NVIDIA B300 SXM6 AC

The NVIDIA B300 SXM6 AC, built on the Blackwell Ultra architecture, represents a decisive step forward for server-class compute. With a benchmark score of 369,831 points, it sits at the 100th percentile of all GPUs, making it the reference point against which all other accelerators are currently measured. This analysis examines its data-sheet specifications and benchmark results to provide a clear picture of its capabilities and positioning.

Who Should Consider It

The B300 SXM6 AC is engineered for workloads that demand the absolute maximum compute throughput available, regardless of power or thermal constraints. The benchmark data shows a Geekbench OpenCL score of 369,831, which places it at the top of the performance hierarchy. This is not a product for general-purpose rendering or consumer-grade tasks; its "No outputs" display configuration confirms it is a compute-only accelerator.

Given its 288 GB of HBM3e memory and 8.19 TB/s bandwidth, this GPU is suited for large-scale artificial intelligence training, scientific simulation, and data center workloads where memory capacity and bandwidth are the primary bottlenecks. The 1100 W TDP and 1500 W suggested PSU indicate it is intended for dense server racks with dedicated power infrastructure, not workstation builds. Users running inference on massive language models or processing datasets that exceed 200 GB will find the capacity essential, while those with smaller models may find the power draw excessive. For high-resolution rendering or real-time graphics, this card is not appropriate; its lack of display outputs and N/A API support for DirectX, OpenGL, and Vulkan makes it exclusively a backend compute device.

Ray Tracing and Feature Set

The data pack lists no RT cores for this GPU, and the API support is uniformly "N/A" for DirectX, OpenGL, and Vulkan. This indicates a complete absence of traditional graphics pipeline capabilities. The B300 SXM6 AC is not designed for real-time ray tracing or rasterized game rendering; it is a pure compute accelerator.

Instead of RT cores, the card relies on 592 Tensor Cores, which drive its FP16 and FP32 performance. The FP32 throughput is 76.99 TFLOPS, with FP16 delivering an identical 76.99 TFLOPS at a 1:1 ratio. This symmetry between precision levels suggests an architecture optimized for mixed-precision machine learning workloads where FP16 is often sufficient for training and inference. The 592 tensor cores are the primary workhorses, handling matrix operations that form the basis of neural network layers. The lack of graphics APIs means software must interact with the hardware directly via compute frameworks like CUDA or OpenCL, which is standard for server accelerators of this class.

Memory Subsystem

The memory subsystem is one of the most defining features of the B300 SXM6 AC. It comes equipped with 288 GB of HBM3e memory on a massive 8192-bit bus, yielding a bandwidth of 8.19 TB/s. This is a substantial increase over previous generations, allowing the GPU to feed its 18,944 shading units and 592 tensor cores without starvation.

For high-resolution workloads, the implications are clear. A dataset or model that fits within 288 GB can be processed entirely on-chip, avoiding slow PCIe transfers to host memory. The 8.19 TB/s bandwidth means that large matrix multiplications and convolution operations can access data at unprecedented speeds. Compared to the NVIDIA B200, the B300’s bandwidth advantage contributes to its 7% higher average benchmark score. The 8192-bit bus width is double that of many consumer cards, reflecting a design philosophy that prioritizes data movement above all else. The memory clock is listed as 2000 MHz with 8 Gbps effective transfer rate, which, when multiplied across the bus, achieves the stated 8.19 TB/s figure. This configuration is specifically tailored for large batch sizes and high-resolution spatial data, common in scientific computing and advanced AI models.

Power and Cooling

The B300 SXM6 AC has a Thermal Design Power (TDP) of 1100 W, which is among the highest for any single accelerator module. The suggested power supply unit is rated at 1500 W, indicating that the system should have headroom above the GPU’s peak draw to accommodate other components like CPUs and memory.

The slot width is listed as "SXM Module," meaning it is not a standard PCIe card but rather a module designed for proprietary server chassis with integrated cooling solutions. No power connectors are specified, as SXM modules typically receive power through the motherboard backplane rather than external cables. The bus interface is PCIe 6.0 x16, which provides high-bandwidth communication with the host system, though the compute workloads will primarily rely on the on-board memory. Cooling must be handled by the server’s chassis fans or liquid cooling loops, as the 1100 W TDP generates substantial heat that requires active management. Data center planners must account for this power density when designing power delivery and thermal systems, as a single server housing multiple B300 modules could exceed 5 kW of GPU-only power draw.

How It Compares

The B300 SXM6 AC outperforms its nearest rivals by significant margins, as shown in the benchmark data. Each comparison below uses the average scores and delta percentages from the nearestRivals field.

NVIDIA B200: The B200 scores 345,482 points, which is 7% lower than the B300. This is the closest competitor, indicating that the B300 is an iterative but meaningful improvement over its direct predecessor in the Blackwell family. The 7% delta suggests that the B300 offers higher clock speeds or better memory efficiency, though the core architecture is similar. For users with existing B200 deployments, the upgrade to B300 yields a modest but measurable performance gain.

NVIDIA H200 NVL: The H200 NVL achieves 334,891 points, trailing the B300 by 10.4%. This represents a previous-generation product (Server Hopper) that still performs admirably but cannot match the Blackwell Ultra’s throughput. The 10.4% gap is substantial enough to justify upgrading for performance-critical applications, especially when considering the B300’s larger memory capacity and newer architecture.

AMD Instinct MI300X: AMD’s flagship scores 317,994 points, which is 16.3% behind the B300. This is a significant margin that underscores NVIDIA’s continued dominance in the server accelerator market. The MI300X is a capable competitor, but the B300’s superior memory bandwidth and tensor core performance give it a clear edge in compute-heavy benchmarks.

NVIDIA L40S: The L40S scores 295,763 points, a full 25% behind the B300. This card is positioned as a lower-tier option, likely with reduced memory capacity and fewer compute units. The 25% delta is the largest among the rivals, reinforcing that the B300 is at the top tier of NVIDIA’s lineup, while the L40S serves a different segment focused on more cost-sensitive or less demanding workloads.

FAQ

Q: What is the primary benchmark score for this GPU?

A: The B300 SXM6 AC scores 369,831 points in Geekbench OpenCL, placing it at the 100th percentile of all GPUs.

Q: Does this card support DirectX or Vulkan?

A: No. The API support is listed as N/A for DirectX, OpenGL, and Vulkan, confirming it is a compute-only accelerator with no graphics outputs.

Q: How much memory does it have, and what type?

A: It has 288 GB of HBM3e memory on an 8192-bit bus, providing 8.19 TB/s of bandwidth.

Q: What power supply is recommended?

A: A 1500 W PSU is suggested, given the 1100 W TDP of the module.

Q: How does it compare to the NVIDIA B200?

A: The B300 is 7% faster than the B200, which scores 345,482 points.

Q: Is it suitable for gaming or graphics rendering?

A: No. It has no display outputs and no graphics API support, making it exclusively for server-side compute tasks.

Benchmark Performance

The Geekbench OpenCL result of 369,831 points is the sole benchmark data point, but it provides a robust indicator of raw compute capability. The percentile rank of 100 means it outperforms every other GPU in the database, a testament to its absolute performance ceiling.

Against its nearest rivals, the deltas are clear: 7% over the B200, 10.4% over the H200 NVL, 16.3% over the MI300X, and 25% over the L40S. These percentages translate to absolute score differences of 24,349 points versus the B200, 34,940 points versus the H200 NVL, 51,837 points versus the MI300X, and 74,068 points versus the L40S. The scaling is not linear with hardware specifications, but the pattern is consistent: the B300 leads by a widening margin as you move down the competitor list.

The FP32 throughput of 76.99 TFLOPS and FP16 of 76.99 TFLOPS indicate that the architecture does not halve throughput for FP16, which is unusual and beneficial for AI workloads that use mixed precision. The texture rate of 1,202.9 GTexel/s and pixel rate of 48.77 GPixel/s are nominal compute metrics, though with no display outputs, these are more reflective of internal processing capability than actual rendering speed. The 18,944 shading units are the integer and FP32 execution units, while the 592 TMUs and 24 ROPs handle texture and raster operations, respectively. The low ROP count (24) relative to shading units is typical for compute-focused accelerators, as they do not need high fill rates for display output.

Architecture and Design

The B300 SXM6 AC is built on the GB110 chip, which is fabricated on a 5 nm process at TSMC. The chip contains 208,000 million transistors on a die size of 1628 mm², resulting in a transistor density of 127.8 million transistors per square millimeter. This is a massive die, reflecting the enormous compute and memory resources packed into the module. The architecture is Blackwell Ultra, which is a refinement of the Blackwell design, and it is part of the "Server Blackwell (Bxx)" generation.

The core configuration includes 18,944 shading units, 592 TMUs, and 24 ROPs. The tensor core count is 592, which matches the TMU count, suggesting a design where each tensor core is paired with a texture mapping unit in a unified compute cluster. The base clock is 1665 MHz, with a boost clock of 2032 MHz, allowing for dynamic frequency scaling under thermal and power headroom. The memory clock is 2000 MHz with 8 Gbps effective data rate, which is standard for HBM3e. The predecessor is listed as Server Hopper, and the successor is Server Rubin, indicating a clear generational roadmap. The production status is Active, with a release date of September 10, 2025. The bus interface is PCIe 6.0 x16, which offers double the bandwidth of PCIe 5.0, though the compute workloads will primarily stay on the 288 GB of on-board memory.

The lack of a launch MSRP field in the data pack means no pricing information is available, and the analysis avoids any cost discussion. The die size of 1628 mm² is exceptionally large, near the reticle limit for photolithography, which explains the high transistor count and the need for advanced cooling. The 5 nm process is mature, but the sheer scale of the chip requires careful yield management. The absence of a codename or series field suggests this is a standalone product without a broader family grouping, though the generation field places it firmly in the Server Blackwell line. The architecture is designed for scale-out data center deployment, where multiple modules work in tandem to solve massive problems, and the B300’s specifications are tailored to that role.