Synthara offers ‘easy transition’ compute-in-memory
Synthara AG (Zug, Switzerland), a 2019 spin off from the Institute of Neuroinformatics (University of Zurich and ETH Zurich), is claiming it has a superior compute-in-memory architecture.
The technique is generally applicable, but is particularly relevant to compute-intensive applications such as machine learning and AI, the company said. “It’s a good solution for LLMs [large language models],” CEO Manu Nair told eeNews Europe.
The company claims that the integration of its technology into an established computational logic architecture can produce a 50x improvement in performance, without requiring alterations to the general chip architecture or to software flows.
Synthara has disclosed that Bosch is an early client. It also has Sean Mitchell, co-founder and CEO of the early AI and vision processing company Movidius, as its chairman of the board. Movidius was sold to Intel for about €300 million in 2016.
Manu Nair, co-founder and CEO of Synthara AG.
Founded by Nair and Alessandro Aimar – who serves the company as CTO – Synthara has developed a method of inserting in-memory computation into pre-existing hardware-software platforms and supporting legacy applications. The technology can also be used to develop hardware-software platforms from scratch.
In-memory compute is something the electronics industry is going to have to adopt for reasons of energy efficiency, Nair said. “The bottleneck in almost all computation systems is the movement of data, so the approach here is to reduce data movement,” said Nair. The key thing about Synthara’s offering is it allows companies to transition effectively including supporting legacy software.
Synthara’s offering is in the form of semiconductor IP called ComputeRAM and an accompanying software addition to the development tool chain.
Agnosticism
ComputeRAM is currently based on SRAM but is agnostic to the memory type. It could be applied to ReRAM or other memory types. The semiconductor IP makes use of a memory array with additional peripheral circuitry that adds about 5 to 10 percent to the area, allowing the memory to perform math functions. It is a fully CMOS-compatible process.
“The ComputeRAM system offers a slightly different abstraction than a multiply accumulate. Its lower computational primitive is that of a dot product,” said Nair. He added: “We prefer to work with memory bit-cells provided by the foundry. We can also work with compiled memory arrays,” said Nair.”We will be creating a ComputeRAM compiler.”
The software addition into the tool-chain is also key. “We automate the partition of algorithms and the synchronization, solving the problem for the systems integrator,” said Nair.
“Synthara’s innovation is in the architecture that allows us to make in-memory computation a programming problem rather than a hardware problem. The technology is processor type agnostic – we can support RISC-V, ARM or x86. It is memory technology agnostic. Our software stack enables the programmer to both run old and new algorithms without rewriting any ComputeRAM-aware code,” the company states on its website.
The technology is also data-type agnostic. “We don’t yet support full floating-point but integer datatypes and blockFP we do support,” said Nair. “Also, it is run-time dynamically configurable.”
Test chip
Synthara has produced an SRAM-based ComputeRAM test chip on the 22nm 22FDX manufacturing process from GlobalFoundries Inc. Nair said that ComputeRAM does not require the body-biasing capabilities of the fully-depleted silion-on-insulator (FDSOI) 22FDX process. The test chip is intended as a demonstrator with samples going out to a few customers.
Nair said he expects to be porting ComputeRAM to a process with foundry TSMC by the end of 2024 although the selection of the particular process will depend on lead customers’ requirements.
Synthara raised about CHF 3.1 million (about US$3.5 million) in seed funding and has been supported with about CHF 5.5 million (about US$6.3 million) of support in the form of grants from the European Union and Swiss organizations. The company is currently fundraising
Related links and articles:
News articles:
Edge AI processor tackles demands of Generative AI, video and LLMs
In-memory AI startup EnCharge raises another US$22.6 million
Tachyum cuts LLMs costs up to 100x to bring them mainstream
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
