Krisp is the global leader in background noise cancellation for online calls. The company emerged from UC Berkeley’s Skydeck accelerator in 2018 and rapidly reached over 100 employees by early 2021. After growing t’s user-base 20x in 2020, Krisp then closed a $14.5M Series A round.
The company applies proprietary Deep Learning algorithms to audio in real time to isolate human voice. Their AI-based system then removes unwanted background noise and echo from all participants in a conference call. The software is also optimized to require minimal computational overhead, so it can perform on any device including smartphones. Krisp currently features amongst the world’s most promising AI companies in the Forbes’ AI 50 list.
Krisp’s technology is based on deep neural networks and is trained using thousands of hours of purchased audio recordings. Like most AI start-ups, they began training on GPU-servers using the free computing credits provided by AWS, GCP and Azure. However, by the end of 2020, Krisp had run out. They required more cost-effective solutions. Krisp initially attempted to deal directly with individual data centers; but these could not offer the scalability and server choice needed to meet their ever evolving needs.
Hoping to find both a cost-efficient and scalable solution, Krisp reached out to FluidStack. During our initial discovery call, their CTO emphasized the need for a highly secure network to handle their training data. Indeed, the datasets used to develop their proprietary IP cannot legally be disclosed to 3rd parties under any circumstances. Additionally, Krisp performs data augmentation to enrich their audio datasets. This requires a combination of powerful GPUs, strong CPUs and high RAM specifications.
Following our initial discovery call, Krisp’s Head of Security was immediately added to our Slack workspace to accelerate onboarding. Working directly with our CTO, a solution was devised in under 5 days. Krisp’s engineering team were connected directly to our most secure Tier 4, 99.995% uptime data center partner in India. (One of only six organizations in the world certified by SAP® in infrastructure operations services.) This allowed Krisp to directly assess the data center’s internal network security as well as it’s access controls, data storage and encryption systems. Legally binding documents requested by Krisp were also duly signed by the data center partner and FluidStack.
Having demonstrated our data center’s security and capacity to match Krisp’s computing requirements, the following free, bare-metal server test server was provided for a week to assess performance:
- 1x Nvidia RTX 2080Ti 11GB,
- 1x Intel Xeon Silver 4208 (8 cores / 16 threads),
- 128GB RAM,
- 3.1TB SATA SSD,
- 300Mbps bandwidth.
Krisp’s testing revealed a stronger GPU performance than anticipated. CPU and RAM requirements, however, were higher than expected, so their engineering team requested higher specifications. Being a flexible startup with deep-rooted datacenter partnerships , we quickly catered to these needs, doubling each server’s CPU and RAM specifications (2x Intel Xeon Silver 4208 and 256GB RAM). This solution was found in less than 6 hours, meeting all of Krisp’s computing requirements. They then asked to deploy 4 such servers. We deployed each with our custom Deep Learning images to further accelerate their setup.
FluidStack’s GPU-servers now serve as key pillars to Krisp’s technological advancements and financial prosperity. Indeed, the results have been significant:
- Competitive rates - Krisp cut their GPU cloud computing costs by 76% while matching the same training speeds as those achieved on AWS. We understand that costs are measured in time spent switching between providers - as well as finances. So our tech team helped Krisp move to FluidStack in under 24 hours. Better yet, with our unique, all-inclusive, and transparent prices, they no longer need to account for data-transfer costs.
- High reliability - For the past 6 months, Krisp’s bare-metal RTX 2080Ti servers have experienced just 7 minutes of downtime. All penetration tests confirmed the high security level of our data centres.
- Dedicated support - By integrating Krisp’s engineers into FluidStack’s Slack support channel, any request from their team is answered on average in less than 17 minutes. A response-time twice as fast as AWS and GCP’s slow ticketing systems. And this is free!
- Bespoke solutions - Servers were tailored to meet Krisp’s high CPU and RAM requirements, with a second processor and additional RAM modules to each server. Krisp soon required a machine with more Tensor and CUDA cores for some extreme neural network training. They were able to test a wide variety of servers ranging from an RTX 3080 to Tesla A100 40GB. Following these tests, they elected to deploy 4x Tesla V100 16GB systems at another Tier 4 facility in Europe.
- Unlimited scalability - Satisfied by our server quality, network security and fast technical support, Krisp began scaling rapidly on short notice. Knowing that any delay in transferring data or setting up new environments was critical, we successfully deployed these new machines in less than 2 hours. We are now helping them shift towards the latest generation of NVidia cards including the RTX 3090 and RTX A6000.
What They Say
“Large GPU power is key to reduce time to market - for deep tech machine learning products like ours. However, GPU costs are also key. We found a great balance at FluidStack.
FluidStack specialists are always fast, they typically react in a dedicated Slack channel within a few
minutes for any request. Once we have an issue with a hardware failure and FluidStack managed to
move the server disk into an identical 4xA100 sever, thus reducing downtime to under an hour.
FluidStack staff, starting from co-founder, heads of IT and Operations, are extremely professional,
open, and willing to solve any issue in the best way for customers.
It’s always a pleasure to do business with them and see how rapidly FluidStack is growing!"
Tigran Sargsyan - Sr. Director of Engineering, ML at Krisp