Palo Alto-headquartered PsiQuantum is taking aim at the data center with a new product that it says could lower power use, increase bandwidth, and eventually link up all data traveling inside the data center with light. Does that mean it's putting the brakes on building a useful quantum computer? Tune in here for that answer.
For bullet points (including the 'specs'), a print story, a Q&A with Robert Soderbery who will lead the new optical circuit switching business, and transcript of the video, sign up with QuantumJane for FREE. All it costs is your email and you will be alerted to the full interview with PsiQuantum CEO publishing next week and a factory tour of PsiQuantum that will follow.
<<THE BULLETS>>
- PsiQuantum is launching an optical circuit switch (OCS) for data centers that can cut power use and improve bandwidth by moving data with light.
- Some customers are already testing a prototype, according to PsiQuantum.
- Victor Peng, CEO: calls it "a multi-billion dollar opportunity". He joined this year, previously ran chip company Xilinx, and was deeply involved in AMD's AI and data center business after AMD bought Xilinx.
- Pete Shadbolt, co-founder and Chief Scientific Officer: says PsiQuantum's life's work is still building a quantum computer, and the data center product is not a pivot.
- Shadbolt: says building the switch uses technology similar to the quantum computer's, which helps the supply chain and lowers the cost of the eventual quantum machine.
- Robert Soderbery, Executive Vice President of the photonics and technologies business: joined four months ago to run the OCS business. He says Broadcom, Nvidia, Cisco, Marvell and Arista would be partners because the OCS links their existing switches into bigger systems. He calls it "the icing on the cake".
- Soderbery: says the first stage connects racks of GPUs. The long-term vision is light all the way down to the GPU, which needs BTO (barium titanate) for fast switching. BTO is not part of this first product.
- Courtesy of PsiQuantum, the 'specs' of the OCS, and what these words mean:
- 1.2 dB of insertion loss: Insertion loss is the light signal that gets weakened as it passes the OCS – lower loss means more of the signal reaches its destination.
- 40 dB of cross talk: Cross talk is unwanted signal “leaking” from one light path into another, like hearing a nearby conversation through a thin wall.
- Less than 3 ms of reconfiguration time: Reconfiguration time is how long an optical switch takes to change a connection from one path to another, like a traffic light switching lanes.
<<PRINT STORY>>
Quantum computing startup PsiQuantum on Wednesday said it’s launching a product that can help lower power use and improve data bandwidth in data centers.
The product, an optical circuit switch (OCS), is born from a decade of research and development to build a quantum computer using photonics technology, basically light. Some customers are already testing the prototype, said PsiQuantum.
“This is a multi-billion dollar opportunity,” PsiQuantum CEO Victor Peng told QuantumJane in an exclusive interview. “We’re absolutely going after the opportunity in optical circuit switching.”
Peng, who joined PsiQuantum earlier this year, was the CEO of chip company Xilinx, which was bought by AMD in 2022 for about $50 billion. Peng was then deeply involved in AMD’s AI and data center business. He’s not a physicist but said he came back from retirement to join PsiQuantum as he is fascinated by the future of computing.
But the multi-billion dollar market opportunity in the data center doesn’t mean PsiQuantum is abandoning the ultimate dream.
“The reason I’m on Earth…my entire life’s work has been to build the bloody quantum computer. That’s what we’re here for,” Pete Shadbolt, co-founder and Chief Scientific Officer, told QuantumJane. “Why now and how do we end up in this position? I think one of the biggest effects is actually just the enormous unprecedented mind-blowing build-out of infrastructure for conventional computing and AI has taken I think most people by surprise.”
Shadbolt said building the optical switch, which is similar to the technology that PsiQuantum uses in its quantum computer, helps build up the supply chain and lower the cost for the quantum computer.
An optical circuit switch essentially helps link racks in the data center so they can act like one giant, powerful brain, passing data between them as light. Today the data centers are still full of copper wires that electrons carrying data run around on. The devices that link those up today are dominated by companies like Broadcom, Marvell, Cisco, NVIDIA, and Arista.
Robert Soderbery, Executive Vice President of the photonics and technologies business at PsiQuantum, who was brought on four months ago to run the optical circuit switch business, said those companies would likely be partners of PsiQuantum. “They can essentially make a much bigger switching network by using their existing packet switches, like Broadcom’s Tomahawk is the most well-adopted switch ever in the history of time. We can essentially take many Tomahawks and link them together using an optical circuit switch to build an even more powerful system,” Soderbery said, calling PsiQuantum’s switch “the icing on the cake”.
The world has been using light to transfer data at long distances for decades through fiber-optic cables. Optical switches will be able to help solve one of the biggest problems in data centers today — reducing power use. Power-hungry data centers are the new political, social, and environmental hot potato with AI.
In fact, tech giant Google already engineered a data center where information travels in light between the server racks. It uses tiny mirrors to steer laser beams, replacing a layer of electrical switches. In 2022, Google said this system lowered its cost to build by 30%, improved bandwidth by 30% and lowered energy use by 40%.
“But there’s still this thing that you’re moving. You’re physically moving a mirror and you have laser beams bouncing around in a box in the data center. It’s kind of miraculous that they made it happen. But there’s also an appetite and an interest to see if we can do that all on chip, no moving parts, solid state, OCS,” said Shadbolt.
He said one of the challenges to building an optical switch that works was extreme precision in attaching fibers to the chips to prevent photon loss. Light escapes easily and high photon loss results in data loss and errors.
“In some ways this is the product of ten years of work and it’s a really impressive piece of technology. In other ways it’s really simple. This is a chip…It’s a [printed circuit board]. It’s some controllers. It’s incredibly simple and that’s actually I think one of the reasons why we’re so excited about it. We’re now able to make this part hit the spec and soon we’ll be deploying this,” Shadbolt said.
PsiQuantum has also developed a way to manufacture chips using a layer of barium titanate, also known as BTO. But that will not be part of this new product.
“The long-term vision is that we go all the way. We take that light signal all the way down to the GPU [chip]. But that’s going to be a second stage of evolution in the industry,” said Soderbery. “There you need to have very fast switching. And BTO is the only really fast switching material that’s available to build a fast optical circuit switch. So it’s a very exciting and transformational technology.”
<<Q&A>>
Here’s a short Q&A with Robert Soderbery, Executive Vice President of the photonics and technologies business at PsiQuantum. He leads the optical circuit switch business. (The interview has been edited for clarity and to delete QuantumJane’s confusing questions.)
QuantumJane:
Was your role newly created for the optical circuit switch business?
Robert Soderbery:
The company had always had this idea that we’re building technologies that could be of use in non-quantum applications and in commercial applications and especially AI. I have a long background in the San Francisco Bay running large tech organizations. And CEO Victor Peng asked me to come on and lead this new function, which is really about commercializing these amazing technologies we’ve built for quantum.
QuantumJane:
So you joined after Victor arrived?
Robert Soderbery:
Yes I’ve been here for four months. So I’m one of the most recent adds to the management team.
QuantumJane:
What market is this product targeting and when PsiQuantum is done with its work, what are the data centers going to look like? Is there going to be an electron running around anywhere outside of a chip when you’re done?
Robert Soderbery:
The big picture is that it’s clear to everybody in the scale up AI world that photonics is going to be very, very important. So every player understands that — the AI rack scale players, the networking folks, and obviously the photonics companies. We’re seeing that first happen in this trend that’s called CPO or NPO, which is really about taking the old kind of clunky pluggable optical transceiver and getting that as close to the chip as possible so that you can go straight from the chip to an optical signal.
QuantumJane:
Right. That’s the Ayar Labs and Lightmatter.
Robert Soderbery:
Exactly. What we’re doing is now building a switch that can natively take that signal coming off of the device it’s talking to, whether it be an electronic switch or a GPU. And the unique thing about what PsiQuantum is doing is that because we have this very low-loss technology, we can use the existing, and these new data center optics. So we can work with that system they’re building. We don’t need to make them change their photonics. And by using light in switching we can build systems that are dramatically higher scale and larger clusters and more effective AI platforms.
QuantumJane:
So in Google’s data center using optics, the thing called the spine, were using mems and lasers, right? But the leaves inside each of the racks were still electronics. Is that where your OCS sits? Does it sit inside the rack and in the data center connecting the racks? Are we able to get rid of all of the electrons outside of the chip?
Robert Soderbery:
Our vision for OCS is that it will start in that spine tier. So we’ll start by connecting multiple AI racks together. So instead of a 72 GPU rack, you could take four of those racks, integrate them with OCS, and have a 288 GPU system. And a bigger system with more GPUs and more bandwidth has huge benefits for both training and inference. So that’s a big win. Now the long-term vision is that we go all the way. We take that light signal all the way down to the GPU. But that’s going to be a second stage of evolution in the industry.
QuantumJane:
Is that the stage where you need the BTO substrate?
Robert Soderbery:
We call that the OCS fabric. You just plug the GPUs into the fabric. And there you need to have very fast switching. And BTO is the only really fast switching material that’s available to build a fast optical circuit switch. So it’s a very exciting and transformational technology. So inside that NVL72 one thing is going to change. But it’s going to essentially have the same number of GPUs and the same number of switches. Whether it’s NVIDIA or AMD or Google, or Marvell or Broadcom, whoever your provider is. What they’re doing is they’re putting optical uplinks on those switches that are inside the rack. And then those optical uplinks, which are using NPO technology, those are going to talk to the optical circuit switches. So you they’re making a single change to what goes on inside the the AI rack and then using the optical circuit switch to stitch together many of these racks to build larger machines. Now above the rack, with the OCS switch, your light signal comes in and it just passively goes straight through to the output port and away it goes.
There’s a lot of excitement and the big bet in CPO and what’s driving Coherent and Momentum Optics and Ayar Labs and everybody’s valuation is that everything’s going to optics.
QuantumJane:
Your product is called an optical circuit switch. Does that make it less flexible than a packet switch? Is it already set in stone or can it be flexible and connect all-to-all?
Robert Soderbery:
It’s a circuit switch and not a packet switch. What a circuit switch means is you have to say I want to connect your device number one to device number two. So you have to give it the map of what you want to connect, and then it just flows the signal straight through that switch. And so that’s a different kind of switch. And it has a lot of benefits. It’s very low power. It’s very low latency. It’s very high bandwidth. But it’s also something new in these architectures. And that’s really the big news — AI architectures are now starting to adopt these. And what Google showed is that these add tremendous value and they published numbers around the benefits that they’ve seen. And then of course it’s really driving the rest of the industry to consider how they should incorporate photonics into their systems.
QuantumJane:
Do you need a packet switch once you have the OCS?
Robert Soderbery:
Not at the spine level. So we just leverage the existing packet switches that are already in those racks. We’re adding the OCS at the spine. So we’re not necessarily replacing anybody. There’s no way to hook these racks together today, because that system is just too costly and too complex. So they’ve been waiting for OCS to come along to add this new layer. The industry is going through a transition from copper to optics. And everybody agrees. The photonic systems are just going to get more and more sophisticated. First our OCS spine, and in the future fabrics, bigger fabrics - a long-term tech evolution to all optical and photonic based networks. That’s the way networks can scale.
QuantumJane:
The product that customers are sampling today is an OCS switch that would connect the racks. In order for the next system that would connect the servers inside the rack to come about, you need to use BTO for much faster switching. And so that’s the second stage later down the road?
Robert Soderbery:
CPO packaging is an important part of that next step. With CPO we have to go all optics all the way down to the GPU. We have to deliver optical circuit switches. There’s got to be innovation in laser technology. So this is part of the natural evolution of photonics. So BTO is very important. But we’re not the only player that has to innovate to deliver photonic fabrics. That’s an industry phenomenon.
QuantumJane:
And the industry is not ready for BTO, because that’s nanosecond switching. If the industry is not ready for that can you follow the industry with the technology you have today at the rack-to-rack level? Can you bring that down to the server-to-server level?
Robert Soderbery:
This is the natural way things evolve. First you introduce the basic capability, then you introduce the more sophisticated capability, then there’ll be things that we haven’t even talked about that come after that. So think about this as just the evolution from today where it’s almost all copper in these racks to the future which will be an all-photonic world. And that’s going to take five to ten years to happen.
QuantumJane:
And today if I want to install your OCS, am I taking something out and putting that in or am I just adding that?
Robert Soderbery:
You’re adding it because you have those racks and you’re adding this and then you’re using it to stitch the racks together. So it’s the icing on the cake, let’s say.
QuantumJane:
How will network switch providers like Nvidia, Broadcom, Cisco, Marvell and Arista be impacted?
Robert Soderbery:
Those companies would be partners because they can essentially make a much bigger switching network by using their existing packet switches, like Broadcom’s Tomahawk is the most well-adopted switch ever in the history of time. We can essentially take many Tomahawks and link them together using an optical circuit switch to build an even more powerful system.
Soderbery says he’s best known for being the senior vice president of enterprise networking at Cisco when John Chambers was CEO. He has a lot of experience in managing large infrastructure businesses and was excited by the opportunity Victor Peng described to him at PsiQuantum. When we were speaking this week he was “poking around doing quantum-y things” at Caltech as he loves quantum stuff as well.
<<TRANSCRIPT FOR VIDEO PIECE>>
Building a useful working quantum computer is a monumental engineering challenge still several years out. But the innovations that come from chasing that future goal are giving birth to technologies that could propel all computing forward, today.
Silicon Valley-headquartered PsiQuantum, which made waves with a giant $1 billion funding round last year, including investors like Nvidia, is launching a new product that can help reduce data center power use and improve bandwidth with light. PsiQuantum CEO Victor Peng says some companies are already evaluating prototypes.
(SOUNDBITE: Victor Peng, CEO PsiQuantum) “We get here because of what we've done with quantum computing. And then with a fairly small technical effort, we can leverage that into traditional computing. And this is a multi-billion dollar opportunity, of course. Companies, many companies are just focused on this, and we get that out of our focus on photonics and driving things at unbelievable low loss levels.///We're absolutely going after the opportunity in optical circuit switching.”
Peng joined PsiQuantum this year, but he’s not a physicist. He ran a multi-billion dollar chip company called Xilinx, which AMD took over for about $50 billion in one of the biggest chip industry deals. There, he was intimately involved in the AI and data center business and has over four decades of semiconductor industry chops. He’s fascinated by the future of computing but sees a big opportunity with what the PsiQuantum team has already created: An optical circuit switch, or OCS, that helps link the racks in a data center so they can act like one giant, powerful brain, passing data between them as light.
(SOUNDBITE: Victor Peng, CEO PsiQuantum) "AI data centers, but eventually all data centers, whether they're focused on AI or not, is going to transition from copper connectivity to optics. And the reason is because copper is finally run out of gas. People again cannot get the bandwidth they need out of it."
Copper wire is what electrons carrying data run around on chips and servers. And the devices that link those up today are dominated by companies like Broadcom, Marvell, Cisco, Nvidia, and Arista. But instead of electrons, using light for data transfer, the way we’ve been using fiber optic cables for decades, is more efficient.
That may sound like a geeky technology in the weeds, but optical switches can work towards solving one of the biggest problems in data centers today --reducing power use. Power hungry data centers are the new political, social, and environmental hot potato with AI.
So why are optical switches better? Here’s how PsiQuantum co-founder and Chief Scientific Officer Dr. Pete Shadbolt explains it.
(SOUNDBITE: Pete Shadbolt, Chief Scientific Officer) “Imagine that you have a whole bunch of optical network traffic in fiber. And now you want to switch that traffic around. One thing you can do is use pluggable transceivers, the little blue or rubbery pluggable devices. And so you convert from light to electricity. Then you have electronic network traffic. Switch that electronically with a regular chip. And then convert back to light with another pluggable. That works, but you can imagine it's not the most efficient thing to do. It's power hungry. You've got a lot of these transceiver modules. It's not very reliable. So everyone would love to do all optical switching.”
In fact, tech giant Google already engineered a data center where information travels in light between the server racks. They use tiny mirrors to steer laser beams, replacing a layer of electrical switches. In 2022 it said this system lowered its cost to build by 30%, improved network movement by 30% and lowered energy use by 40%.
(SOUNDBITE: Pete Shadbolt, Chief Scientific Officer) “But there's still this thing that you're moving, you're physically moving a mirror and you have laser beams bouncing around in a box in the data center It's kind of miraculous that they made it happen, but there's also an appetite and an interest to see if we can do that all on chip, no moving parts, solid state, OCS.”
So what keeps an optical circuit switch from going mainstream already? Shadbolt says the challenge is light. It likes to escape and when that happens data can go missing too. It requires extreme precision of connecting the optical fibers to chips and any other connection points to keep photon loss to a minimum. You can crank up the lasers to offset the loss, but that makes it power-hungry and can create failures due to a hot, hot laser.
(SOUNDBITE: Pete Shadbolt, Chief Scientific Officer) “In some ways this is the product of ten years of work and it's a really impressive piece of technology. In other ways it's really simple. This is a chip. /// It's a PCB, it's some controllers, it's incredibly simple and that's actually I think one of the reasons why we're so excited about it is that we're now able to make this part hit the spec and soon we'll be deploying this.///they're not sexy like BTO, but they work great.”
Sexy like BTO? Some might say BTS is sexy, but BTO? That’s barium titanate, a material that can make ultra fast optical switches. If that sounds sexy to you, tune in for part two of the PsiQuantum story of bigger is better in quantum computing and fast and furious optics. For now BTO isn’t part of the new product launch.
A data center product for today raises an obvious question: is PsiQuantum giving up on quantum computers?
(SOUNDBITE: Pete Shadbolt, Chief Scientific Officer) "If you're making ultra high performance integrated photonics, it doesn't take a genius to say, wow, you could make some products here for conventional computing and so on. I'd actually say we've been incredibly restrained to not do that for a decade. And that's because the reason I'm on Earth and the reason that my entire life's work has been to build the bloody quantum computer. That's what we're here for. If we go off and do fifteen different things in parallel, we're probably not gonna succeed in focusing on the delivery. Why now and how do we end up in this position? I think one of the biggest effects is actually just the enormous unprecedented mind-blowing build-out of infrastructure for conventional computing and AI has taken I think most people by surprise.///the thing that we're really excited about is that, of course, in addition to actually being able to provide a product and get involved in an industry, there's also going to be this reciprocal effect. For us, where us going out and building huge volumes, we hope, of these parts de-risks our own supply chain, yield, volume, reliability, because we're just comfortable operating at that kind of scale, which is important as we go on ahead and build these very, very large quantum computing systems."
THAT'S IT! Thanks for reading to the end.