SRAM has many advantages as the fastest read/write memory mediumFor example, it can be compatible with more advanced process nodes, with the computing advantages of high energy efficiency, without the endurance issues of other non-volatile storage, and more importantly,The SRAM-based design solution can support purely digital design, which can solve the accuracy problem of many application scenarios.
The SRAM solution has three advantages at this stageSRAM is the fastest of all storage types and has no limit on the number of writes, making it almost mandatory for scenarios that pursue fast response. SRAM is compatible with advanced manufacturing processes, resulting in higher energy efficiency and areal efficiency; Compared with new types of memory, SRAM has a high degree of process maturity and can achieve technology implementation and mass production relatively quickly.
There are a variety of technical paths for in-memory computing based on different storage mediaSuch as SRAM, Flash, and other new types of memory. The SRAM process has a high degree of maturity, high compatibility with advanced nodes, and unlimited erase and write times, and can achieve lossless accuracy and short read and write latency during computing, which is suitable for scenarios such as autonomous driving and drones that require high computing accuracy and response speed.
18 times faster than GPT-4, the world's fastest large model GROQ is here! 500 tokens per second broke the record, and the self-developed LPU is 10 times that of NVIDIA GPU
chatgpt-3.5. The generation rate is only 40 tokens per second.
Some netizens compared it with GPT-4 and Gemini to see how long it takes for them to complete a simple debugging problem.
Unexpected,The GroQ completely crushes both, delivering 10 times faster output than Gemini and 18 times faster than GPT-4. (In terms of the quality of the answers, though, gemini is better.) ）
Unlike NVIDIA GPUs, which rely on high-speed data transmission, GroQ's LPUs do not use high-bandwidth memory (HBM) in their systems.
It uses SRAM, which is about 20 times faster than the memory used by the GPU.
In contrast, you only need an H200 (equivalent to the density of 1 4 server racks) to run these models fairly efficiently.
This configuration may perform well in scenarios where only one model needs to be run and there are a large number of users. However, once you need to run multiple models at the same time, especially if you need to do a lot of model fine-tuning or use high-level LoRa, etcThis configuration is no longer applicable.
To sum up, the author has also sorted out the 5 leading enterprises of the "SRAM" concept, which are worthy of everyone's attention and collection!
1. Beijing Junzheng
The company's memory chips are divided into three categories: SRAM, DRAM and FLASH, which are mainly for the automotive, industrial, medical and other industry markets and high-end consumer markets. Groq's large-model inference chip is making a splash at 500 tokens per second, surpassing traditional GPUs and Google TPUs. The GroQ chip is manufactured on a 14nm process and is equipped with a large 230MB SRAM to ensure memory bandwidth, and the on-chip memory bandwidth is up to 80TBS.
2. Ninestar
Ninestar and the Chinese Academy of Sciences jointly developed and mass-produced China's first radiation-proof series of chips - phase change memory (PCRAM) products. The company's main business: research, development, production and sales of integrated circuit chips, universal printing consumables and core components and recycled printing consumables.
3. China Power Port
The company is the largest distributor of electronic components in China. According to the prospectus, the company has the products of the world's leading memory chip manufacturers such as Micron (Micron) and Yangtze River Storage in the memory product market.
Fourth, Guangli Technology
As an international high-tech enterprise based in China, the company has focused on the layout and development of emerging businesses of semiconductor packaging and testing equipment in the past five years, and is committed to becoming a world-class semiconductor equipment enterprise.
5. C Huawei
At present, the company's main products cover two major fields: digital and analog integrated circuits, of which digital integrated circuit products include logic chips, memory chips and microcontrollers represented by programmable logic devices (CPLD FPGAs); Analog integrated circuit products include data conversion (ADC DAC), bus interfaces, and power management.
Send a 6 to the blogger [Envelope] and get it for free.