NVIDIA H100 SXM5 80 GB

Benchmark Performance

The NVIDIA H100 SXM5 80 GB is a compute-first accelerator built on the Hopper architecture, and its benchmark positioning reflects that specialization. With a percentile rank of 50 against all GPUs in the database, this card sits squarely in the mid-pack of the entire GPU landscape—but that figure is misleading without context. The H100 SXM5 is an end-of-life server part, and its 50th-percentile standing indicates that while it is not the fastest accelerator ever cataloged, it remains a formidable compute engine that outpaces most consumer and prosumer hardware in raw throughput metrics.

The FP32 performance of 66.91 TFLOPS is the headline number for general compute. This places the H100 SXM5 in a class where traditional rasterization-focused GPUs cannot compete. The shading units (16,896) and texture mapping units (528) are arranged to feed this compute throughput, and the texture rate of 1,045.4 GTexel/s confirms that the card can sustain heavy shader workloads without bottlenecking on texture fetch. The pixel rate of 47.52 GPixel/s, however, is comparatively modest—a direct result of having only 24 ROPs. This is a deliberate design choice: the H100 is not built to drive displays or rasterize frames; it is built to process data.

The FP16 throughput of 267.6 TFLOPS (4:1 ratio) is where the H100 SXM5 truly separates itself from general-purpose GPUs. This 4:1 ratio means the card can deliver quadruple the FP16 work compared to its FP32 rate, a feature that is essential for AI training and inference workloads that rely on mixed-precision arithmetic. The tensor cores—528 in total—are the engines behind this capability. When the data shows FP16 at 267.6 TFLOPS, that is the tensor-core-accelerated path, not the standard shader pipeline.

Because the nearestRivals field is empty in the fact pack, no direct percentage deltas against competing accelerators can be cited. However, the percentile rank of 50 against all GPUs provides a reference point: half of all GPUs in the database score higher on aggregate benchmarks, and half score lower. This is a surprisingly low standing for a 700 W server part, which suggests that the aggregate benchmark suite may weight gaming or consumer-oriented workloads more heavily than compute tasks. For the H100 SXM5, the relevant comparison is not against gaming cards but against other server accelerators—and in that context, the FP32 and FP16 figures indicate a top-tier compute device.

Ray Tracing and Feature Set

The H100 SXM5 has no dedicated ray tracing cores listed in the fact pack. The architecture is Hopper, which is NVIDIA's compute-focused generation, and the absence of RT cores means hardware-accelerated ray tracing is not a feature of this card. The APIs section shows no DirectX, OpenGL, or Vulkan support data—these fields are null. This is consistent with a server accelerator that has no display outputs and is not intended for graphics rendering.

The tensor cores (528) are the defining feature set element. These are not the same as RT cores; they are matrix-math accelerators designed for deep learning and scientific computing. The FP16 4:1 ratio (267.6 TFLOPS) is the practical manifestation of these tensor cores. The memory subsystem supports this compute focus: 80 GB of HBM3 on a 5120-bit bus delivers 3.36 TB/s of bandwidth. That bandwidth is critical for feeding the tensor cores during large matrix operations, and the 80 GB capacity allows for large model residency without constant host-device transfers.

The bus interface is PCIe 5.0 x16, which provides a high-bandwidth host connection. The card uses an SXM module slot, not a standard PCIe slot form factor. The display outputs field is "No outputs," confirming that this is not a graphics card in the traditional sense. The predecessor is listed as Server Ada and the successor as Server Blackwell, which places the H100 SXM5 in a clear generational line of NVIDIA server accelerators.

Power and Cooling

The H100 SXM5 has a TDP of 700 W. This is a substantial power draw that requires serious thermal management. The fact pack lists a suggested PSU of 1100 W, which means the rest of the system—CPU, motherboard, storage, and any additional accelerators—must be accounted for within that 1100 W budget. The power connectors field is "None," which is expected for an SXM module: power is delivered through the motherboard or baseboard connector, not through auxiliary PCIe power cables.

The slot width is "SXM Module," which means this is not a card you can install in a standard tower chassis. It requires a server platform designed for SXM modules, with dedicated cooling solutions—typically passive heatsinks with high-speed server fans or liquid cooling loops. The 700 W TDP is the thermal design point, and any cooling solution must dissipate that heat continuously under sustained compute loads. The 5 nm process node (TSMC foundry) with 80,000 million transistors on an 814 mm² die means the thermal density is high; the transistor density of 98.3M per mm² is among the highest in the database.

For a PC builder, the practical takeaway is that the H100 SXM5 is not a retrofit option. It demands a server platform with SXM support, a power delivery system rated for 700 W per module, and cooling infrastructure capable of handling that thermal load. The absence of power connectors on the card itself simplifies cabling but shifts the burden to the motherboard's VRM design.

FAQ

Q: What is the FP32 performance of the H100 SXM5?

A: The FP32 throughput is 66.91 TFLOPS, which is the standard single-precision compute rate for the card.

Q: Does the H100 SXM5 support hardware ray tracing?

A: No dedicated ray tracing cores are listed in the specifications. The card has 528 tensor cores but no RT cores, and it has no display outputs.

Q: What memory type and capacity does the H100 SXM5 use?

A: It uses 80 GB of HBM3 memory on a 5120-bit bus, providing 3.36 TB/s of memory bandwidth.

Q: What is the power consumption and PSU requirement?

A: The TDP is 700 W, and the suggested PSU is 1100 W. The card uses no auxiliary power connectors; power is delivered through the SXM module interface.

Q: What is the process node and die size?

A: The H100 SXM5 is manufactured on TSMC's 5 nm process, with 80,000 million transistors on an 814 mm² die. The transistor density is 98.3M per mm².

Q: What is the bus interface and form factor?

A: The card uses a PCIe 5.0 x16 bus interface but is physically an SXM module, not a standard PCIe card. It has no display outputs.

How It Compares

The nearestRivals field is empty in the fact pack, so no direct rival comparisons with specific score deltas can be made. However, the percentile rank of 50 against all GPUs provides a general positioning. The H100 SXM5 is an end-of-life product, with the predecessor being Server Ada and the successor being Server Blackwell. This places it in the middle of NVIDIA's server accelerator lineage—older than the current Blackwell generation but newer than the Ada-based server parts.

Against server accelerators that lack tensor cores or HBM3 memory, the H100 SXM5's 267.6 TFLOPS FP16 and 3.36 TB/s bandwidth would be decisive advantages. Against the newer Server Blackwell parts, the H100 would likely trail in raw compute, but the fact pack does not provide specific numbers for that comparison. The 50th-percentile ranking suggests that in the aggregate database, the H100 SXM5 is not a runaway leader, but that ranking likely includes consumer GPUs with higher rasterization scores that do not reflect this card's compute strengths.

In the absence of rival data, the most honest comparison is against the card's own specifications. The FP32-to-FP16 ratio of 4:1 is the key differentiator. A GPU with 66.91 TFLOPS FP32 but only 2:1 FP16 would deliver half the mixed-precision throughput. The H100's 4:1 ratio means it is optimized for the precision reduction that AI workloads tolerate, making it a specialized tool rather than a general-purpose one.

Who Should Consider It

The H100 SXM5 is for users who need massive FP16 throughput and large memory capacity, and who have the server infrastructure to support an SXM module. The 267.6 TFLOPS FP16 and 3.36 TB/s bandwidth are the specifications that matter for AI training, inference, and scientific computing. If your workload involves transformer models, large matrix multiplications, or any mixed-precision compute, the H100 SXM5's tensor cores are the right tool.

The 80 GB of HBM3 memory is a practical capacity for model sizes that exceed the 24 GB or 48 GB found on consumer or prosumer cards. The 5120-bit bus ensures that memory bandwidth is not a bottleneck when the tensor cores are saturated. For users working with datasets or models that fit within 80 GB, the H100 SXM5 can keep the entire working set on the card, avoiding PCIe transfers that would slow down smaller-memory alternatives.

The 50th-percentile aggregate ranking means that for gaming or traditional graphics workloads, this card is not a sensible choice. It has no display outputs, no RT cores, and a low ROP count (24) that would cripple rasterization performance. The pixel rate of 47.52 GPixel/s is far below what even entry-level gaming GPUs achieve. This is a compute accelerator, and it should only be considered if your primary workload is FP16 or FP32 compute.

The 700 W TDP and 1100 W suggested PSU mean this card is for professional workstations or data center servers, not desktop builds. The SXM form factor requires a compatible motherboard or baseboard, and the cooling must be designed for continuous 700 W dissipation. If you have that infrastructure, the H100 SXM5 offers production-ready compute performance at a mid-pack percentile—but in the compute domain where it operates, the FP16 and memory bandwidth figures place it at the high end. For users running AI training or inference at scale, the H100 SXM5 is a proven, end-of-life workhorse that still delivers competitive throughput. For anyone else, the lack of graphics features and the server-only form factor make it a poor fit.