Product Demo Video Production
Astera Labs Leo CXL Memory Pooling — OCP 2022 Demo
Read Full Transcript of This Video
Transcript: Astera Labs Leo CXL Memory Pooling — OCP 2022 Demo
Hi everyone.
**Aanchal Sharma, Senior Director of Project Management, Astera Labs:** Hi everyone, I'm Aanchal Sharma Product Manager at Astera Labs and I'm super excited to be here in person at OCP 2022.
One of the hottest topics of discussion here has been CXL, Compute Express Link, and how it's going to enable the next generation of cloud server architecture. Astera Labs has been taking a leadership role in taking this from concept to reality.
Earlier this year we announced the pre-production sampling of our Leo CXL memory connectivity platform, the only solution in market to enable memory expansion pooling and sharing using CXL 1.1 and 2.0 specification.
We have already demonstrated memory expansion with our real live ASIC in several other platforms and conferences today. The start of the show is going to be the industries' first ever memory pulling with our Leo smart memory controllers.
So let's Jump Right In and talk about this demonstration.
So as you can see here, this is one of our custom platforms that is connected to our Leo system validation board. Let's actually go to the screen to check out the configuration before we run the demo.
So on the screen what we are showing here is the configuration for industries first CXL memory pulling using our Leo CXL smart memory controller connected to two CPUs from Intel using 2x8 CXL links.
Each of these CPUs are populated with half terabyte of local memory, whereas the Leo cxl Asic is populated with a total of 128 gigabyte cxl attached memory.
What we are going to show in this demonstration is the ability to share or pool memory between two CPUs using the CXL links.
So now let's take a look at how the CXL attached memory gets pulled between two CPUs and running it live.
So even though we have one Leo memory device connected to the server with the two CPUs when we do run as LSPCI, we see that two CXL devices appear here.
Now, let's go ahead and run NUMA control.
So when we run NUMA control we can see that node 0 here shows the CPU node 1 here shows another CPU, so there are two CPUs on the motherboard. And then we were supposed to see just one node with the CXL-attached memory with a total of 128 gigabyte, but what we are seeing on the screen here is the ability to show that we can pool memory between two CPUs and you see them appear as two different NUMA nodes.
So NUMA node 2 here shows the CXL device pooling 64 gigabytes of memory attached to CPU 0 and node 3 here shows 64 gigabytes again attached to CPU 1.
Now, what we demonstrated right now is the ability to share and pool memory resources between 2 CPUs using our Leo smart memory controller using the CXL attach links.
Now, let's go ahead and run an industry standard stress app that's going to stress the CXL attached memory.
So as you can see on the screen here, the status is passed and there are no errors detected.
So we have successfully demonstrated Industries first memory pulling now.
The reason why CXL-based memory pulling is so important is because there is a lot of memory that goes unused in a server behind CPUs. The inability to share or pool memory resources between CPUs and GPUs also in a heterogeneous environment is leading to underutilization of memory and an overall increase in total cost of ownership.
Leo CXL memory controllers from Astera Labs have been paving the paths to enable CXL-based memory pulling, and if there is one takeaway I can ask you guys to take from this whole demonstration is that we are the first to M\market to enable memory pooling, expansion, and sharing and are already deploying to several customers and is also ready for next generation of cloudscale deployment.
Thank you.
About This Video
At OCP 2022 (Open Compute Project Summit), Anil Sharma from Astera Labs demonstrated the industry's first-ever live CXL memory pooling between two CPUs — using the Leo CXL smart memory controller. The demo showed a single Leo device appearing as two separate numactl nodes, demonstrating that 128GB of CXL-attached DRAM could be pooled and shared between two Intel CPUs. The system passed industry-standard stressapp testing with no errors. The significance: CXL memory pooling addresses a fundamental problem in modern servers — memory tied to individual CPUs goes underutilized, increasing total cost of ownership. Leo's pooling capability lets multiple processors share a single pool of memory, improving utilization across heterogeneous environments. "This is the first to market to enable memory pooling, expansion, and sharing." For <span class="strong">product demo video production</span> in the semiconductor space, a conference floor demo needs to be captured with clean audio, good screen visibility, and clear narration — the technical audience is watching to verify the claims, not just understand them.
How do you film a live technical demo at an industry conference like OCP?
Plan for noise and set up as close to the demo hardware as possible. Conference floors are loud. A wired lapel mic or a directional shotgun mic positioned close to the presenter will handle the ambient noise better than an on-camera microphone. The screen needs to be legible on video — if you can't read the terminal output in the footage, the demo doesn't work as a video. Consider recording the screen separately if the physical monitor will be hard to capture. Narration should explain what's on screen at each step.
What should a CXL memory pooling demo video include?
The problem statement (memory underutilization in multi-CPU servers), the system configuration (how Leo is connected to the CPUs), the live demonstration (lspci showing CXL device enumeration, numactl showing memory pooling between CPU nodes), and the verification step (stressapp passing with no errors). Each step should be narrated: what are you doing, what does it show, why does it matter. A demo that just shows screens without narration loses viewers who can't parse the terminal output.
How do you produce a product demo video that works both as a live capture and as an edited asset?
Shoot with editing in mind: get a clean take of the narration, make sure the screen is legible throughout, and capture a clear conclusion. At OCP, the demo was done in a booth environment — the edited version can tighten the narration, cut pauses, and add text overlays to explain what's on screen. The raw conference capture is evidence that the demo happened live; the edited version is the asset that lives on the company's website or YouTube channel.
How long should a live conference product demo video be?
Five to eight minutes for a complete technical demo with context, live demonstration, and summary. Shorter is fine if the demo itself is tight — but cutting context to get under five minutes usually hurts the video, because the audience needs to understand the problem before they can evaluate the solution. The demo sections should run as long as they need to — don't cut the verification step just to hit a runtime target. For a semiconductor audience, credibility matters more than brevity.
