Mixxtech’s Plans For India’s Nascent Semiconductor Space | In Conversation With Cofounder Vivek Raghuraman | NDTV Profit
Category
🗞
NewsTranscript
00:00Hello and welcome to NTG Profit.
00:01Joining us today is Mr. Vivek Raghuraman.
00:03He's the co-founder of Mixtec, a US-based startup
00:06involved in the AI and HPC sectors.
00:09He's in India at the DAM Capital Investor Summit,
00:13Investor Meet, with a new AI chip, so to speak.
00:19Mr. Raghuraman, welcome to NTG Profit.
00:21Thank you.
00:22So first of all, we went to your website
00:24to check what you do.
00:26So your website says siliconization in progress.
00:29So let's start the conversation there.
00:32Sure.
00:33So what we as a company are focused on
00:37is improving the interconnects between the server
00:42chips, the high-performance computing chips,
00:44and the new GPUs or the AI accelerator
00:47chips that are being designed.
00:49So today, as you are all aware, most of the development
00:53or the largest GPUs that are being made
00:56are by companies like NVIDIA.
00:58And then eventually, these need to be scaled
01:02to big-size data centers.
01:05Right now, they are very small clusters.
01:07But then as you start scaling them,
01:09you run into significant power issues
01:11in terms of the network power consumption.
01:15So we are trying to address that space
01:17by using a technology called silicon photonics.
01:21Now, this technology has, so optic fibers
01:24have always existed in data centers.
01:27But the density requirements of these GPUs
01:31to communicate with each other, the latency requirements,
01:34are all driving for a paradigm shift
01:37in terms of implementing these new silicon photonics-based
01:41chips that can be integrated with the GPUs.
01:45So the newness in this is basically
01:48today's GPUs are interconnected to each other,
01:51predominantly using electrical links.
01:54These are high-speed electrical links
01:57where the data goes between the GPUs or GPU to memory,
02:01and the communication happens for whatever
02:03the large-language models or the generative AIs
02:05are being developed.
02:07You have something actually on the table
02:09that in front of us, there's actually a chip
02:11that you have brought us.
02:12Just talk to us about this particular chip.
02:14So, yes, so you talked about siliconization,
02:17about what we are going to be doing.
02:19So this is a prototype of the siliconization
02:21that we are looking at, wherein what you see
02:24is a very large-body GPU ASIC
02:28that is constructed using the most advanced
02:31packaging technologies, 2.5D, 3D integration
02:35packaging technologies, and the silicon photonics
02:38integrated with that packaging.
02:40And this is going to be sort of the future
02:43of how the GPUs are going to be designed
02:46and how the GPUs are going to communicate with each other.
02:49So this is a photon-based transmission
02:51rather than electron-based?
02:52That's correct, that's correct.
02:54So this uses photons, which are directly converted
02:59at the package and then transmitted out.
03:02In that sense, of course, they're going to consume
03:04less power because that's the biggest concern
03:06about AI chips, AI supercomputers,
03:09their power consumption.
03:10So how much does this particular chip reduce
03:13or address that concern?
03:15So the major power that is consumed in the AI chip
03:20is actually the core silicon,
03:23but the other aspects of it,
03:25this is basically the networking, the memory,
03:27these are other aspects that consume power.
03:30And then there is an aspect of density.
03:32So with this solution, it provides two benefits.
03:36One is the density, and the other is the reach.
03:40And combined, it saves about 30% of the power
03:44for coming out of a GPU ASIC,
03:46compared to the technologies that are available today.
03:49All right, very interesting, sir.
03:51Let's talk about semiconductor manufacturing in India
03:53and your plans also to scale up this particular chip.
03:56Does India feature in those plans?
03:59What are your plans for India in that sense?
04:01Sure, I've been very passionate about developing
04:05technologies out of India.
04:06And even through my last 15, 20 years,
04:09we've actually developed some of these
04:11state-of-the-art niche technologies out of Bangalore.
04:14And we continue to plan and expand on that.
04:19And we will have offices out of Bangalore
04:22that will be focused on the complete vertical integration,
04:25everything from silicon, to the packaging, to the systems.
04:29And this is where the next piece comes in,
04:32which is basically the manufacturing
04:34that's being looked at for semiconductors in India, right?
04:38So now we have an opportunity where the design
04:40is very close to manufacturing,
04:42and we'll be able to enable very short time to market
04:47for the manufacturing of these products,
04:49because design and manufacture are going to be hand-in-hand.
04:52So is it going to be an OSAT project for you,
04:54or are you looking at even a foundry, maybe?
04:56So I would love to have a foundry,
05:00but I don't think it is there yet.
05:03And us being startups, we are time-bound
05:07to develop the technology and productize it.
05:10So for the foundry piece of it,
05:12we are going to exist on the existing supply chain.
05:16And with the goal that eventually
05:18there will be technologies
05:19that can be manufactured in India.
05:22And likewise with the advanced packaging,
05:26that we will look at both what is available today,
05:30versus what is being developed and manufactured in India.
05:33And I think I'm very confident
05:36about the systems piece of it,
05:38which is where the existing players
05:39like Kane, Semicon, come in,
05:41who have phenomenal facilities
05:44that can take these advanced ASICs
05:46and manufacture them into products.
05:50Finally, how are you funding this?
05:52Are you looking at funding rounds
05:54to fund your manufacturing processes?
05:57So yes, both the design and the manufacturing.
06:01So we have a set of strategic investors
06:05who have come in,
06:06and who are supporting us through this endeavor.
06:09And we are in the midst of our Series A,
06:12which is expected to close in the next couple of months.
06:15But as you all know,
06:16semiconductor startups are not like software.
06:20And it does require patience.
06:22It does require a lot of investment.
06:25So it's going to be a long-term effort,
06:28where we'll be developing this technology
06:30over multiple product generations.
06:33Finally, how novel is this idea?
06:35How many other people are involved in this particular space?
06:39So I'd like to break it into three pieces.
06:43One is the core silicon photonics technology,
06:47which is basically the science of transmitting optics,
06:51or photons, through the silicon.
06:54That has been existing for maybe 20, 30 years now.
06:58But the implementation of this solution into a product,
07:02the first commercialization happened through Intel,
07:05back in 2017,
07:06where Intel was the first volume supplier
07:10of this technology.
07:11And then, most recently, Broadcom developed this technology
07:17and brought into market
07:18the world's first co-packaged optics.
07:20So these two companies have been very successful,
07:22while there are other players, like Cisco,
07:24who've been doing this technology for the data center,
07:29for the U.S. data centers, as well.
07:31And in this new niche space, targeting the AI,
07:35there are quite a few startups,
07:37as well as the large corporations
07:39who are looking at this technology,
07:41and who, right now, it is becoming inevitable.
07:45So that's why there's a lot of investment
07:47that is going in this area.
07:49And we saw the benefits of that at companies like Broadcom,
07:55and now we are kind of expanding on that.
07:57All right, thank you so much, Vivek, for speaking with us.
08:00Thank you viewers for tuning in.
08:01This is Tushar for NTV Profits.
08:05Thank you.