$200 'socketed' Nvidia AI GPU for servers hacked into a PCIe card with custom PCB and 3D-printed cooling — modded Tesla V100 SMX data center GPU runs AI LLMs and is more efficient than many modern midrange offerings in AI inference - Tom's Hardware

**Skyler Monroe** (0:10)
Hey, everyone, welcome back to the AI Hardware Show. I'm your host Skyler Monroe, and this is the show where we get into the nuts and bolts, the silicon and the solder of everything powering the AI revolution. Big thanks to our sponsors today, Ada, who helps businesses actually integrate AI into their real-world workflows. Ago Consulting, that's Ago, your go-to for silicon development from AI chips to full SoC design. And Zen Semiconductor, an incredible umbrella company building the processors and AI fabric behind modern data centers, with ventures like their Sierra RISC vCPUs and Loom AI fabric, plus Ada for deploying AI at scale in any business. Today we've got a seriously fun lineup, a $200 data center GPU getting a wild DIY makeover, a new challenger claiming their chip is 10 times faster than a GPU, China cooking up a memory bandwidth monster, Nvidia teaming up with Silvaco on chip design and a deep dive into TSMC's advanced packaging tech. Let's get into it. All right, first up, and honestly, this one made me smile. Someone has taken an old Nvidia Tesla V100 Smx data center GPU, which you can now pick up for around $200 on the second hand market and turned it into a fully functional PCIe card using a custom PCB and wait for it, 3D printed cooling. I love this community. So let's back up a second and talk about what the V100 Smx actually is. The V100 was Nvidia's flagship data center GPU, back in 2Otherwin 17, 2Otherwin 18 It was the workhorse of early AI training. We're talking Volta architecture, 640 tensor cores, 32 gigabytes of HBM2 memory with around 900 gigabytes per second of memory bandwidth. When this thing launched, it cost somewhere between $8,000 and $10,000.
So seeing it go for $200 now is a wild reminder of how fast this industry moves. Now the SMX form factor is the interesting part here.
SMX is a proprietary connector format Nvidia uses for their NvLink capable server GPUs. It's designed to plug into specialized server boards. You're not just dropping it into your desktop PCIe slot. So this person essentially built a bridge. They designed a custom PCB that converts the SMX connector into a standard PCIe interface that any motherboard can talk to. That is not a trivial engineering task, by the way. PCIe lane routing, power delivery, signal integrity, that stuff matters. And then for cooling, instead of the big blower coolers you'd typically see on server GPS, they went with a 3D printed shroud and what sounds like a custom heat sink solution. The V100 SMX pulls around 300 watts, so thermal management is no joke. Getting that heat out without a proper server chassis is a real engineering challenge, and the fact that they made it work is genuinely impressive. But here's where it gets really interesting from an AI perspective. The V100 is actually holding its own and LLM inference workloads against a lot of modern mid-range consumer GPUs, and that shouldn't be entirely surprising if you think about it. Inference is extremely memory bandwidth hungry, you're constantly loading model weights, doing matrix multiplications, and the V100's HBM2 memory gives it a significant edge over GDDR6-based cards in that regard. Think of it like this.
If training AI is like building a library, inference is like a librarian rapidly pulling books off the shelves and reading them out loud.
You need fast access to the stacks. HBM is like having the books right next to you on a desk. GDDR is like having them in a room down the hall. The implication here is pretty meaningful for the broader community. There are thousands of these V100 Smx cards sitting in decommissioned data center hardware right now, and at $200 a pop, this mod could open up serious AI capability for hobbyists, researchers, and small labs who can't afford current generation of hardware. If the custom PCB design gets open-sourced, which I hope it does, this could become a real thing. It also makes a broader point about the secondhand GPU market.
Enterprise hardware has this weird depreciation curve where the price collapses once it gets a few generations old, even though the raw compute capability is still genuinely useful. The community is getting smarter about exploiting that gap, and I think we're going to see a lot more of this kind of creative hardware hacking. Next story, Nvidia and Silvaco have announced a partnership aimed at elevating semiconductor design using AI.
Now, Silvaco might not be a household name for everyone listening, but in the EDA world, that's electronic design automation, they're well known. Silvaco makes software tools that chip designers use to simulate, verify, and actually build chip designs. So what does an Nvidia partnership mean here? Essentially, the idea is to bring Nvidia's AI computing capabilities, probably GPU accelerated simulation, and potentially some of their Calytho or AI assisted design tools into Silvaco's existing EDA workflows. This is part of a bigger trend we've been watching where AI is being used to design better AI chips. It's beautifully recursive. Think about what goes into designing a chip. You're talking about billions of transistors, miles of metal interconnects, thermal simulations, signal integrity analysis, power distribution networks. The complexity is almost incomprehensible. Traditional EDA tools are powerful, but they're computationally expensive and slow.