Astera Labs has launched new Leo memory controllers designed to stuff more DDR4 and DDR5 closer to CPUs and hungry AI accelerators.
According to ServeTheHome the new line-up consists of the Leo 2 E-Series, Leo 2 P-Series and Leo X-Series. Leo 2 upgrades Astera’s CPU-attached CXL memory controllers, while Leo X connects memory directly to accelerator fabrics rather than making GPUs rummage through a host CPU.
Leo 2 moves from PCIe Gen5 and CXL 2.x to PCIe 6 and CXL 3.2. PCIe 6 doubles per-lane bandwidth, giving each memory expander considerably more room to shift data without becoming an expensive bottleneck bolted to a server.
CXL 3.x brings multi-level switching, direct peer-to-peer memory access and support for multiple memory expanders behind a PCIe root port. Astera offers a single PCIe 6 x16 connection on the E-Series and dual x8 connectivity on the P-Series.
Astera has doubled the memory channels from two to four to keep the wider interface fed. Each Leo 2 device can handle up to 768GB of DDR4 or 4TB of DDR5, doubling capacity and bandwidth compared with the previous generation.
One of the stranger attractions is DDR4 reuse. Astera has a four-channel, three-DIMMs-per-channel design that can swallow 12 DDR4 DIMMs, giving hyperscalers a useful place to put memory pulled from retiring servers instead of feeding it to the recycling pile.
Old DRAM is not immortal, so Astera has added hardware testing and reliability features to keep reused DIMMs trustworthy. Leo 2 includes hardware test engines, memory-health management and automated repair, with ServeTheHome reporting that Astera layers soft ECC on top of the DIMMs’ hardware ECC.
The E-Series is the simpler part, intended for memory expansion behind one host CPU. Leo 2 P-Series adds pooling and sharing between two hosts, letting systems move capacity around rather than permanently assigning every chunk of DRAM to one processor.
Leo X is paired with Astera’s Scorpio X-Series fabric switches; it lets memory expanders sit directly on an AI accelerator fabric, creating a dedicated tier for KV caches and agent context without routing every memory request through CPU-attached DRAM or NVMe storage.
The controller supports PCIe and platform-specific interfaces, giving hyperscalers room to attach it to different accelerator architectures. Astera has demonstrated a Leo X expander with four DIMM slots that supports both DDR4 and DDR5.
Astera claims its Leo X arrangement can cut time to first token by as much as 62 per cent and increase token throughput by up to 22 per cent. There is a fairly chunky footnote attached to those numbers: ServeTheHome says Astera tested against a PCIe-based Nvidia H200 system rather than a newer Blackwell platform.
The timing is convenient for Astera, with new CXL 3.x servers arriving while DRAM capacity remains painfully expensive. The company says the expanded Leo family is now sampling with hyperscalers, with customers being pitched everything from cheap DDR4 reuse to near-GPU KV-cache capacity for long-context AI workloads.
