NVIDIA Unveils Blackwell Ultra GB300: A Leap Forward in AI Accelerator Technology

NVIDIA Introduces Groundbreaking GB300 Blackwell Ultra GPU

NVIDIA has unveiled its latest AI accelerator, the GB300 Blackwell Ultra, marking a significant advancement in GPU technology for artificial intelligence and scientific computing. This new chip builds upon the capabilities of its predecessor, the GB200, offering substantial improvements in compute resources, memory capacity, and communication speed 1

1

.

Architectural Innovations

The GB300 employs a dual-chip approach, combining two silicon chips that together pack an impressive 208 billion transistors. Manufactured using TSMC's 4NP process, these chips are linked by NVIDIA's NV-HBI technology, providing 10 TB/s of bandwidth between them 1

1

. This design allows the two chips to function as a unified GPU, simplifying programming and maximizing throughput.

Enhanced Compute Capabilities

At the heart of the GB300 are 160 streaming multiprocessors, each containing 128 CUDA cores, totaling 20,480 cores. This represents an increase from the 144 streaming multiprocessors and 18,432 CUDA cores found in the GB200 2

2

. The GPU also features 5th generation Tensor Cores, which accelerate matrix math for AI training and inference, supporting multiple precision modes including FP8, FP6, and the new NVFP4 format 1

1

.

Memory and Bandwidth Advancements

NVIDIA has equipped the GB300 with eight HBM3E stacks, providing a total of 288 GB of memory directly on the GPU package. This represents a significant increase from the 192 GB found in the GB200 3

3

. The memory bandwidth reaches an impressive 8 TB/s over an 8192-bit bus, spread across 16 memory channels 1

1

.

Connectivity and Scalability

The GB300 features advanced connectivity options, including NVLink 5 for GPU-to-GPU links, delivering 1.8 TB/s of bidirectional bandwidth per accelerator. For CPU connections, it uses NVLink-C2C, linking directly with Grace CPUs at 900 GB/s. Additionally, the accelerator supports PCIe 6.0 x16, doubling throughput over PCIe 5.0 to 256 GB/s 1

1

.

Performance and Power Considerations

The GB300's enhanced performance comes at the cost of increased power consumption, with a thermal graphics power (TGP) reaching 1400 W, up from 1200 W in the GB200 1

1

. This presents significant engineering challenges for cooling and power delivery in data center and supercomputing applications.

Real-World Performance

CoreWeave, a cloud services provider, has demonstrated the GB300's capabilities in a benchmark using the DeepSeek R1 reasoning model. The test showed that a system with just 4 GB300 GPUs could deliver 6 times higher raw throughput per GPU compared to a 16-GPU cluster of H100s 4

4

. This significant performance gain is attributed to reduced tensor parallelism and improved inter-GPU communication.

Implications for AI and Scientific Computing

The GB300 Blackwell Ultra represents a major step forward in GPU technology for AI and scientific computing. Its increased memory capacity and bandwidth, coupled with advanced Tensor Cores and the new NVFP4 format, enable the handling of larger AI models and more complex scientific simulations 2

2

. This advancement is particularly crucial as AI models continue to grow in size and complexity, with some reaching trillions of parameters.

As NVIDIA continues to push the boundaries of GPU technology, the GB300 Blackwell Ultra sets a new standard for AI accelerators, promising to drive innovation in fields ranging from natural language processing to scientific research and beyond.