Frangipani
Regular
Hi TECH,
just an attempt to sort out the terminology…
The following is how I understand BrainChip’s nomenclature, but I could be wrong, so everybody please feel free to brainstorm and chip in…
In early 2022, shortly after Sean Hehir had joined BrainChip as CEO, “AKD1000” was still used as an umbrella term to describe everything our company had for sale at the time (chip, IP, PCIe board), as evident by the following investor conference presentation slides:
These days, however, a distinction appears to be made between the generational iterations of the Akida processor technology platform (categorised as akida, akida 2.0…) as opposed to the physical reference chips (so far AKD1000 and AKD1500).
The AKD1000 and AKD1500 SoCs are both silicon implementations of the Akida technology embodied in BrainChip’s 1st Generation Edge AI neuromorphic processor platform akida (technically speaking akida 1.0).
AKD1000 was implemented with TSMC at 28nm, whereas AKD1500 was taped out on GlobalFoundries’ MCU-friendly 22nm fully depleted silicon-on-insulator (FD-SOI) process, aka GF’s 22FDX technology. As reference chips, they were primarily meant to target prospective IP licencees (as a proof of concept), but at the same time the AKD1000 (on a PCIe board or inside a Dev Kit) benefitted individual developers (professional hardware engineers in companies or academic settings as well as advanced hobbyists) who were not interested in mass production of edge devices and the signing of an IP licence, but instead may have only required a single PCIe Board or Dev Kit for their projects or research (take note that it says on the BrainChip shop website in bold capital letters that development kits are not intended to be used for production purposes). Meanwhile, AKD1000 chips have also been integrated into the VVDN Edge AI Boxes, and the Unigen Cupcake Edge AI Server will soon be offered with a new configuration based on the AKD1500 (?) as an AI option.
Akida 2.0, the neural processing system’s enhanced 2nd Generation, was announced and introduced last year, but purely as an IP offering, productised in three different variations: akida-E, akida-S and akida-P, depending on where in the Edge AI spectrum (sensor edge < server edge) its prowess is required.
AKD2000, however, doesn’t exist - at least not yet.
The way I understand it, AKD2000 would be the name of BrainChip’s (hypothetical) reference chip based on Akida Gen 2, which may or may not materialise.
Let’s recall what was said earlier this year:
In an interview with Jim McGregor from TIRIAS Research during CES 2024 (January 9-12), Todd Vierra replied the following to his interview partner’s comment “And correct me if I’m wrong, but this is the Akida 2?”
Todd: “This is actually all ran [sic] on Akida 1 hardware. Akida 2, erm - we are in the process of taping out and we’ll get that silicon back a little bit later, but these are all just Gen 1...” (from 9:26 min). His statement about an imminent tapeout seemed to confirm what (according to FF) Sean Hehir had told select shareholders in the November 2023 Sydney “secret meeting”.
Surprisingly, a mere seven weeks later, during the Virtual Investor Roadshow (February 27), neither Sean Hehir nor Tony Lewis mentioned anything at all about a tape-out in progress, but instead argued a second generation reference chip as proof of concept was unnecessary, while at the same time not totally excluding a potential future tape-out complementing their core IP business. However, our CTO made it clear that it is definitely not their intention to manufacture chips on a large scale:
From 43:17
Roger Manning’s question: “Given BrainChip’s business model is to largely sell its akida IP, will it be necessary to prove each new version of akida in silicon?”
Tony: “So I think the big question was, will this event-based paradigm yield, erm can it be done and will it yield some benefits to customers, and I think we achieved that by taping out our earlier generation of products, and so we’ve already achieved that. And it’s my belief that there is only marginal benefit in taping out the next generation, we already proved the main points of it. And clearly we don’t want to start to manufacture chips on a large scale. We’d be competing with our customers and that would really break our business model right now.”
Sean: “Yeah, and to be clear, a lot of work has gone on with our ability to simulate workloads in Generation 2 as well. As Tony said, we certainly have reference chips in Generation 1, and the typical engagement course that we work with IP license prospects is we allow them to run models on there and/or simulate them in our simulation tools that we have for Generation 2, so Roger, I would say stay tuned, we may or may not, erm, but right now, there is no need for us to do that.”
Now the way I understood the “We’d be competing with our customers” comment is not for fear of treading on their customers’ or a specific customer’s toes (as other TSE users have interpreted it), but because they would shoot themselves in the foot by doing that: after all, it would be less profitable for our company when customers could just buy those chips off the shelves to utilise them in their in-house development (and hence save a lot of money, as there won’t be any follow-up costs for them) rather than having to pay an initial IP license fee and future royalties.
Does my interpretation make sense or am I overlooking something here?
And as for the next iteration of the Akida processor family, we don’t even know, yet, whether it will be called akida 3.0 or akida 2.X…
From min 47:41 min:
Sean Hehir: “… but round Generation 3, if you will: You know, I mentioned earlier in my slide that when I talked about our product planning and our execution cycle, if you will, you could see we are always planning on this product and we are always looking for improvements on our IP offering right there. Now whether we call it formally “Generation 3” or “2 something x” - but yes, we are in the middle of a planning cycle right now to make some changes and we’ll make announcements over time.”
So what we can say with confidence (provided we believe our CEO’s words) is that the advent of the Akida neural processor family’s 3rd generation is approaching… Whether there will ever be an AKD3000, though, is uncertain.
Regards
Frangipani
P.S.: I still can’t get my head around Todd Vierra’s statement, though. Would they really have backed out at the last minute, eg due to financial constraints?
Or was he possibly referring to a tapeout not of AKD2000 as a reference chip, but to a tape-out of a SoC by a specific customer that includes Akida 2.0 IP, along the lines of what DingoBorat said?
But then again, would he really say “…we’ll get that silicon back a little later”?
I suppose only if it was a joint development, such as possibly one with Socionext or Tata Elxsi (which could explain the different colour of the mysterious Custom SoC on that presentation slide)? As otherwise, wouldn’t such a SoC be taped out by the customer itself rather than by / in collaboration with BrainChip? I am a bit confused here.
It seems too early as a mere placeholder for the planned integration of Akida IP into the Frontgrade Gaisler SoC, which according to ESA’s Laurent Hili “We aim to tape out ideally before the end of the year, beginning of next year” (from 47:25 min shortly before the end of the mid -March BrainChip podcast Episode 31). After all, they could have labelled it “prospective Customer Custom SoC”, but the way it is presented on the slide, I agree with DingoBorat that it does appear to be an already existing physical implementation, indeed. Mmmmh…
Last edited:
