NVIDIA H800 PCIe 80 GB

NVIDIA H800 PCIe 80 GB is a server-class accelerator built on the Hopper architecture, using the GH100 chip fabricated on TSMC’s 5 nm process. It packs 80,000 million transistors on an 814 mm² die, yielding a transistor density of 98.3M per mm². The card is currently in active production, was released on 2023-03-20, and sits as the successor to Server Ada and predecessor to Server Blackwell. With no benchmark scores listed, its percentile versus all GPUs is 50, and its average benchmark score is 0, meaning quantitative performance comparisons must rely on architectural specifications rather than measured results.

Who Should Consider It

The H800 PCIe 80 GB targets workloads where memory capacity and bandwidth dominate over raw rasterization speed. With 80 GB of HBM2e memory and a 5120-bit bus width, the card delivers 2.04 TB/s of bandwidth — a figure that suggests it is engineered for large datasets, high-resolution compute tasks, and AI inference or training batches that exceed the VRAM of typical consumer cards. The shading units number 14,592, with 456 TMUs and only 24 ROPs; the low ROP count relative to shading units indicates this is not a card optimized for traditional pixel-pushing or high frame-rate gaming. Instead, the pixel rate of 42.12 GPixel/s and texture rate of 800.3 GTexel/s point toward compute-heavy scenarios where texture sampling and math throughput matter more than final raster output.

For resolution-based recommendations, the data implies that at 1080p or 1440p, the H800’s strengths would be wasted, as its 24 ROPs would bottleneck fill-rate-limited scenes. At 4K or beyond, the massive memory pool and bandwidth could shine in scenarios like large-scale scientific simulations, deep learning model training with huge batches, or rendering workloads that precompute massive textures. The lack of display outputs — the card has none — confirms it is not intended for direct monitor connection, reinforcing its role as a compute accelerator in servers. Users who need to process datasets larger than 24 GB or 48 GB would find the 80 GB capacity compelling, especially when memory bandwidth is the limiting factor. Conversely, anyone seeking real-time graphics performance or low-latency rendering should look elsewhere, as the architecture prioritizes throughput over latency.

Ray Tracing and Feature Set

The FACT PACK lists no RT cores for the H800 PCIe 80 GB, and the RT core field is null. This absence is notable: while the card does include 456 tensor cores, which are dedicated to matrix math and AI workloads, there is no dedicated hardware for ray tracing acceleration. The tensor cores deliver FP16 performance of 204.9 TFLOPS (4:1 ratio), indicating a strong focus on mixed-precision compute common in neural network training. The FP32 throughput is 51.22 TFLOPS, which is substantial for general compute but irrelevant for ray tracing without dedicated cores.

API support is entirely absent from the FACT PACK — DirectX, OpenGL, and Vulkan fields are all null. This suggests the card is not marketed for graphics APIs, and its driver stack may be oriented toward CUDA or compute frameworks rather than real-time rendering. The architecture is Hopper, which in the broader product line includes features like NVLink and advanced tensor operations, but the FACT PACK does not enumerate those specifics for this SKU. The tensor core count of 456 aligns with the TMU count, hinting that each texture unit may pair with a tensor unit for fused operations. For ray tracing, the data indicates no hardware support, meaning any RT workloads would fall back to compute shaders on the 14,592 shading units, which would be inefficient. The lack of RT cores and null API fields strongly imply that this is a pure compute product, not a graphics card.

Benchmark Performance

No benchmark scores are provided in the FACT PACK, and the nearestRivals array is empty. The average benchmark score is 0, and the percentile versus all GPUs is 50, which is a neutral midpoint — but without actual scores, this percentile cannot be interpreted as a performance ranking. The absence of data forces reliance on theoretical peak rates: FP32 at 51.22 TFLOPS, FP16 at 204.9 TFLOPS, and memory bandwidth of 2.04 TB/s. These numbers, when compared to typical server GPUs, would place the H800 in a high-compute tier, but the FACT PACK does not provide rival names or delta percentages to quantify that position. The pixel rate of 42.12 GPixel/s and texture rate of 800.3 GTexel/s are derived from the clock speeds and unit counts, but they are not benchmark results.

The base clock is 1095 MHz, boosting to 1755 MHz, with memory clocked at 1593 MHz (3.2 Gbps effective). These clocks, combined with the 5 nm process and 350 W TDP, suggest a power-efficient design relative to older nodes, but again, no comparative data exists. The fact that the card has 24 ROPs is striking — that is a low count even for a compute card, and it will limit any non-compute output. In synthetic compute benchmarks, the FP16 TFLOPS figure of 204.9 is likely the headline number, as it reflects tensor core throughput. Without measured scores, any statement about "performance" must be couched in architectural terms: the H800’s strengths are memory bandwidth and FP16 compute, not rasterization or ray tracing.

How It Compares

The nearestRivals array is empty, so there are no direct competitor comparisons available from the FACT PACK. The predecessor is Server Ada and the successor is Server Blackwell, but the FACT PACK provides no specifications for either. This absence means the H800’s position relative to other GPUs cannot be quantified with deltas or percentiles. The percentile of 50 versus all GPUs is the only positional data, but with an average score of 0, it is unclear whether that percentile reflects a median ranking or a placeholder. The card’s die size of 814 mm² and transistor count of 80,000 million are large numbers that suggest a flagship-class part, but without rival data, no competitive analysis is possible. In practical terms, the H800 appears to sit between generations — after Server Ada and before Server Blackwell — but its exact performance envelope relative to those is unspecified.

The lack of rivals also means no pricing comparisons can be made, and the launch MSRP field is null. The data indicates the card is dual-slot, 268 mm long (10.6 inches), and 111 mm tall (4.4 inches), but these physical dimensions do not translate to performance rankings. The only inference from the empty nearestRivals is that this database entry has not been populated with comparison points, so any conclusion about how it stacks against competitors must remain qualitative. The H800’s design choices — 80 GB HBM2e, 5120-bit bus, 456 tensor cores — clearly target a niche that may not have direct rivals in the same form factor.

FAQ

Q: Does the NVIDIA H800 PCIe 80 GB support ray tracing?

A: No. The FACT PACK lists no RT cores (the field is null), and the API support fields for DirectX, OpenGL, and Vulkan are also null, indicating no dedicated ray tracing hardware.

Q: What is the memory bandwidth of this card?

A: The memory bandwidth is 2.04 TB/s, achieved via 80 GB of HBM2e memory on a 5120-bit bus, with memory clocked at 1593 MHz (3.2 Gbps effective).

Q: How many tensor cores does it have?

A: It has 456 tensor cores, which deliver FP16 performance of 204.9 TFLOPS (4:1 ratio). The FP32 throughput is 51.22 TFLOPS.

Q: Can I connect a monitor to this card?

A: No. The FACT PACK states "No outputs" for display outputs, so it is a compute-only accelerator without any video connectors.

Q: What is the power requirement?

A: The TDP is 350 W, and the suggested PSU is 750 W. It requires a single 16-pin power connector and is a dual-slot card.

Q: What process node is the chip built on?

A: The GH100 chip is built on TSMC’s 5 nm process, with 80,000 million transistors on an 814 mm² die, giving a transistor density of 98.3M per mm².

Power and Cooling

The H800 PCIe 80 GB has a TDP of 350 W, which is a significant but manageable power draw for a server accelerator. The suggested PSU is 750 W, providing headroom for the card and supporting system components. Power is delivered via a single 16-pin connector, which is the modern high-current interface. The card is dual-slot in width, meaning it will occupy two expansion slots in a chassis, and its dimensions are 268 mm in length (10.6 inches) and 111 mm in height (4.4 inches). These physical specs are important for server rack compatibility, but the FACT PACK does not specify cooling type — it only states the slot width and power connectors. The 5 nm process from TSMC likely contributes to thermal efficiency, but no specific cooling solution or noise levels are provided. The 350 W TDP, when combined with the 750 W PSU recommendation, suggests that system builders should ensure sufficient airflow, but the dual-slot design likely includes a passive heatsink or blower style cooler. Without explicit cooling details, the data implies that the card is designed for server environments with forced airflow rather than open-air desktop cases.

Memory Subsystem

The memory subsystem is the H800’s defining feature: 80 GB of HBM2e on a 5120-bit bus, delivering 2.04 TB/s of bandwidth. This is an enormous pool — larger than most GPUs — and the bandwidth of 2.04 TB/s is essential for feeding the 14,592 shading units and 456 tensor cores. The memory clock is 1593 MHz, translating to 3.2 Gbps effective per pin. The 5120-bit bus width is exceptionally wide, which is why bandwidth reaches such high levels despite modest memory clocks. For high-resolution workloads, this means that data movement is rarely a bottleneck; the card can load and store large tensors or textures at rates that exceed what most CPUs can generate. The 80 GB capacity allows entire models or datasets to reside in VRAM, avoiding PCIe transfers. However, the 24 ROPs and 42.12 GPixel/s pixel rate will limit any rendering tasks, so the memory advantage is purely for compute and data-intensive applications. The HBM2e type is notable because it stacks memory dies, enabling the 80 GB capacity in a compact form factor, but the FACT PACK does not specify the number of stacks or the memory configuration. The 2.04 TB/s bandwidth is roughly 40 times higher than typical consumer cards, but no comparative numbers are available in the FACT PACK. This subsystem is clearly designed for AI training, scientific computing, and large-scale data analytics where memory capacity and bandwidth are critical constraints.