Startup aims to make CPUs 100x more energy efficient with 'reconfigurable' chips - Efficient Computer reveals Monza chip
100x more efficient than MCUs, 1,000x lower power than GPUs too, Efficient Computer claims.
Chip startup Efficient Computer has emerged from stealth mode and unveiled a processor microarchitecture that promises to be 100 times more energy efficient than currently available general-purpose CPUs. The reconfigurable processor architecture is tailored for specific use cases, particularly in low-power embedded and edge computing, and requires a proprietary software stack (compiler) that supports general programming languages. Reuters reports that Efficient already has its first test chip, called Monza.
Traditional general-purpose processors are architected to handle virtually all workloads possible and to be backwards compatible with software released decades ago, which greatly increases their complexity and eventually power consumption. As Efficient Computer puts it, they are overdesigned for generality and spend loads of power on inessential internal data movement and instruction control overheads. Efficient's Fabric architecture is a reconfigurable dataflow processor architecture that can execute specially optimized code with parallelism on its 'computing fabric.' The architecture was developed in over seven years of research at Carnegie Mellon University.
Recompiling software is a must for this architecture, so software compatibility will be a limitation of Fabric-based processors on the mainstream market. Efficient's software stack supports major embedded languages, so developers of actual applications will be able to quickly recompile their code for the fabric architecture.
Efficient provides few details regarding how its processor architecture works, but based on how the company describes it, it looks like the CPU resources can be adjusted by software for a particular workload, which can greatly enhance efficiency. Meanwhile, Efficient claims that Fabric can handle general-purpose data processing computations, data analytics, and be used for AI and ML, which suggests that we are dealing with an inherently parallel architecture. Efficient says its Fabric architecture is 100 times more efficient than microcontroller units (MCUs) and consumes 1,000 times less power than GPUs. What the company has not revealed are actual performance numbers for chips based on the Fabric architecture.
Yet Efficient is wasting no time to remind us of several key advantages of processors with low power consumption. In general, Fabric enables more efficient on-device computing solutions (which eliminates or reduces costly communication to the cloud), lower operational costs, larger fleet sizes, and a reduced environmental impact. Ultimate energy efficiency could create new use cases and even new classes of devices, which will diminish its software compatibility issues as those devices will need new software anyway.
"Energy consumption impacts nearly everything in modern computing, from where devices are located to the capabilities they offer and the scale of their deployment," said Brandon Lucia, co-founder and CEO of Efficient Computer. "We are removing the energy barrier from computing at the edge, while giving developers the freedom and flexibility to quickly build devices and applications at scale. Efficient hardware and software will significantly reduce energy consumption for computing, creating entirely new categories of use cases."
Efficient Computer has received its first round of funding of $16 million led by Eclipse, a venture capital firm that has funded Cerebras, FlexLogix, and Tenstorrent.
Stay On the Cutting Edge: Get the Tom's Hardware Newsletter
Get Tom's Hardware's best news and in-depth reviews, straight to your inbox.
"The technology community's long-held secret of highly inefficient general-purpose processors has slowed innovation and limited applications, particularly at the edge," said Greg Reichow, partner at Eclipse. "More than just closing this gap, the Efficient team is introducing an entirely new category of processor that is enabling organizations to reconsider what is possible. With its unmatched energy efficiency, the software-agnostic processor is capable of powering a variety of smart devices with additional capabilities designed to improve the user experience, data consumption, and overall serves as a catalyst for innovation moving forward."
Anton Shilov is a contributing writer at Tom’s Hardware. Over the past couple of decades, he has covered everything from CPUs and GPUs to supercomputers and from modern process technologies and latest fab tools to high-tech industry trends.
-
edzieba That's nice. Go stand with the bones of all the other companies promising their totally-not-an-FPGA/PLA would be super duper efficient if only everyone would abandon all their existing applications and write new ones targeting our special new architecture.Reply -
ekio I always had this kind of tech phantasm of a chip that is like a general CPU that can be programmed like a FPGA to emulate dynamically any chip/isa configuration, huge single thread, super multithreaded, gpu etc, with on die memory to replace the need of external ddr for speed and efficiency. Is that what is it ? If yes, I want to invest immediately.Reply -
thisisaname I guess they are looking for some funding as $16million is not going to go very far.Reply -
bit_user "The technology community's long-held secret of highly inefficient general-purpose processors has slowed innovation and limited applications, particularly at the edge,"
It's not a "secret" that general-purpose CPUs are inefficient! Why do you think people bother to use FPGAs or design ASICs??
So, this idea of reconfigurable computing is hardly new. Aside from FPGAs, which can't be reconfigured very frequently and take a long time to do so, there have been other attempts. One that I know a little about was developed and brought to market by a company called Stretch, Inc. (later acquired by Max Linear).
https://www.maxlinear.com/stretchinc
The field of reconfigurable computing in embedded systems was already mature enough in 2006 that I ran across this survey paper, while searching for more info on Stretch:
https://citeseerx.ist.psu.edu/document?repid=rep1&type=pdf&doi=4fe2d039554c2ce473ca6afe668f206e41c1daa8 -
Avro Arrow
Yep, Intel Itanium comes to mind.... ;)(y)edzieba said:That's nice. Go stand with the bones of all the other companies promising their totally-not-an-FPGA/PLA would be super duper efficient if only everyone would abandon all their existing applications and write new ones targeting our special new architecture. -
bit_user
I'm sure the " totally-not-an-FPGA/PLA" part was sarcasm. Itanium doesn't remotely belong to this class of reconfigurable computers.Avro Arrow said:Yep, Intel Itanium comes to mind.... ;)(y) -
Conor Stewart
I was about to comment this myself, this is not a secret and never has been. It just seems like a typical startup thinking they have done something totally unique and know more than anyone else.bit_user said:It's not a "secret" that general-purpose CPUs are inefficient! Why do you think people bother to use FPGAs or design ASICs??
So, this idea of reconfigurable computing is hardly new. Aside from FPGAs, which can't be reconfigured very frequently and take a long time to do so, there have been other attempts. One that I know a little about was developed and brought to market by a company called Stretch, Inc. (later acquired by Max Linear).
https://www.maxlinear.com/stretchinc
The field of reconfigurable computing in embedded systems was already mature enough in 2006 that I ran across this survey paper, while searching for more info on Stretch:
https://citeseerx.ist.psu.edu/document?repid=rep1&type=pdf&doi=4fe2d039554c2ce473ca6afe668f206e41c1daa8
I highly doubt this will see any kind of mainstream adoption. Especially if most software wont work with it and would need rewritten. Another major issue is vendor lock in. If you invest time in developing your program to work with their architecture then you are completely at the mercy of a startup, which isn't a good position to be in. Your code will not work on any other processor.
It's very interesting that they compare it's efficiency to microcontrollers rather than desktop CPUs and that they talk about embedded a lot. In the embedded industry a chip that can be reconfigured for loads of different tasks really isn't that useful and is often expensive. If you are making a HMI, then use a chip designed for that purpose, if you are making a motor controller then use a chip with the required peripherals, you don't typically go and use a large general purpose chip that could do everything, you use the cheapest chip that is suitable. They seem to market the concept to general computing but then change and talk about microcontrollers and embedded development.
The say the efficiency is 100x higher than a microcontroller, but that is extremely vague, that is everything from a small 8 bit, 16 MHz MCU to a 1 GHz capable cortex M7 or similar or low power optimised MCUs. So again it seems like some dodgy claims. If they were being honest and actually know what they are doing then they wouldn't have just compared it to a microcontroller without giving at least details of which MCU they are comparing it to.
They also say it uses 1000x less power than a GPU which again is a nonsense claim. For all we know it has 10000x less performance than a GPU. You could say a 0.1 W LED uses 600 times less power than a 60 W halogen bulb, yes you would be correct but the LED is putting out much less light too. Or saying an MCU uses however many times less power than a desktop CPU, it really doesn't actually tell you much. What they have not claimed is that it can match the performance of a GPU whilst using 1000x less power. This all makes me think that this will be a low performance chip.
Overall this is nothing new and will likely fail like pretty much every other time it has been tried and it seems full of the usual startup marketing nonsense. -
bit_user
To have gotten even seed funding, I'm sure they're not that naive. It sure would be interesting to hear exactly why they think they can succeed where others haven't.Conor Stewart said:It just seems like a typical startup thinking they have done something totally unique and know more than anyone else.
That's not actually a deal-breaker for embedded applications, depending on whether it truly requires a rewrite or just a recompile.Conor Stewart said:I highly doubt this will see any kind of mainstream adoption. Especially if most software wont work with it and would need rewritten.
That Stretch, Inc. company had a C compiler that could use profiling data to automatically determine which hot spots of your program should be synthesized into logic gates, while the rest ran on a general purpose core. So, at least their sales pitch was that it just needed a recompile and not a full rewrite.
The cool thing it could do is reconfigure itself about 100 times per second (which is orders of magnitude faster than you could reconfigure a typical FPGA). So, different parts of your program could use different configurations of the logic array. This was an actual shipping product they sold that could do video compression, among other things. And yet, market uptake was relatively low, with most sales going to people treating it as effectively a black box video compression ASIC.
The vendor lock-in thing is mainly an issue just because they're a startup that might go out of business before they can make enough chips to use in your product. Otherwise, it's really just a question of how much you need to customize your code for it. In the embedded world, firmware is typically customized to a given chip, to some degree.Conor Stewart said:Another major issue is vendor lock in. If you invest time in developing your program to work with their architecture then you are completely at the mercy of a startup, which isn't a good position to be in. Your code will not work on any other processor.
Well... you're sort of overlooking FPGAs. So, maybe they've got something that can take on FPGAs in one or more respects. Or, maybe they've got something like Stretch had, where the reconfiguration happens while the program is running so you can get more mileage out of a limited logic budget.Conor Stewart said:It's very interesting that they compare it's efficiency to microcontrollers rather than desktop CPUs and that they talk about embedded a lot. In the embedded industry a chip that can be reconfigured for loads of different tasks really isn't that useful and is often expensive. ... you don't typically go and use a large general purpose chip that could do everything,
If it can let you do big things on a microcontroller-scale power budget and for similar cost, then it's an apt comparison.Conor Stewart said:They seem to market the concept to general computing but then change and talk about microcontrollers and embedded development.
There's a reason they're not giving us these details, and it's not because they're so incompetent. They're announcing themselves because they want to raise money. If you were a potential investor, you'd get to see their full pitch deck, which would include all of the sorts of details we're talking about.Conor Stewart said:If they were being honest and actually know what they are doing then they wouldn't have just compared it to a microcontroller without giving at least details of which MCU they are comparing it to.
However, what they've announced to the general public is just enough to attract the interest of investors, but not enough to tip off potential competitors. Once they have an actual product they're trying to sell, that's when we'll get all the juicy details. Right now, it's just too early.