Because the world rushes to utilize the most recent wave of AI applied sciences, one piece of high-tech {hardware} has turn into a surprisingly scorching commodity: the graphics processing unit, or GPU.
A top-of-the-line GPU can promote for tens of thousands of dollars, and main producer Nvidia has seen its market valuation soar past $2 trillion as demand for its merchandise surges.
GPUs aren’t simply high-end AI merchandise, both. There are much less highly effective GPUs in telephones, laptops, and gaming consoles, too.
By now you’re in all probability questioning: What’s a GPU, actually? And what makes them so particular?
What Is a GPU?
GPUs had been initially designed primarily to shortly generate and show complicated 3D scenes and objects, akin to these concerned in video video games and computer-aided design software program. Fashionable GPUs additionally deal with duties akin to decompressing video streams.
The “mind” of most computer systems is a chip known as a central processing unit (CPU). CPUs can be utilized to generate graphical scenes and decompress movies, however they’re sometimes far slower and fewer environment friendly at these duties in comparison with GPUs. CPUs are higher suited to basic computation duties, akin to phrase processing and looking net pages.
How Are GPUs Totally different From CPUs?
A typical trendy CPU is made up of between 8 and 16 “cores,” every of which might course of complicated duties in a sequential method.
GPUs, alternatively, have hundreds of comparatively small cores, that are designed to all work on the similar time (“in parallel”) to realize quick general processing. This makes them well-suited for duties that require numerous easy operations which could be completed on the similar time, reasonably than one after one other.
Conventional GPUs are available two most important flavors.
First, there are standalone chips, which frequently are available add-on playing cards for giant desktop computer systems. Second are GPUs mixed with a CPU in the identical chip package deal, which are sometimes present in laptops and sport consoles such because the PlayStation 5. In each circumstances, the CPU controls what the GPU does.
Why Are GPUs So Helpful for AI?
It seems GPUs could be repurposed to do greater than generate graphical scenes.
Most of the machine studying methods behind synthetic intelligence, akin to deep neural networks, rely closely on varied types of matrix multiplication.
This can be a mathematical operation the place very massive units of numbers are multiplied and summed collectively. These operations are well-suited to parallel processing and therefore could be carried out in a short time by GPUs.
What’s Subsequent for GPUs?
The number-crunching prowess of GPUs is steadily rising because of the rise within the variety of cores and their working speeds. These enhancements are primarily pushed by enhancements in chip manufacturing by corporations akin to TSMC in Taiwan.
The scale of particular person transistors—the fundamental elements of any pc chip—is lowering, permitting extra transistors to be positioned in the identical quantity of bodily area.
Nevertheless, that isn’t the complete story. Whereas conventional GPUs are helpful for AI-related computation duties, they don’t seem to be optimum.
Simply as GPUs had been initially designed to speed up computer systems by offering specialised processing for graphics, there are accelerators which might be designed to hurry up machine studying duties. These accelerators are also known as information middle GPUs.
Among the hottest accelerators, made by corporations akin to AMD and Nvidia, began out as conventional GPUs. Over time, their designs advanced to raised deal with varied machine studying duties, for instance by supporting the extra environment friendly “brain float” quantity format.
Different accelerators, akin to Google’s tensor processing units and Tenstorrent’s Tensix cores, had been designed from the bottom in control up deep neural networks.
Information middle GPUs and different AI accelerators sometimes include considerably extra reminiscence than conventional GPU add-on playing cards, which is essential for coaching massive AI fashions. The bigger the AI mannequin, the extra succesful and correct it’s.
To additional velocity up coaching and deal with even bigger AI fashions, akin to ChatGPT, many information middle GPUs could be pooled collectively to kind a supercomputer. This requires extra complicated software program to correctly harness the obtainable quantity crunching energy. One other strategy is to create a single very massive accelerator, such because the “wafer-scale processor” produced by Cerebras.
Are Specialised Chips the Future?
CPUs haven’t been standing nonetheless both. Current CPUs from AMD and Intel have built-in low-level directions that velocity up the number-crunching required by deep neural networks. This extra performance primarily helps with “inference” duties—that’s, utilizing AI fashions which have already been developed elsewhere.
To coach the AI fashions within the first place, massive GPU-like accelerators are nonetheless wanted.
It’s attainable to create ever extra specialised accelerators for particular machine studying algorithms. Lately, for instance, an organization known as Groq has produced a “language processing unit” (LPU) particularly designed for operating massive language fashions alongside the traces of ChatGPT.
Nevertheless, creating these specialised processors takes appreciable engineering assets. Historical past exhibits the utilization and recognition of any given machine studying algorithm tends to peak after which wane—so costly specialised {hardware} could turn into shortly outdated.
For the typical client, nonetheless, that’s unlikely to be an issue. The GPUs and different chips within the merchandise you employ are prone to preserve quietly getting quicker.
This text is republished from The Conversation underneath a Inventive Commons license. Learn the original article.
Picture Credit score: Nvidia
