Gradually, alternative architectures are being developed in the processor manufacturing industry, on the basis of which very remarkable chips are created. So, the Esperanto Technologies startup introduced its own version of the RISC-V processor, and not a conventional one, but a high-performance one designed to work with AI systems and machine learning. We have already written about experiments in this industry more than once, let’s see what kind of novelty it is.
A bit about architecture
As far as the RISC-V architecture is concerned,
adapt the creation of specific chips with the ability to choose which instruction set extensions will be used. The architecture is based on a free command system and, in fact, the processor architecture itself based on the RISC concept.
As for its features, the description of the architecture contains about 50 standard instructions. If you use extensions, you can get an additional 53 instructions.
Chips based on RISC-V have been produced not so long ago – since about 2017, not counting any prototypes and laboratory samples. Since 2018, the RISC-V Foundation has been working closely with The Linux Foundation.
RISC-V is great for parallel computing applications. These include, among other things, machine learning, as well as AI systems. That is why several companies are developing processors based on RISC-V at once. Not to say that everything is going smoothly for everyone, but there are also successful examples. One can hope that Esperanto will become one of them.
Okay, so what kind of processor is this?
According to the developers, the chip is great for working in the field of AI and machine learning.
The name of the model is ET-SoC1 (Esperanto Technologies Supercomputer-on-Chip 1), its developers call AI Inference Accelerator. The processor did not appear yesterday, it is already being developed more than a year, so the chip can be conditionally called a novelty. But now it has only just begun to be tested by third-party companies. Features include 1,088 power-efficient, general-purpose 64-bit RISC-V cores with vector/tensor computing modules to optimize and accelerate operations related to AI and machine learning. In addition, the chip includes four high-performance RISC-V cores, 160 million bytes of on-board SRAM (152 megabytes), plus interfaces for connecting flash memory and external DRAM modules. As far as we know, there are 23.8 billion transistors in ET-SoC1.
One of the advantages of the new development is high energy efficiency, that is, reduced energy consumption with very high performance. It is clear that “each cricket praises its hearth, but the developers claim that they have created the fastest processor of this type at the moment based on the RISC-V architecture.
The maximum efficiency of the chip is observed when working in recommender systems based on machine learning. If the processor itself is as good as the developers say about it, then it can find application in hyperscale data center systems of the largest companies in the world. But, of course, in order for this to happen, the processor must be tested – in the “field conditions”, and not in the laboratories of the company that developed it. And this is already being done – for example, Samsung.
If you are reading our blog, you may be interested in these texts:
What Samsung says
According to them, the chip really does a good job with the tasks. The test team from Samsung SDS stated that the processor demonstrates high scalability. In general, this was discussed above when we discussed RISC-V, and according to Samsung representatives, as the number of chips in the data processing system increases, its performance increases rapidly. It is worth noting that this does not always happen with traditional chips – even with a very large number of them, the system performance grows up to a certain point, after which further growth either slows down very much or stops altogether. Systems based on the new chip
According to the results of the tests, it turned out that ET-SoC1, showing very high performance with models such as ResNet 50, DLRM and Tranformer, consumed (per processor) only 20 watts.
In addition to Samsung, the processor was also tested at Cambrian-AI Research.
representatives of this company, the performance indicators of the processor during the tests showed results close to those declared by the developers. The head of the test team, Karl Freund, said that the RISC-V cores do the bulk of the work without shifting the matrix multipliers to the MAC cores or GPU.
Positive feedback is, of course, good, but the commercial use of the chip among large corporations is still far away. Esperanto has now announced that it is inviting new partners to test its chip for a variety of tasks, and not just use it in recommender systems. The testing program allows partners to test different models, data types and computing systems up to 32 clusters.
Until the biggest corporations take notice of Esperanto (which, in principle, is already happening little by little), AI system manufacturers like Nvidia have nothing to worry about. But this is only for now, the situation can change at any time.