Blueshift Memory: accelerating memory access

Blueshift Memory is a proprietary new architecture design enabling up to 1,000 times faster memory access.

Its proprietary chip design optimises the memory architecture for more efficient handling of large data sets and time-critical data, enabling up to 1,000 times faster memory access for specific data[1]focused applications. These include high performance computing, artificial intelligence (AI), machine vision for augmented and virtual reality (AR/VR), 5G Edge connectivity, and the Internet of Things (IoT).

The problem

“The challenge that I set out to solve at Blueshift Memory is one that has been well-documented over the years – the ‘Von Neumann Bottleneck’,” said Peter Marosan, CTO and Founder, Blueshift Memory.

This is the phenomenon that data transfer between the core and the memory has become the limiting factor in computational speed. Not only has this effect never been mitigated effectively, but as computing tasks grow more data-hungry it becomes increasingly more of an obstacle to efficiency.

The root of the problem is that communication between the processor core and the main memory (usually DRAM) is ineffective and is holding back progress. Earlier attempts to improve this include augmenting the central processor (CPU) with a graphics processor (GPU) or using a hardware accelerator in the form of an FPGA card or an ASIC.

New memory types such as MRAM and FeRAM have also sought to make the stored data more readily available. However, the growth in data-intensive calculations like big data, AI, and in-memory databases is increasing the effect of the bottleneck more quickly than these remedies can keep up with. Using a larger number of cores to cope with the calculations turns out to be counter-productive, as the extra data demands just make the bottleneck worse.

Marosan began to realise the extent of this problem when a project he was working on required him to speed up a big data calculation at a software level.

“It quickly became obvious that our CPU cores were unable to run at their maximum speed, as there was not enough data available to them at the optimum time. Caches simply cease to be effective when data volumes become too large, and the processor stays idle while waiting for new data. When I strengthened the machines by adding extra cores, the waiting time did not change. Increasing the number of cores doesn’t solve the problem – it just results in more wasted CPU capacity,” he notes.

Lightbulb moment

Having learnt that he had encountered a known hardware limitation, he accepted it and moved on.

“But the creative process is strange, and it sometimes needs an external influence to generate a lightbulb moment,” Marosan said. “One evening when I was helping my teenage son to change the memory chips in his computer, I started describing to him how it worked, and suddenly – with the physical picture in front of me – I began to understand how we could solve this problem.”

Over the following few weeks, Marosan investigated the new idea from many angles, applying his knowledge of computer science and maths, and he had some very inspiring conversations with his friends.

“It was the start of a fantastic journey. Having checked my calculations over and over again, I shared them with two very good friends who are experienced in supercomputing and semiconductors, and they confirmed my assumptions and assured me it was an interesting solution that could work,” he said.

The focus of Blueshift Memory’s technology is the Cambridge Architecture, the next-generation technology for stored-program machines, designed to replace the currently-used modified Harvard architecture and to overcome the traditional constraints of the von Neumann bottleneck.

The company’s solution offers significant performance benefits:

  • Zero Latency Memory
  • Reduces energy consumption by 30–50%
  • Performance Improvements 2x–1000x
  • Linear performance can become logarithmic, and logarithmic performance can become constant 10x–100x
  • Can extend product life cycle of older DDR generations at lower cost

The highs and lows of a deeptech startup

“[Establishing the company] was probably the scariest part, and it was almost certainly harder work than creating the idea itself,” Marosan notes. “I had run an R&D startup before, but each one is unique. By definition, you are working on a problem that no-one has previously addressed, and you have to communicate your proposition to investors and potential customers. You also have to protect your IP, since at this stage that is where the entire value of the company lies.

“We are still a young company, but we are making exciting progress!”

In 2021, Blueshift Memory was named by StartUs Insights as one of the Top 5 RAM solutions startups worldwide.

In 2022, the company was awarded an Innovate UK Smart Grant, against fierce competition – this project is one of only 71 out of a total of 1,072 applications that were successful in securing funding in that round. The 13-month project allowed the team to configure a full-size FPGA demonstrator with the new architecture, optimising it for faster performance and better power efficiency in a computer vision AI application.

More recently, Blueshift Memory announced that it is using Codasip Studio tools to integrate RISC-V processor IP into an FPGA-based design for memory architecture, modifying the Codasip core to maximise memory bandwidth.

“Our ultimate goal is simple but ambitious: we would like to see the Cambridge Architecture adopted as an industry standard in servers, computers, mobile, and the IoT,” concludes Marosan.

This article originally appeared in the May/June 2023 issue of Startups Magazine. Click here to subscribe