David Driggers is the Chief Technology Officer at Cirrascale Cloud Services, a leading provider of deep learning infrastructure solutions. Guided by values of integrity, agility, and customer focus, Cirrascale delivers innovative, cloud-based Infrastructure-as-a-Service (IaaS) solutions. Partnering with AI ecosystem leaders like Red Hat and WekaIO, Cirrascale ensures seamless access to advanced tools, empowering customers to drive progress in deep learning while maintaining predictable costs.
Cirrascale is the only GPUaaS provider partnering with major semiconductor companies like NVIDIA, AMD, Cerebras, and Qualcomm. How does this unique positioning benefit your customers in terms of performance and scalability?
As the industry evolves from Training Models to the deployment of these models called Inferencing, there is no one size fits all. Depending upon the size and latency requirements of the model, different accelerators offer different values that could be important. Time to answer, cost per token advantages, or performance per watt can all affect the cost and user experience. Since Inferencing is for production these features/capabilities matter.
What sets Cirrascale’s AI Innovation Cloud apart from other GPUaaS providers in supporting AI and deep learning workflows?
Cirrascale’s AI Innovation Cloud allows users to try in a secure, assisted, and fully supported manner new technologies that are not available in any other cloud. This can aid not only in cloud technology decisions but also in potential on-site purchases.
How does Cirrascale’s platform ensure seamless integration for startups and enterprises with diverse AI acceleration needs?
Cirrascale takes a solution approach for our cloud. This means that for both startups and enterprises, we offer a turnkey solution that includes both the Dev-Ops and Infra-Ops. While we call it bare-metal to distinguish our offerings as not being shared or virtualized, Cirrascale fully configures all aspects of the offering including fully configuring the servers, networking, Storage, Security and User Access requirements prior to turning the service over to our clients. Our clients can immediately start using the service rather than having to configure everything themselves.
Enterprise-wide AI adoption faces barriers like data quality, infrastructure constraints, and high costs. How does Cirrascale address these challenges for businesses scaling AI initiatives?
While Cirrascale does not offer Data Quality type services, we do partner with companies that can assist with Data issues. As far as Infrastructure and costs, Cirrascale can tailor a solution specific to a client’s specific needs which results in better overall performance and related costs specific to the customer’s requirements.
With Google’s advancements in quantum computing (Willow) and AI models (Gemini 2.0), how do you see the landscape of enterprise AI shifting in the near future?
Quantum Computing is still quite a way off from prime time for most folks due to the lack of programmers and off-the-shelf programs that can take advantage of the features. Gemini 2.0 and other large-scale offerings like GPT4 and Claude are certainly going to get some uptake from Enterprise customers, but a large part of the Enterprise market is not prepared at this time to trust their data with 3rd parties, and especially ones that may use said data to train their models.
Finding the right balance of power, price, and performance is critical for scaling AI solutions. What are your top recommendations for companies navigating this balance?
Test, test, test. It is critical for a company to test their model on different platforms. Production is different than development—cost matters in production. Training may be one and done, but inferencing is forever. If performance requirements can be met at a lower cost, those savings fall to the bottom line and might even make the solution viable. Quite often deployment of a large model is too expensive to make it practical for use. End users should also seek companies that can help with this testing as often an ML Engineer can help with deployment vs. the Data Scientist that created the model.
How is Cirrascale adapting its solutions to meet the growing demand for generative AI applications, like LLMs and image generation models?
Cirrascale offers the widest array of AI accelerators, and with the proliferation of LLMs and GenAI models ranging both in size and scope (like multi-modal scenarios), and batch vs. real-time, it truly is a horse for a course scenario.
Can you provide examples of how Cirrascale helps businesses overcome latency and data transfer bottlenecks in AI workflows?
Cirrascale has numerous data centers in multiple regions and does not look at network connectivity as a profit center. This allows our users to “right-size” the connections needed to move data, as well as utilize more that one location if latency is a critical feature. Also, by profiling the actual workloads, Cirrascale can assist with balancing latency, performance and cost to deliver the best value after meeting performance requirements.
What emerging trends in AI hardware or infrastructure are you most excited about, and how is Cirrascale preparing for them?
We are most excited about new processors that are purpose built for inferencing vs. generic GPU-based processors that luckily fit quite nicely for training, but are not optimized for inference use cases which have inherently different compute requirements than training.
Thank you for the great interview, readers who wish to learn more should visit Cirrascale Cloud Services.
Credit: Source link
