Intel’s former CEO deposits money to a less known hardware initiative who wants to restore Nvidia
- It is supported by the UK -based Fraktile NATO and wants to create faster and cheaper memory artificial intelligence
- Nvidia’s BruteForce GPU approach consumes a lot of power and is kept back with memory
- Fractile’s numbers focused on the set of H100 GPU comparison, not the main current H200
Nvidia sits on the AI hardware food chain, high -performance GPUs and high -performance GPUs and CUDA software piles, which have become the default tools for large AI models – but this dominance is a cost – that is to grow on its back.
Hyper scales such as Amazon, Google, Microsoft and Meta are shedding resources to improve their own special silicones to reduce their dependence on NVIDIA’s chips and reduce costs. At the same time, the wave of an AI hardware initiative is trying to increase the increasing demand for specialized accelerators, hoping to provide more efficient or affordable alternatives and ultimately displace Nvidia.
You may not have heard the UK -based Fracilil Nevertheless, the initiative, which claims the revolutionary approach to the information process, may have some remarkable supporters, including the cost of the world’s largest language models 100x faster and at 1/10, including NATO and former Intel CEO Pat Gelsinger.
Removing every bottleneck
“We are building hardware that will raise the fastest possible inference of the largest transformer networks, Frağ says Fractile, Fractile.
“This means a universe with completely new talents and possibilities for how we will work with the world’s largest LLMs and superhuman intelligence models that will work almost than you can read.”
Before being very excited, it should be noted that Fractile’s performance numbers are based on comparisons with NVIDIA H100 GPU clusters using 8-bit quantitation and tensorrt-llm.
In a LinkedIn shipment, Gelsinger, who recently joined the VC company Playground Global as a general partner, wrote: “The removal of Frontier AI models is bottled by the hardware. Before the scaling of the test-time calculation, the cost and delay were great difficulties for the distribution of large-scale LLM.
Orum I am pleased to share that I invest in Fractile, an AI hardware company established by the UK, which has recently followed a radical path to offer such a leap, ”he said.
“In -memory calculation approaches for inference acceleration, together with two bottles, overcome both memory bottlenecks that protect today’s GPUs, the biggest physical restriction we encounter when it increases the capacity of the data center, in a luxurious study at Stanford University, some of the ideas come from the exploration of some of the ideas.
You can also like
