We explored a network-attached KV-cache for consumer GPUs to offset their limited VRAM. It doesn’t make RTX cards run giant models efficiently. Still, for workloads that repeatedly reuse lengthy prefixes—such as chatbots, coding assistants, and multi-turn threads—it delivers a 2–4× speedup in RPS and time-to-first-token on 7B and 70B models.
How it works:
On return visits, instead of re-running the prompt through the model, we fetch previously computed KV blocks from network storage and skip re-computing those tokens (i.e., we avoid re-running prefill on repeated prefixes). This is helpful when VRAM can’t hold all sessions, and users pause between messages, which is almost always the case.
Why RTX benefits:
Prefill is the computationally intensive part (quadratic attention, numerous reductions, and inter-GPU traffic). Without NVLink, PCIe becomes the choke point in multi-GPU setups. KV-caching cuts repeated prefill, leaving mostly the lighter decoding step—something PCIe-only RTX nodes handle well.
Results & endpoint:
- 2–4× speedup on multi-turn benchmarks (RPS & TTFT) with RTX 4090.
- We’ve opened one free public endpoint for demos, not production grade (https://console.cloudrift.ai/inference?modelId=meta-llama%2F...). Ping us at hello@cloudrift.ai if you need a reliable setup.
Technical Notes:
- Works with consumer and data-center GPUs. In theory, you can even split roles: NVLink boxes do prefill, while cheaper RTX pods serve as decoders using stored KV.
- We use special hardware to reduce fetch overhead and offload the CPU, but you can reproduce this at home with a regular NAS (with lower peak performance).
- For a more in-depth walkthrough of the math and architecture of a KV-cache solution, please watch this video from the KV-cache solution vendor (https://www.youtube.com/watch?si=T69vxku8xPr6p7I0&v=CV4FYMTF...)
We explored a network-attached KV-cache for consumer GPUs to offset their limited VRAM. It doesn’t make RTX cards run giant models efficiently. Still, for workloads that repeatedly reuse lengthy prefixes—such as chatbots, coding assistants, and multi-turn threads—it delivers a 2–4× speedup in RPS and time-to-first-token on 7B and 70B models.
How it works: On return visits, instead of re-running the prompt through the model, we fetch previously computed KV blocks from network storage and skip re-computing those tokens (i.e., we avoid re-running prefill on repeated prefixes). This is helpful when VRAM can’t hold all sessions, and users pause between messages, which is almost always the case.
Why RTX benefits: Prefill is the computationally intensive part (quadratic attention, numerous reductions, and inter-GPU traffic). Without NVLink, PCIe becomes the choke point in multi-GPU setups. KV-caching cuts repeated prefill, leaving mostly the lighter decoding step—something PCIe-only RTX nodes handle well.
Results & endpoint: - 2–4× speedup on multi-turn benchmarks (RPS & TTFT) with RTX 4090. - We’ve opened one free public endpoint for demos, not production grade (https://console.cloudrift.ai/inference?modelId=meta-llama%2F...). Ping us at hello@cloudrift.ai if you need a reliable setup.
Technical Notes: - Works with consumer and data-center GPUs. In theory, you can even split roles: NVLink boxes do prefill, while cheaper RTX pods serve as decoders using stored KV. - We use special hardware to reduce fetch overhead and offload the CPU, but you can reproduce this at home with a regular NAS (with lower peak performance). - For a more in-depth walkthrough of the math and architecture of a KV-cache solution, please watch this video from the KV-cache solution vendor (https://www.youtube.com/watch?si=T69vxku8xPr6p7I0&v=CV4FYMTF...)