Frore's new LiquidJet coldplates are designed to handle up to 4,400W of power from Nvidia's Feynman AI GPUs — new design uses semiconductor manufacturing techniques to create 3D jet-channel microstructures

Frore Systems
(Image credit: Frore Systems)
7

Frore Systems has introduced LiquidJet, a coldplate that supports existing AI GPUs, such as Nvidia's Blackwell, with a 1,400W thermal design power. It can scale its performance for next-generation processors, such as Nvidia's Feynman, which will feature a total power of 4,400W. The new coldplates feature 3D short-loop jet-channel microstructures that increase hotspot power density of a whopping 600 W/cm^2 and reduce pressure loss by four times compared to conventional coldplates. In the end, Frore's LiquidJet is ready for multi-kilowatt AI GPUs set to arrive in the coming years.

Modern AI GPUs like Nvidia's Blackwell AI system-on-chips (SoCs) consume an extreme amount of power and therefore require liquid cooling. Modern copper coldplates feature relatively long 2D microchannels with small cross sections that are CNC-machined or skived from high-purity copper blocks. Because these microchannels are long, liquid must travel farther and rub against more surface area, which increases hydraulic resistance and reduces pressure, which affects performance. Frore says that its LiquidJet coldplates with 3D short-loop jet channel microstructures reduce hydraulic resistance and therefore maintain higher pressure inside to boost performance.

Latest Videos FromTom's Hardware
Anton Shilov
Contributing Writer

Anton Shilov is a contributing writer at Tom’s Hardware. Over the past couple of decades, he has covered everything from CPUs and GPUs to supercomputers and from modern process technologies and latest fab tools to high-tech industry trends.

  • edzieba
    Looks a lot like the old IBM Aquasar warm-loop concept, but applied to gain higher deltas at lower temperatures rather than operate at lower deltas at higher temperatures.
    Reply
  • Jame5
    So glad not a single company involved in any of this is considering how to do more work per watt, rather than just throwing more watts at the problem.
    Reply
  • bolweval
    Jame5 said:
    So glad not a single company involved in any of this is considering how to do more work per watt, rather than just throwing more watts at the problem.
    "You can please some of the people all of the time, you can please all of the people some of the time, but you can't please all of the people all of the time"
    John Lydgate
    Reply
  • rluker5
    Jame5 said:
    So glad not a single company involved in any of this is considering how to do more work per watt, rather than just throwing more watts at the problem.
    They are all trying to get more per watt, but the node improvements aren't what they used to be, so to increase performance they have to increase watts.

    There was also just news about Intel making that low power 160GB vram inference card: https://www.tomshardware.com/pc-components/gpus/intel-unveils-crescent-island-an-inference-only-gpu-with-xe3p-architecture-and-160gb-of-memory
    Reply
  • thestryker
    Jame5 said:
    So glad not a single company involved in any of this is considering how to do more work per watt, rather than just throwing more watts at the problem.
    They are actually optimizing to get more work per watt, but the fact of AI as it stands today is they want maximum performance. This means that while everyone's happy about perf/watt optimizations delivering that with another 50% power is even better. That is what's different about this AI boom than normal datacenters or supercomputers: power consumption is something to build around rather than target for.
    Reply
  • Notton
    Nvidia can absolutely do more work per watts.
    The cost is silicon size, which would be extremely costly for Nvidia.

    Just compare the RTX4000 ADA SFF to 4060Ti 16GB
    It's like 70W vs 165W for similar performance.
    Reply
  • BoobleKooble
    Jet flow in water blocks is normally a performance boost, but true jet flow conditions usually require that a system be operated at high pressure. Jets can break up the boundary layer
    Reply