AI-Native Memory Systems
from Silicon to Tokens
The World's First Memory Processing Unit (MPU) Chip — Unlocking New Frontiers for Training, Inference and Agentic AI.
Learn more about Netpreme X-Mem™Animated hero concept: a unit of two MPU™ memory nodes and two XPU compute nodes meshed through a scale-up fabric, with data streaming between them; the view then pulls back to show the same unit tiling outward across the rack.
X-Mem™ Technology: the Memory Tier Built for AI
- GPUG1HBM10 TB/s · 0.3 TB
- G1.5 & G2.ExpX-Mem™ MPU3.6 TB/s · 8 TB
- G1HBM10–20 TB/s · 300 GB
- G1.5 & G2.ExpX-Mem™ Tier3.6 TB/s · 8 TB
- G2System DRAM128 GB/s · 1–6 TB
- G3Local SSDs64 GB/s · 20 TB
- G4Shared Storage16 - 64 GB/s · Elastic
- XPU-Native Memory Tiering Library
- Transparent to XPU kernels and inference / training runtimes
- Verified solution for KV/prefix caching, long-context inference, multi-model serving, and optimizer/activation offloading
- 7×
faster TTFT than DRAM offload on prefix-heavy workloads. Read more ↗︎
- 4×
higher concurrency with the same number of GPUs.
- 10x
cheaper inference for long-horizon agents. Try it now ↗︎
Industry Perspectives
(formerly VP/GM at Intel Foundry, SVP at Samsung and VP at Micron)
Netpreme has the team and the technology to break the memory wall in AI inference.
Its outstanding MIT-founded team of more than 20 engineers combines silicon design with deep AI systems and software expertise. Its Memory Processing Unit creates a new warm memory tier that extends accelerator memory by up to 100x at scale-up bandwidth, backed by a strong IP portfolio. The results are compelling: on long-context inference, X-Mem delivers 6.7x faster time to first token (1), and ~1.87x more tokens per second per dollar (2). This is transformative for end-user productivity, differentiated services for operators and investor ROIC.”
(1) The 6.7x is X-Mem vs. host DRAM offloading at an 80K-token prompt, from a 10K to 80K prompt sweep. Methodology follows the vLLM KV offloading benchmark.
(2) Netpreme modeling, not a measurement. Kimi K2.6 (MLA), 32K-token input, 95% KV cache reuse, prefill only, prefill/decode disaggregated. Tokens per second per dollar is +87% vs. the same platform without X-Mem.