sdecoret - stock.adobe.com
With a surge in popularity and advancement, AI hardware has become a competitive market. AI hardware companies must have quick turnarounds in product advancements to have the newest and most effective products on the market.
Although these eight AI hardware companies focus on CPUs and data center technologies, their specializations have slowly broadened as the market expands. Now, these companies are competing to create the most powerful and efficient AI chip on the market.
Nvidia became a strong competitor in the AI hardware market when its valuation surpassed $1 trillion in early 2023. The company's current work includes its $10,000 A100 chip and Volta GPU for data centers. Both products are critical technologies for resource-intensive models. Nvidia also offers AI-powered hardware for the gaming sector.
In August 2023, Nvidia announced its latest technological breakthrough with the world's first HBM3e processor. It unveiled the Grace Hopper platform, a superchip with three times the bandwidth and over three times the memory capacity of the current generation's technology. Its design is focused on heightened scalability and performance in the age of accelerated AI.
The company's NVLink technology can connect the Grace Hopper superchip to other superchips. NVlink enables multiple GPUs to communicate through high-speed interconnection.
Intel has made a name for itself in the CPU market with its AI products. Although it hasn't outpaced Nvidia in GPUs, it is a leader in its niche. For example, the 2022 rollout of its Xeon processors made it stand out against its competitors.
Intel's Xeon Platinum series is a CPU with built-in acceleration. The series has nearly three times as much memory capacity and almost two times as much bandwidth as the previous generation.
Lately, Intel has been attempting to gain an edge in the broader AI hardware market. While Intel's CPUs run 70% of data center inferencing globally, the company is looking to expand.
Alphabet, Google's parent company, has various products for mobile devices, data storage and cloud infrastructure. Cloud TPU v5e is purpose-built for large language models and generative AI that's half the cost of the previous generation. Models process data five times faster when running on Cloud TPU v5e.
Alphabet has focused on producing powerful AI chips to meet the demand for large-scale projects. It has also unveiled Multislice, a performance-scaling technology. While hardware limitations typically restrict software scaling, runs show nearly 60% utilization of model floating-point operations per second on multibillion parameter models on TPU v4.
Apple's Neural Engine, specialized cores based on Apple chips, has furthered the company's AI hardware design and performance. The Neural Engine led to the M1 chip for MacBooks. Compared to the generation before, MacBooks with an M1 chip are 3.5 times faster in general performance and five times faster with graphic performance.
After the success of the M1 chip, Apple announced the next generation. It released M2-powered products, boasting a better battery life and performance than devices with the previous iteration. Apple plans to release the M3 as early as 2024.
After the success of its first specialized AI chip, Telum, IBM set out to design a powerful successor to rival its competitors.
In 2022, IBM created the Artificial Intelligence Unit. The AI chip is purpose-built and runs better than the average general-purpose CPU. It has over 23 billion transistors and features 32 processing cores. It's far less memory-intensive and is more efficient than previous generations.
Although Qualcomm is relatively new in the AI hardware market compared to its counterparts, its experience in the telecom and mobile sectors makes it a promising competitor.
Qualcomm's Cloud AI 100 chip beat the Nvidia H100 in a series of tests. One test was to see the number of data center server queries each chip could carry out per watt. Qualcomm's Cloud AI 100 chip totaled 227 server queries per watt, while Nvidia H100's hit 108. The Cloud AI 100 chip also managed to net 3.8 queries per watt compared to the Nvidia H100's 2.4 queries during object detection.
AWS has switched its focus from cloud infrastructure to chips. Its Elastic Compute Cloud Trn1s are purpose-built for deep learning and large-scale generative models. They use AWS Trainium chips, AI accelerators, to function.
Trn1.2xlarge was the first iteration. It only had one Trainium accelerator, 32 GB instance memory and 12.5 GBps network bandwidth. Now Amazon has the trn1.32xlarge instance, which has 16 accelerators, 521 GB of memory and 1,600 GBps bandwidth.
AMD created the next generation of Epyc processors. The company released its latest product, the Zen 4 CPU, in 2022.
The company plans to release the MI300 AI chip by 2024. AMD's MI300 and Nvidia's H100 rival in memory capacity and bandwidth.
Devin Partida is the editor in chief of ReHack.com and a freelance writer. She has steadily increased her knowledge of niches such as biztech, medtech, fintech, IoT and cybersecurity.