

A: The memory bandwidth is 897.0 GB/s, achieved through 16 GB of HBM2 memory on a 4096-bit bus. The memory clock is 876 MHz with 1752 Mbps effective data rate.

Q: How many tensor cores does the V100 have, and what do they do?

A: The card has 640 tensor cores. These are specialized units for matrix multiplication and are the reason the FP16 rate is 32.40 TFLOPS, exactly double the FP32 rate of 16.20 TFLOPS.

Q: Can I use this card for gaming?

A: The data suggests no. It has no display outputs, no RT cores, and a 50th percentile ranking among all GPUs with an average benchmark score of zero. Its design is compute-oriented, not rasterization-oriented.

Q: What is the power requirement for this card?

A: The TDP is 250 W, and the suggested power supply is 600 W. The card uses no external power connectors, drawing power through the PCIe 3.0 x16 slot.

Q: What process node and architecture does the V100 use?

A: It uses TSMC's 12 nm process and the Volta architecture, with the GV100 chip containing 21,100 million transistors on an 815 mm² die. The production status is end-of-life, released on 2018-03-26.