SambaNova claims AI performance comparable to Nvidia and announces offering as a service

SambaNova performs various deep learning tasks such as natural language processing tasks in neural networks with Google’s BERT-Large.

Samba Nova system

The still very young market for artificial intelligence computers is creating interesting business models. On Wednesday, Palo Alto-based startup Samba Nova Systems, which received about $ 500 million in venture capital, announced the general availability of its dedicated AI computer, DataScale, as well as a service to offer. Place the machine in a data center and rent its capacity for $ 10,000 per month.

“This is a quick and easy way to access and use DataScale products as a service with an entry price of $ 10,000 per month,” said Marshall Choy, Vice President of Products at Samba Nova, in the following interview. ZDNet Via video.

“We have one or more racks in the data center to own, manage, and support the hardware, so we can actually use this product as a service.” The managed service is called dataflow-as-a-service. Will be. This is because it emphasizes the marketing of hardware and software rerouting itself based on the flow of AI models built into the system.

DataScale computers compete with Nvidia’s graphics chips, which dominate the training of neural networks.

“This is a change not just for AI, but for the entire computing industry,” said Samba Nova CEO.

Like other startups Graphcore and Cerebras Systems, Samba Nova takes a system approach to AI and does not compete with Nvidia just to sell cards, but custom chips, firmware, software, data and memory I / O subsystems. I built a complete finished machine with. Even Nvidia recently announced a dedicated AI appliance computer.

The DataScale system is touted as comparable to Nvidia’s DGX-2 rack-mounted system 64 running A100 GPUs, but only a quarter of a standard telco rack.

The computer uses a custom chip with reprogrammable logic called the Reconfigurable Data Unit (RDU). There is a unique software system called SambaFlow for laying out convolutions or other deep learning operations in a way that uses multiple RDUs. It has a high-speed fabric for connecting RDUs.

SambaNova has also rewritten some core applications, including natural language processing, to make it more efficient in benchmark tests.

“We consider this to be one of the biggest data center migrations we’ve seen for a very long time,” Rodrigo Liang, co-founder and CEO of the company, said in the same video session. ZDNet I talked to Liang in February when the details of the machine were still obscured. Liang reiterated February’s claim that Samba Nova’s focus is to influence broad and deep changes throughout computing.


Rodrigo Liang, co-founder and CEO of Samba Nova Systems, along with co-founder, is excited to be in a unique position to do this because he “owns all layers of the stack.” “. Kunle Orkotun on the left and Christopher Re on the right.


But at this point, the emphasis is, as Choy says, to make AI “quick and easy” to use.

“We are excited because we have ownership of all layers of the stack and are in a unique position to be able to do this,” says Liang. “We are not only building chips that will be built into someone else’s system, but also integrating software to build everything down to the rack, and then applications.”

As an example of out-of-the-box ease of use, Samba Nova claims better benchmarking results compared to Nvidia. For example, when training Google’s BERT-Large natural language neural network, which is a version of the popular Transformer language model, Samba Nova has a throughput of 28,800 samples per second, compared to 20,086 for Nvidia’s A100-based DGX. Claims to be sample / sec.

This week, the company, which hosts a software tools developer event, has adopted pre-built models such as Hugging Face, one of the most popular chatbots, and offers a pre-trained version that can be downloaded and run. I am. On a Samba Nova machine.

“This problem can be solved in a time that cannot be achieved by the number of GPUs and CPUs,” startup Cerebras tells the Supercomputing Conference.

“With Samba Flow, you can use these existing models to get cutting-edge results in seconds,” says Liang.

Choy says the standalone pricing for owning a DataScale is comparable to Nvidia’s DGX-2.

An early customer who purchased the system completely was Argonne National Laboratory. The Institute, which is part of the US Department of Energy, has worked with Samba Nova on some of the kind of huge projects that are the mission of DoE’s National Institute, including the COVID-19 study.

“Samba Nova is designed in several ways to summarize the performance people usually see from the GPU,” Rick Stevens, Associate Travo Director of Argonne, said in an interview. ZDNet Via video. “You can think of it as scaling up and down the GPU in a very efficient way.”


DataScale consists of multiple custom reconfigurable processors called RDUs that are combined on a special high-speed fabric.

Samba Nova system.

“It also has so much memory that you can train models that don’t fit on the GPU,” says Stevens, who says that the internal architecture that interconnects processors and interleaves memory over time. There is an expanding “headroom”.

Stevens said the lab still understands that some deep learning neural networks have advantages over other networks on the machine. Argonne has a number of problems, especially in cancer research, astronomy, and fusion reactors. They utilize a variety of neural networks, including convolutional neural networks and the form of generated networks called “tomography GANs”.

“It’s definitely better than the GPU on these issues,” Stevens said. The main reason is that, like GPUs, there is no memory plateau that reaches the memory limit of the GPU card. “Samba Nova makes it much smoother and allows you to explore a much wider number of model parameters.” Also, the large memory means that you don’t have to explicitly split your code into parallel operations. To do. This is a particularly burdensome development task.

“We are in the process of learning how to evolve neural architectures to get the most out of our hardware, which can take a year,” he said. Large memory means that not only very large linguistic neural networks such as large transformers are particularly beneficial, but also very large generative models are available. He mentioned the vector quantization autoencoder as another potential beneficiary.

Argonne evaluates SambaNova, Cerebras, and other AI accelerators, with the ultimate goal of making the system available to various collaborators around the world. Stevens predicts that technologies like Samba Nova and accelerators for Cerebras, Graphcore, Grok, or Intel’s Habana units will be built together in exascale systems.

“Large machines of the future may have AI complexes as part of their procurement,” Stevens said. “We are actively working on what we will deploy as a large-scale computing resource for DoE over the next five years,” Stevens said. “And one of the questions is what is the combination of architectures for that?”

“They are just building blocks,” said Stevens of various accelerators, including Samba Nova. “Challenge the integrator to think about how to build a system that connects them. If you want to drive an AI engine from an AI application running on a smaller part of your machine, training or reasoning,” he said. It was. “That’s where things are going.”

Stevens talked with ZDNet earlier this year about Argonne’s work to speed up COVID-19 research using the Cerebras computer CS-1. When asked how to compare the two machines, Stevens refused to compare Samba Nova and Cerebras.

SambaNova claims AI performance comparable to Nvidia and announces offering as a service

Source link SambaNova claims AI performance comparable to Nvidia and announces offering as a service

Back to top button