Accelerated Computing for Cloud and On-Prem Data Centers

Written by Matthew Williams, CTO at Cerio

As workloads evolve, our traditional generalized approaches to system design are no longer sufficient for diverse workload types. So, how do we build systems tailored to our needs, and precisely when we need them?  

The Cerio Accelerated Computing Platform is designed to solve the structural inefficiencies in how server systems are designed and built – and it doesn’t matter if that server is in the cloud, on-prem or a hybrid, the system itself is a restriction.   

On-prem data centers 

Many organizations want cloud-like flexibility without the high cost and security implications of moving their data offsite. In fact, 42% of organizations keep critical workloads off the cloud due to security concerns.  

While the cloud offers the advantage of being able to dynamically acquire compute, servers and GPUs can be very costly to rent on a per-hour basis. What often happens is that these cloud instances keep running and suddenly the budget has been blown. 

On-prem provides more control over spending and security, but it’s difficult to get any level of agility because GPUs are deployed within a fixed system model. This leads to stranded assets which restrict accelerator availability and over-provisioning. When you put GPUs in certain servers, and don’t always use them all, or have the wrong kind of accelerator in the wrong server – not only are the accelerators stranded and driving over provisioning, but the CPUs in those servers and the entire server itself becomes idle and stranded.  

If the accelerator deployed within a server is focused on a subset of use cases, you’re not going to use that server for any other kind of workload, resulting in an expensive asset sitting there delivering zero value. With Cerio, you can add a wide variety of accelerators to your pool of accelerators and assign them dynamically to any server, avoiding needing a dedicated server for each type and having those servers sit idle when those accelerators aren’t being used. 

You get the best of both worlds in that you can be hybrid; you can experiment in the cloud and use cloud services to get the right system with the option to move production back into on-prem without losing agility. You can now have the right system design based on the actual workload and job that you’re running. 

Cloud service providers 

Cloud service providers can also use the Cerio Accelerated Computing Platform to build base systems of physical infrastructure upon which they can spin up a very large variety of instance types or products. Today you may have a small catalogue of instance types – one, two, or four GPUs with fixed memory and CPU ratios – but now you can offer people a variety of GPU instances with different vendors and accelerator models and dynamically move those accelerators into the right server to support the mix of instance types currently required by your customers. 

Like on–prem data centers, stranded assets are a major problem – you don’t want an expensive resource sitting idle. If you have a fixed system model, you can very easily run out of CPU cores or memory. Being able to move GPUs to the right system with the right resources for that VM is important. Anytime you don’t have access to the right resources within the right system, you end up with an expensive asset not making you money because you’re not able to sell services to your customers.  

People who go to the cloud want choices – they want to be able to pick the right combinations, and not be limited to a single vendor’s type of accelerator. There are hundreds of accelerator companies with innovative technology that can provide real value, but it’s hard for service providers to consume it. Previously, cloud service providers couldn’t afford to build a system around a custom accelerator or new accelerator that’s just entering the market, but with Cerio, cloud providers can provide that unique environment with a vendor’s technology that is otherwise not available. Cerio makes it possible to create the right kind of system on the fly, with exactly the right number of GPUs and the right amount of system memory. Those same GPUs can then be reallocated in a different ratio to create another system to fit a different workload. 

As a cloud service provider, you can sell more services to people taking a hybrid approach because you have a greater selection of products and instance types to leverage to make improvements. The better you can tune products to what customers want, the more money you’re going to make.  

Now you can create systems on top of virtual infrastructure and offer a much wider flexibility to end users. On-prem users can use Cerio’s platform to get cloud-like agility, while cloud service providers can create very customized solutions that are truly optimal for their customers. 

The Cerio Accelerated Computing Platform addresses the structural inefficiencies inherent in server system design, regardless of whether it lives in the cloud, on-prem, or in a hybrid environment, effectively surpassing the limitations imposed by the system itself.