Nvidia has officially unveiled its next-generation Blackwell GPU architecture, which offers up to 5 times the performance increase compared to the Hopper H100 GPUs that propelled the company to trillion-dollar status and the position of the world’s third-most valuable company. As competitors strive to catch up with Nvidia, the company has also announced the Blackwell-based new products, the Blackwell B200 GPU and the GB200 Superchip. Here’s everything you need to know:

Nvidia’s Blackwell Architecture and the New Blackwell B200:

Nvidia detailed its new generation AI & Tensor Core GPU architecture codenamed Blackwell at a recent event. As anticipated, Blackwell GPUs represent Nvidia’s first Multi-Chip Module (MCM) design, combining two GPUs on the same package.

The Blackwell architecture, produced on the TSMC 4NP process, boasts a total of 104 billion transistors per compute die. The B100 GPUs feature a high-bandwidth interface of 10 TB/s for super-fast chip-to-chip communication. By combining multiple GPUs on the same package, a total of 208 billion transistors are accessed, offering complete GPU cache coherence.

Compared to Hopper, Nvidia’s Blackwell GPU offers 128 billion more transistors, 5 times the AI performance per chip at 20 petaFlops, and 4 times the embedded memory. The Blackwell GPU itself is equipped with an 8 TB/s memory bandwidth interface and up to 192 GB of 8-layer HBM3e memory.

Nvidia claims that the GB200, utilizing a dual GPU and single CPU solution, offers 30 times more performance for LLM inference workloads and reduces costs and energy consumption by up to 25 times compared to the H100. Nvidia has clearly made a significant leap from one generation to the next. For instance, training a 1.8 trillion-parameter AI model previously required 8,000 Hopper GPUs and 15 megawatts of power. Today, the same task can be accomplished with 2,000 Blackwell GPUs and just 4 megawatts of power.

An important advancement is Nvidia’s second-generation Transformer Engine, which doubles computation, bandwidth, and model size by utilizing four bits per neuron instead of eight. Another significant improvement is evident when multiple GPUs are interconnected: the new 5th generation NVLink enables communication between 576 GPUs at 1.8 terabytes per second bidirectional bandwidth.

In terms of product offerings, Nvidia is introducing the GB200 NVL72 package, which combines 36 CPUs and 72 GPUs on a single liquid-cooled rack, providing a total of 720 petaflops of AI training performance or 1,440 petaflops of inference. The company plans to offer NVL72 to cloud service subsidiaries of Amazon, Google, Microsoft, and Oracle. Additionally, Nvidia will provide DGX Superpod for DGX GB200, which combines eight systems to deliver a total of 288 CPUs, 576 GPUs, 240 TB of memory, and 11.5 exaflops of FP4 computation for HPC applications. Nvidia also announced the HGX B200 platform for HPCs, which includes Blackwell Tensor Core GPUs and provides up to 15 times more inference performance compared to the previous generation.

In summary, Nvidia has announced two new solutions: GB200, a combination of Grace CPU and Blackwell B100 GPU, and Blackwell B200 GPUs. These solutions are expected to be released in 2024. Each GB200 Grace Blackwell Superchip is supported by two B200 AI GPUs and a single Grace CPU with 72 ARM Neoverse V2 cores. The platform will offer 40 petaflops of computing performance (INT8) and a massive 864 GB memory pool with 16 TB/s memory bandwidth. In terms of performance, the GB200 Grace Blackwell Superchip platform will offer a 30-fold increase in AI performance compared to the H200 Grace Hopper platform.

Nvidia has not shared specific release dates and pricing information yet. However, the H100 GPU was priced between $25,000 and $40,000 per chip, and the cost of a system based on these chips could reach up to $200,000. Demand for these systems was so high that wait times remained around 11 months for a considerable period. Demand for the Blackwells also appears to be high already.

Stay plugged into the gaming pulse, and anticipate the thrill of new adventures and the nostalgia of departing classics. For the freshest updates, keep an eye on Lavegagames.com – where gaming news comes to life.