MC,
That was a great article. I think it's worthwhile for everyone to see this following part of the article as it adds a lot of weight to the previous discussion we were just having:
https://www.crn.com/slide-shows/components-peripherals/pat-gelsinger-intel-will-be-more-ecosystem-friendly-than-nvidia/5think about the overall hardware versus software industry. Thirty years ago, hardware was 2x the size of software. Here we are today: Software is 2.5x the size of hardware. That‘s a pretty simple observation that hardware has grown, software has grown dramatically, and software and SaaS and some of the things that we were already talking about in that respect.
And if you go back to the answer to the last question: Are we going to have more software revenue, software products at Intel going forward? Yes. Customers have moved from being focused on the hardware level to the software-level interfaces, and our job now and some of the things that I‘ve learned in my 11-year vacation is delivering silicon that isn’t supported by software is a bug. We have to deliver the software capabilities, and then we have to empower it, accelerate it, make it more secure with hardware underneath it. And to me, this is the big bit flip that I need to drive at Intel.
To deliver a hardware product that doesn‘t have the full support of the software ecosystem already in place. Why did you waste the transistors? Why did you waste the validation time? Why are you wasting this power budget for our customers if you haven’t enabled the software ecosystem? Software is more important. Those APIs are more important. The developer is more important. And in that sense, I have to essentially create that flip at Intel to fully realize what we do in silicon. Not only do we need to deliver the software, the BIOS, the firmware, the p-code, all of those things, the PyTorch support, etc., concurrently, but in fact we need those delivered quarters or years ahead of time so software development can be in place by the time we deliver the hardware that enhances it.As you can see, Intel under Pat is now refocussing on getting software delivered
before their hardware comes out. So from what Mike Davies has previously said, Loihi is a good 5 years away. But he also recently said that software is the hardest part. The fact that he's mentioning software in this context indicates he's already been told to focus his attention on software, then once that's ready [maybe] you can go back to your developing your hardware (Loihi). This is the Pat Gelsinger way, get your software done first and then move onto providing the hardware that can do the job. This way customers can start practicing or simulating the hardware out, get a better idea on how they want to implement it and then eventually test it out on a silicon.
If Intel is suddenly spending so much time and money on software, they want to make sure that their customers will use it. But they know Brainchip have a neuromorphic chip coming out that also has Python capabilities. So again, who would want to buy their hardware or use their software if Loihi is going to be obsolete?
Mind you, Brainchip's current software is limited in the SNN space because they've been focussing on commercialising their existing chip. To maximise commercial opportunities the main priority is making Akida compatible with existing machine learning models, like CNN's. Brainchip haven't had much chance to look at coding in SNN capabilities, and they've also got popular platforms other than Tensorflow (like Onnx, Pytorch, Caffe etc....) that they want to focus their time and money on. By focussing on these platforms Brainchip can get more of their customers up and running as fast as possible. SNN capabilities are a want, not a need at this stage. Plus they've already got their advanced SNN algorithm built into MetaTF which is arguably the most powerful SNN algorithm out there because it can do simple feedforward learning at the edge.
So assume Brainchip have now been approached by Intel. Intel don't want to develop software and hardware that won't be used so they've approached Brainchip, offering to make this software (Lava) compatible with Akida. This saves Brainchip from spending money and time they don't currently have on developing these additional capabilities which they don't currently need for commercialisation anyway. However, this Lava software could be handy because there's been billions (I'm guessing) spent on all this alternative SNN research. Some of the researchers may be onto something novel and don't want to throw it all away. So what Intel are actually offering is to provide Akida with a platform suitable for researchers. If Brainchip say yes, which we're assuming they did, it's a win-win situation.
Intel can now develop software which is guaranteed to be used and therefore not wasted. Intel may even get some future revenue from this software. Meanwhile, Brainchip get a research platform for free at a period of time where they can't focus their energies on it, so other future customers (researchers) are being supported for free and will automatically be able to switch from their existing Loihi chip to an Akida chip. Win-win.
Assuming this is actually what happened, it's not clear what would then happen to Loihi. Intel may continue to develop it further if they can find a section of the market (like the research community) where it can clearly differentiate from Akida. Intel don't want to waste all the time and effort spent on it so they may continue down this research path, hoping to develop something with better features for an alternative market stream. After all, it's naive to think there can only be one type of neuromorphic chip. Just like there are many different CPU or GPU or FPGA manufacturers (think AMD, Intel, NVIDIA, ARM, Lattice etc.) there may be room for many types of neuromorphic chips. Brainchip are working on Akida 1000 and 2000, but there may well be other architectures not yet conceived with other benefits and Intel could use this software to pursue that direction.
Either that or they're giving the market the impression Intel is still a neuromorphic contender to make people use their software, while in reality Intel will actually dump Loihi (Pat doesn't want to spend money developing something 5 years behind Akida). They are now focusing on their new larger market, software. Intel had already spent the time and money researching and developing Loihi 2. So Pat said throw it altogether and if it doesn't work, throw it in the bin. This approach allowed Intel to test out their latest pre-production 7nm node and gave a good marketing opportunity to brand Intel as a company for neuromorphic capabilities even if they no longer do the hardware.
Pure speculation, DYOR
(20min delay)