RAM
A reinforcement learning post-training method for diffusion models proposed by TU Munich doctoral student Andreas Bergmeister. Derived from first principles, it achieves a reported 50-fold speedup in post-training by formulating a scalable training objective through Reinforce Adjoint Matching.
Timeline (last 90 days)
Sharing memory remotely over the internet is infeasible because RAM needs to be tightly integrated with the GPU and has extremely high responsiveness requirements for disk I/O.
An air-gapped computer can exfiltrate data via radio signals generated by RAM under software control, with no physical hardware implantation or network connection required.
Past research has used the electromagnetic radiation of CPUs, GPUs and even USB interfaces for air-gap exfiltration, but cases of using RAM for signal emission are relatively rare.
This RAM electromagnetic leakage attack method is still at the laboratory stage and remains far from actual weaponization.
The radio signals generated by RAM are very weak, with an effective reception distance usually within a few meters, and require high-sensitivity equipment and interference-resistant modulation schemes.
An attacker can generate regular electromagnetic pulses by repeatedly reading and writing specific memory addresses, and modulate these pulses to encode binary data.
Polina's rule of thumb is: an ANR accompanied by low available RAM is ironclad evidence of a memory leak or bloat.
Counterpoint attributes the decline in PC shipments to surging RAM prices, which have forced PC makers to pass costs on to consumers.
When RAM ≥ 70% of the expert weights, the MoE model runs at almost the same speed as running entirely in memory; when RAM is between 40% and 70%, the speed fluctuates with the prompt; when RAM < 40%, it is usually not worth running.
The expert FFN weights of MoE models occupy most of the volume, but only a small portion is read each time, so this part should be placed in RAM, and when RAM is insufficient the operating system pages it in from the SSD.
3 more timeline events
All Facts (14)
显存带宽与系统内存带宽之间相差大约10倍
85%UnverifiedSharing memory remotely over the internet is infeasible because RAM needs to be tightly integrated with the GPU and has extremely high responsiveness requirements for disk I/O.
50%UnverifiedPast research has used the electromagnetic radiation of CPUs, GPUs and even USB interfaces for air-gap exfiltration, but cases of using RAM for signal emission are relatively rare.
50%UnverifiedAn air-gapped computer can exfiltrate data via radio signals generated by RAM under software control, with no physical hardware implantation or network connection required.
50%UnverifiedAn attacker can generate regular electromagnetic pulses by repeatedly reading and writing specific memory addresses, and modulate these pulses to encode binary data.
50%UnverifiedThe radio signals generated by RAM are very weak, with an effective reception distance usually within a few meters, and require high-sensitivity equipment and interference-resistant modulation schemes.
50%UnverifiedThis RAM electromagnetic leakage attack method is still at the laboratory stage and remains far from actual weaponization.
50%UnverifiedPolina's rule of thumb is: an ANR accompanied by low available RAM is ironclad evidence of a memory leak or bloat.
50%UnverifiedCounterpoint attributes the decline in PC shipments to surging RAM prices, which have forced PC makers to pass costs on to consumers.
50%UnverifiedThe expert FFN weights of MoE models occupy most of the volume, but only a small portion is read each time, so this part should be placed in RAM, and when RAM is insufficient the operating system pages it in from the SSD.
50%UnverifiedWhen RAM ≥ 70% of the expert weights, the MoE model runs at almost the same speed as running entirely in memory; when RAM is between 40% and 70%, the speed fluctuates with the prompt; when RAM < 40%, it is usually not worth running.
50%UnverifiedWhen a model runs, its weights need to be loaded into RAM (system memory) or VRAM (video memory), and having large storage capacity does not mean the model can actually run.
50%UnverifiedRunning AI models on-device requires more RAM, which is driving up demand for memory chips across the entire industry.
50%Unverified机器学习笔记本内存建议至少16GB，系统内存与显存是完全独立且不能互相替代的两个资源
50%