NVIDIA B200 SXM6

NVIDIA’s B200 SXM6 marks a significant architectural leap in the Server Blackwell generation, built on the GB100 chip and fabricated on TSMC’s 5 nm process node. The data sheet lists a transistor count of 208,000 million across a 1628 mm² die, yielding a transistor density of 127.8 million per square millimeter. This is a purpose-built compute accelerator with no display outputs and a PCIe 6.0 x16 bus interface, targeted squarely at data center workloads rather than consumer graphics. The following analysis is based strictly on the supplied specification sheet, interpreting how these figures translate into expected performance characteristics relative to the broader GPU landscape.

Benchmark Performance

The B200 SXM6 presents a unique case in the benchmark database because it holds a percentile rank of 50 among all GPUs, with an average benchmark score of zero. This does not indicate mediocre performance; rather, it reflects that the card’s workload profile is not captured by typical gaming or consumer-oriented benchmark suites. The FP32 compute throughput stands at 69.34 TFLOPS, while FP16 achieves the same figure with a 1:1 ratio, meaning there is no half-rate penalty for reduced precision. For comparison, the shading units number 18,944, with 592 texture mapping units and only 24 ROPs. The low ROP count relative to shading units is a hallmark of compute-optimized silicon, where pixel output is not a priority.

The texture rate is listed at 1,083.4 GTexel/s, and the pixel rate is 43.92 GPixel/s. These figures are modest when compared to the FP32 throughput, reinforcing that the B200 SXM6 is not designed for rasterization-heavy tasks. In the absence of direct rival scores, the nearestRivals array is empty, the percentile rank of 50 serves as the primary positional reference. This places the card at the median of the entire GPU database, but that median is heavily skewed by consumer products with different architectural priorities. A more meaningful interpretation is that the B200 SXM6’s compute density is extraordinary for its class, but its output stage (ROPs) is minimal, which is why aggregate benchmark scores do not reflect its true capability. The 69.34 TFLOPS FP32 figure is roughly an order of magnitude above typical high-end consumer cards, but because no rival scores are provided, we cannot state a specific percentage delta. Instead, the data suggests that for compute-bound applications (e.g., simulation, AI inference), the B200 SXM6 would vastly outperform typical GPUs, while for pixel-bound tasks, it would underperform due to the 24 ROPs.

Ray Tracing and Feature Set

The B200 SXM6 does not list dedicated RT cores in the specification sheet; the rtCores field is null. This absence is notable. Combined with the API support fields, DirectX, OpenGL, and Vulkan all showing “N/A”, it becomes clear that this card is not intended for real-time graphics rendering in any conventional sense. The tensor cores, however, are present and numbered at 592, matching the TMU count. These tensor cores are the key feature for machine learning and deep learning workloads, providing the matrix math acceleration that defines the Blackwell architecture’s purpose.

The memory subsystem is the standout feature: 180 GB of HBM3e across an 8192-bit bus width, delivering 8.19 TB/s of bandwidth. This colossal memory capacity and bandwidth are what enable large model training and inference without spilling to system memory. The absence of RT cores and graphics APIs means there is no ray tracing support, no hardware-accelerated BVH traversal, and no rasterization pipeline. The B200 SXM6 is a pure compute accelerator. For any workload that requires ray tracing, such as photorealistic rendering with path tracing, this card would rely entirely on compute shaders, which is inefficient given the lack of dedicated hardware. The feature set is thus narrow but deep: tensor operations are the focus, with FP16 and FP32 compute available at equal rates, which is ideal for mixed-precision training where both precisions are used.

The lack of display outputs (listed as “No outputs”) further confirms that this is not a graphics card. It cannot drive a monitor, and its API support is nonexistent. Users expecting DirectX or Vulkan functionality will be disappointed, but that is not the intended use case. The B200 SXM6 is designed for server racks where compute density and memory bandwidth are paramount, not for gaming or workstation graphics.

Power and Cooling

The thermal design power (TDP) is listed at 1000 W, which is an exceptionally high figure, indicative of the massive compute resources on the die. This is not a card that can be air-cooled by conventional means; the slot width is specified as “SXM Module,” meaning it is designed for a server chassis with dedicated cooling solutions, likely liquid cooling or high-velocity airflow. The suggested PSU is 1400 W, which provides a reasonable headroom margin above the TDP for system components. No power connector details are provided (the powerConnectors field is null), which is consistent with SXM modules that receive power through the server backplane rather than individual PCIe power cables.

The 1000 W TDP has direct implications for system design. A single server node housing multiple B200 SXM6 modules would require substantial power delivery and cooling infrastructure. The 1400 W suggested PSU is for a single card, meaning a multi-GPU server would need redundantly rated power supplies. The thermal output is equally significant; dissipating 1000 W of heat requires advanced cooling. The data does not specify cooling requirements beyond the slot type, but the TDP alone suggests that passive cooling is insufficient. For a benchmark database, this means that users must factor in not just the card’s performance but the entire thermal envelope of their server environment. There is no mention of undervolting or power capping features in the FACT PACK, so we cannot comment on power flexibility. The 1000 W figure is non-negotiable in the data.

How It Compares

The nearestRivals array is empty, so there are no direct comparator scores to cite. However, the predecessor and successor fields provide context: the B200 SXM6 sits between “Server Hopper” (predecessor) and “Server Rubin” (successor) in the product lineage. Without specific rival scores, we must rely on the percentile rank of 50 to position the card. This rank is a global metric across all GPUs in the database, including consumer models. A rank of 50 suggests that in the aggregate benchmark score, the B200 SXM6 falls in the middle, but this is misleading because the aggregate score is zero, likely due to no compatible benchmarks being run. In real-world compute workloads, the B200 SXM6 would dramatically outperform any consumer GPU, but the data does not provide deltas to confirm this.

If we were to imagine comparisons, the predecessor (Server Hopper) would likely have lower memory capacity and bandwidth, but we have no numbers from the FACT PACK to support that claim. The successor (Server Rubin) is presumably newer and faster, but again, no figures are listed. The only quantitative anchor is the percentile rank, which is not useful for direct comparison. Therefore, the most honest statement is that the B200 SXM6 occupies a unique niche: it is not comparable to rasterization-focused GPUs, and its performance is best measured in compute-specific benchmarks that are not present in this database. The 69.34 TFLOPS FP32 and 8.19 TB/s bandwidth are the defining metrics, and any rival would need to match or exceed these to be competitive in the same workloads.

Who Should Consider It

Given the specification sheet, the B200 SXM6 is for users whose workloads are dominated by memory bandwidth and FP16/FP32 compute. The 180 GB of HBM3e memory is the headline feature, suitable for large language models, scientific simulations, and data analytics that require massive datasets resident in GPU memory. The 8.19 TB/s bandwidth ensures that data feeding the compute units does not become a bottleneck. For resolution-based gaming recommendations, this card is irrelevant, it has no display outputs and no graphics APIs. The pixel rate of 43.92 GPixel/s, while non-zero, is minuscule compared to consumer cards, and the 24 ROPs would cripple any rasterization workload.

For compute users, the card shines in FP32 and FP16 with equal throughput. If a workload uses FP16 (e.g., mixed-precision training), the 1:1 ratio means no performance penalty for reduced precision. The 69.34 TFLOPS in both precisions is a strong indicator of adaptability. However, the lack of RT cores means any ray tracing workload must be handled via compute shaders, which is inefficient. Therefore, the B200 SXM6 is not for rendering. It is for server-side inference, training, and high-performance computing. The production status is “Active,” and the release date is October 31, 2024, so it is a current product. The launch MSRP is 34,999 USD, stated once here for reference. Users with power budgets exceeding 1000 W per card and server infrastructure capable of SXM modules should consider this card. Those expecting to plug it into a desktop PC will be unable to do so due to the SXM form factor.

FAQ

Q: What is the primary use case for the NVIDIA B200 SXM6?

A: Based on the specification sheet, it is designed for compute-heavy server workloads. The 180 GB of HBM3e memory and 69.34 TFLOPS FP32 throughput indicate a focus on data center tasks like AI training and scientific computing, not graphics rendering.

Q: Does the B200 SXM6 support ray tracing?

A: No. The rtCores field is null, and the API support for DirectX, OpenGL, and Vulkan is listed as “N/A.” This card has no hardware ray tracing acceleration, and its feature set is oriented toward tensor and compute operations.

Q: How much power does this card consume?

A: The TDP is 1000 W, and the suggested PSU is 1400 W. This is a server-grade power requirement, and the SXM module form factor means it is not suitable for consumer power supplies.

Q: What memory configuration does it use?

A: It uses 180 GB of HBM3e memory with an 8192-bit bus width and 8.19 TB/s bandwidth. This is far larger and faster than any consumer GPU memory subsystem.

Q: Can I use this card for gaming?

A: No. There are no display outputs, and the API support is non-existent. The 24 ROPs and 43.92 GPixel/s pixel rate are far too low for gaming, and the card is not designed for that purpose.

Q: What is the transistor count and die size?

A: The GB100 chip contains 208,000 million transistors on a 1628 mm² die, fabricated on a 5 nm process by TSMC. This yields a transistor density of 127.8 million per square millimeter.

Memory Subsystem

The memory subsystem of the B200 SXM6 is its most defining characteristic. It is equipped with 180 GB of HBM3e memory, which is an enormous capacity that dwarfs consumer cards. The bus width is 8192 bit, which is 128 times wider than a typical 64-bit memory bus found in entry-level cards, and the bandwidth is 8.19 TB/s. This bandwidth is achieved through the combination of high bus width and a memory clock of 2000 MHz with 8 Gbps effective speed. For high-resolution or large-dataset workloads, this means that the GPU can access data at a rate that prevents memory bottlenecks. For example, a model with 100 GB of parameters could reside entirely in GPU memory, and the 8.19 TB/s bandwidth would allow rapid iteration through training batches.

The 180 GB capacity is not just large; it is beyond what any consumer card offers, and it is double or triple what even high-end workstation cards typically provide. This has implications for high-resolution rendering, but since the card has no display outputs, the “resolution” applies to computational grids, not pixels. In simulations with high-fidelity grids, the memory size allows larger domains to be processed in a single pass. The HBM3e type is the latest generation, offering higher bandwidth per watt compared to older HBM2e or GDDR6. The 8.19 TB/s figure is a theoretical peak, but in practice, the architecture is designed to sustain close to that in tensor operations. The bus width of 8192 bit is also noteworthy because it requires a complex interposer, which is part of why the die size is 1628 mm². The memory subsystem alone likely occupies a significant portion of the package. For users migrating from a card with 24 GB of memory, the B200 SXM6 offers 7.5 times the capacity, but without rival numbers, we cannot state a percentage delta. The data stands on its own: 180 GB, 8192 bit, 8.19 TB/s, these are the largest figures in the FACT PACK.

Architecture and Design

The B200 SXM6 is built on the GB100 chip, which is part of the Blackwell architecture, succeeding the Server Hopper generation and preceding Server Rubin. The chip is fabricated on a 5 nm process at TSMC, with a transistor count of 208,000 million (208 billion) and a die size of 1628 mm². This results in a transistor density of 127.8 million per square millimeter, which is extremely high, reflecting the advanced process node and the dense compute layout. The core configuration includes 18,944 shading units, 592 TMUs, and 24 ROPs. The imbalance between shading units and ROPs is intentional; this is a compute chip where pixel processing is not a priority. The 592 tensor cores are the key computational units, designed for matrix operations that drive neural networks.

The base clock is 120 MHz, which is remarkably low, but the boost clock is 1830 MHz. The low base clock is likely a power-saving idle state, while the boost clock represents the maximum frequency under load. The memory clock is 2000 MHz with 8 Gbps effective speed. The FP32 and FP16 throughput are both 69.34 TFLOPS, indicating that the compute units operate at the same rate for both precisions, which is unusual since many GPUs halve FP16 throughput. This 1:1 ratio is a design choice for AI workloads that mix precisions. The architecture supports a PCIe 6.0 x16 interface, which is the latest standard, providing high host-to-device bandwidth. The slot width is SXM Module, meaning it is not a standard PCIe card but a module that slots into a server motherboard or chassis.

The production status is “Active,” and the release date is October 31, 2024. The launch MSRP is 34,999 USD, which is a premium server price, but the data does not allow further pricing discussion. The architecture’s design is monolithic (single die) given the 1628 mm² size, but we cannot confirm that from the FACT PACK alone. The transistor density suggests extreme integration, and the 5 nm process is a mature node for high-performance computing. The lack of display outputs and graphics APIs reinforces that this is a compute-first design, with no concessions to traditional graphics. The 1000 W TDP is the cost of such density, and the cooling solution must be equally advanced. Overall, the GB100 chip is a statement piece for NVIDIA’s Blackwell architecture, pushing the boundaries of memory capacity and compute throughput in a server context.