#superchip

Jensen Huang just revealed the NVIDIA GB200 NVL72, a monster data center superchip boasting 72 Blackwell GPUs and 1.4 exaFLOPS of compute power! We break down these mind-blowing specs and explore what they mean for the future of AI and high-performance computing.

Murex MX.3 Benefits Form NVIDIA Grace Hopper Superchip

Pursuing optimal speed and energy efficiency, Spatz utilised NVIDIA’s CUDA and GPU-accelerated computing as head of quantitative research at Paris-based Murex, a trading and risk management software company.

The team of quants at the corporation has started testing the NVIDIA Grace Hopper Superchip, always looking to push the boundaries of technology. Helping clients more effectively price and manage the exposure to credit and market risk associated with derivatives contracts is the main goal of the initiative.

High-performance computing (HPC) and large-scale artificial intelligence (AI) applications were the inspiration behind the creation of the ground-up NVIDIA GH200 Grace Hopper Superchip. With the superchip, researchers and scientists may solve the most difficult issues in the world with unprecedented ease. Applications processing terabytes of data can benefit from up to 10X higher performance.

Examine the Superchip More Closely

A CPU and GPU coherent memory paradigm for accelerated AI and HPC applications is provided by the NVIDIA GH200 Grace Hopper Superchip, which combines the NVIDIA Grace and Hopper architectures via NVIDIA NVLink-C2C. 7 times quicker than PCIe Gen5, the superchip has 900 gigabytes per second (GB/s) of coherent interface. It further boosts fast computing and generative AI with GPU memory that includes HBM3 and the HBM3e. NVIDIA AI Enterprise, the HPC SDK, and Omniverse are just a few of the NVIDIA software stacks and platforms that are compatible with the GH200.

Up to three times the bandwidth and 3.5 times the GPU memory capacity of an H100 in a single server can be obtained with the Dual GH200 Grace Hopper Superchip, which fully connects two GH200 Superchips via NVLink.

The NVIDIA Grace Hopper architecture combines the adaptability of the NVIDIA Grace CPU with the ground-breaking performance of the NVIDIA Hopper GPU into a single superchip, linked with the high-bandwidth, memory-coherent NVIDIA’s Chip-2-Chip (C2C) connection technology.

Superchips can use the memory-coherent, high-bandwidth, low-latency NVIDIA NVLink-C2C interface. At the centre of the GH200 Grace Hopper Superchip, it can give a total bandwidth of up to 900 gigabytes per second (GB/s), which is seven times faster than the PCIe Gen5 lanes that are often found in accelerated systems. Through NVLink-C2C, programmes can directly use the high bandwidth memory of the Grace CPU and oversubscribe the GPU’s memory.

The GPU has direct access to 7X more fast memory than HMB3 or nearly 8X more fast memory than HBM3e, depending on how the GH200 memory is configured, thanks to the up to 480GB of the LPDDR5X CPU memory per GH200 Grace Hopper Superchip. Standard servers may easily be equipped with GH200 to execute a range of applications that require a lot of compute and memory, such as data analytics and inference. The GH200 can also be used in conjunction with the NVIDIA NVLink Switch System, allowing all GPU threads to operate on up to 256 GPUs linked to the switch and high bandwidth access to up to 144 terabytes (TB) of memory.

With NVIDIA’s software stacks, the NVIDIA GH200 Grace Hopper Superchip. In addition to case studies and interactive demonstrations, it also features large language models (LLM) and real-world applications.

Murex MX.3 user manual pdf

Each day, the Murex MX.3 platform serves over 60,000 users across 65 countries. In terms of trading, risk management, and asset class operations, MX.3 helps banks, asset managers, pension funds, and other financial institutions.

Led by Grace Hopper, Managing Risk with MX.3

In order to perform risk models on large datasets for pricing and risk calculations, as well as to enable real-time decision-making, financial institutions require high-performance computing infrastructure.

Under MX.3, the trading book is fundamentally reviewed, x-valuation adjustment (XVA), BASEL capital regulations, and credit and market risk are all covered. Derivative contract valuation adjustments of many kinds, including funding valuation adjustments, margin value adjustments, and credit value adjustments (CVA), are handled by XVA.

Murex is evaluating Grace Hopper for XVA computations on the MX.3 platform, along with other functions like pricing assessment, sensitivity analysis, market risk calibration, and profit and loss computations across many asset classes.

The Murex platform benefits from Grace Hopper’s quicker calculations and power savings.

With its heterogeneous design and special combination of CPU and GPU computations, Grace Hopper is ideal for counterparty credit risk workloads like CVAs, according to Spatz. “Grace is the fastest and most power-efficient processor in the trading world, enabling green IT to become a reality in risk calculations.”

A 4x reduction in energy usage and a 7x boost in performance when compared to CPU-based systems is what the Murex research and development lab has observed when executing XVA workloads in MX.3.

With Grace Hopper, Pricing FX Barrier Options in MX.3

Using its flagship and most recent stochastic local volatility model, Murex has priced foreign currency (FX) barrier options. It has also shown significant performance gains when utilising Grace Hopper. The payout of a barrier option is contingent upon the underlying asset price reaching or beyond a predetermined threshold over the option contract’s duration.

An Arm-based NVIDIA Grace CPU in the GH200 is more cost-effective for the 2D partial differential equation used in the pricing study. When using MX.3 on Grace Hopper, pricing this derivative is 2.3 times faster than when using an Intel Xeon Gold 6148 processor.Image Credit to NVIDIA

When it comes to FX barrier calculations, the NVIDIA Grace CPU provides 5 times better power efficiency per watt of power used.

Murex points to the aforementioned results to demonstrate how NVIDIA’s next-generation accelerated computing architecture is enabling energy efficiency and cost savings for high-performance computing for quantitative analytics in financial markets.

Nvidia grace hopper superchip price

NVIDIA doesn’t release Grace Hopper chip pricing, although several sources estimate system pricing:

Some Grace Hopper Superchip (GH200) system manufacturers offer prices. Example: Hyperscalers sells an NVIDIA Grace Hopper MGX Architecture machine for $65,400 [USD]. This pricing may contain the chip, memory, and other components, but it’s for the entire system.

Grace Hopper chips are likely more expensive than consumer-grade graphics cards because they target high-performance computing and data centres.

Alright, so I spent most of the day working on setting up unit testing on my home machine, so I could also run it locally as well. So far so good!

I don't have Jenkins setup yet, so I'm not going to begin to set that up yet, so the wait will be a bit longer before I start doing CI/CD completely.