In an office building in Austin, Texas that goes unnoticed, two tiny rooms are occupied by a small number of Amazon personnel who are engineering two varieties of microchips for training and increasing generative AI. These personalized chips, namely Trainium and Inferentia, provide AWS customers a different option to training their comprehensive language models using Nvidia GPUs, which have rapidly become difficult and pricey to acquire. Adam Selipsky, Chief Executive Officer of Amazon Web Services, stated in a June interview with CNBC: "The entire world would prefer access to more chips for doing generative AI, be it GPUs or Amazon's own chips that we're designing. We are in a better position than anyone else on the planet to give the power that our customers in total are wanting." Regardless of this, others have acted faster and invested more to take advantage of the generative AI surge. When OpenAI released ChatGPT in November, Microsoft got attention for hosting the popular chatbot as well as investing an estimated $13 billion in OpenAI. It then swiftly decided to utilize generative AI models in its own products, including Bing, in February. The same month, Google introduced its own large language model called Bard, followed by a $300 million investment in Anthropic, OpenAI's rival. It was only in April that Amazon announced its line of extensive language models named Titan and a service called Bedrock to help developers enhance software by using generative AI. Chirag Dekate, VP analyst at Gartner, commented: "Amazon is not accustomed to chasing markets. Amazon usually creates markets. I think for the first time in a long time, they find themselves behind and are working to catch up." Recently, Meta also unveiled its own LLM, Llama 2. The ChatGPT open-source counterpart is now available for users to try on Microsoft's public cloud, Azure.
In the future, according to Dekate, Amazon's purpose-built silicon could give it an advantage in generative AI. "The real separator is the technical proficiency that they acquired. Microsoft, of course, does not own Trainium or Inferentia," he added. Before this year began, AWS had already started making its own silicon with Nitro, its specially designed hardware. CNBC has reported that each AWS server holds at least one of these chips, and more than 20 million are in use around the world.
In 2015, Amazon acquired Israeli chip startup Annapurna Labs. This ultimately led to the introduction of Graviton, an Arm-based server processor rivaling x86 CPUs from the likes of AMD and Intel. According to Stacy Rasgon, senior analyst at Bernstein Research, "probably high single-digit to maybe 10% of total server sales are Arm, and a good chunk of those are going to be Amazon. So on the CPU side, they've done quite well". Around the same time, Amazon unveiled its AI-oriented chip. This followed Google's revelation of its first Tensor Processor Unit (TPU) two years prior, and preceded Microsoft's announcement of Athena AI chip, which has been said to have been developed in partnership with AMD.
CNBC obtained access to a tour of Amazon's chip lab in Austin, Texas, where Trainium and Inferentia are crafted and examined. VP of product Matt Wood specified what these chips are designed for: "Machine learning breaks down into these two different stages. So you train the machine learning models and then you run inference against those trained models". Trainium was launched in 2021, following Inferentia in 2019. Wood declared that Trainium enables customers to provide "very, very low-cost, high-throughput, low-latency, machine learning inference".
At present, Nvidia's GPUs are still the preeminent technology when it comes to training models. In July, AWS released new AI acceleration hardware powered by Nvidia H100s. Reflecting on this, Rasgon commented that "Nvidia chips have a massive software ecosystem that's been built up around them over the last like 15 years that nobody else has. The big winner from AI right now is Nvidia.
AWS' strong cloud presence has set them apart from the competition. According to technology industry researcher Gartner, AWS was the world's biggest cloud computing provider in 2022, commanding 40% of the market share, with their operating income being lower year-over-year for three quarters in a row, however, AWS still accounted for 70% of Amazon's overall operating profit of $7.7 billion in the second quarter. With AWS having larger operating margins than those of Google Cloud, they have been building up a portfolio of developer tools focused on generative AI. Bedrock gives customers access to models from multiple providers as Sivasubramanian, AWS' VP of database, analytics and machine learning said, “We don't believe that one model is going to rule the world, and we want our customers to have the state-of-the-art models from multiple providers because they are going to pick the right tool for the right job.” With millions of AWS customers being familiar with the platform, it is a question of velocity to see how quickly they can benefit from generative AI applications.
Amazon has recently unveiled AWS HealthScribe, a service which uses generative AI to help doctors create patient visit summaries. The company's SageMaker machine learning hub provides algorithms and models, and CodeWhisperer coding companion has reportedly enabled developers to complete tasks 57% faster. AWS has additionally established a $100 million "center" for generative AI. CEO Jassy is personally overseeing the development of expansive language models, and more than 20 machine learning services have been offered. These services are being used by entities such as Philips, 3M, Old Mutual and HSBC. The vast growth of AI has sparked worries that confidential information might be exposed in the training data of public large language models. To address such concerns, AWS's Bedrock service will guarantee that anything done through Bedrock will remain encrypted and secure in a private cloud environment. According to CNBC, more than 100,000 customers are currently using AWS's machine learning services, and analysts believe this number could potentially increase as Amazon customers tend to explore its ecosystems extensively.
コメント