SSD
A solid-state drive based on flash memory storage, with read and write speeds higher than those of hard disk drives but lower than those of video memory and main memory, and relatively poor performance for small random-block reads.
Timeline (last 90 days)
A fivefold expansion of tape capacity does not mean tape will replace disks or flash; online business, real-time inference, and hot data caching still rely on SSDs and memory.
Through four optimizations—expert-aligned chunking, hot expert residency, compute splitting and scheduler preloading—Wi-Sci raises effective SSD bandwidth from 1 GB per second to about 11.5 GB per second, enabling DeepSeek 4.1 Flash to reach a generation speed of about 8-15 tokens/second on the Wi-Sci box.
As Apple fully switched to solid-state drives (SSD), SMS's hard drive protection function lost its original meaning, but the accelerometer itself remains on the motherboard.
DeepSeek uses a dedicated data cluster equipped with about 100 PB of SSD storage to support its long-duration caching strategy.
SMART attributes are a set of health metrics that SSD firmware periodically collects and reports to the operating system, covering multiple dimensions such as temperature, power-on hours, write wear, error rate, and bad sectors.
Local SSDs have microsecond-level random read latency and a fixed hardware cost model, which is entirely different from S3's billing model.
The SSD (State Space Duality) framework proposed by Mamba-2 reveals that SSMs with a specific structure are mathematically equivalent to a masked form of linear attention, allowing training to be parallelized using matrix multiplication.
A post on Hacker News shows streaming expert weights from an SSD to run large models that exceed memory capacity, receiving 18 points.
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.
Monitoring the rMB/s and maj_flt growth rate of iostat -xm 1 nvme0n1 can determine whether the SSD has become a bottleneck; if vmtouch reports that the resident ratio has long hovered around 40% and iostat continuously shows large read traffic, it can basically be confirmed that the SSD is the bottleneck.
22 more timeline events
All Facts (20)
内存访问比SSD快约1000倍，比机械硬盘快约10万倍
85%Unverified业界通常建议至少保留10%-20%的SSD空间为空闲状态以维持最佳性能
80%UnverifiedSSD当存储空间接近满载时性能会显著下降，写入速度可能降至正常水平的50%甚至更低
70%UnverifiedA fivefold expansion of tape capacity does not mean tape will replace disks or flash; online business, real-time inference, and hot data caching still rely on SSDs and memory.
50%UnverifiedThrough four optimizations—expert-aligned chunking, hot expert residency, compute splitting and scheduler preloading—Wi-Sci raises effective SSD bandwidth from 1 GB per second to about 11.5 GB per second, enabling DeepSeek 4.1 Flash to reach a generation speed of about 8-15 tokens/second on the Wi-Sci box.
50%UnverifiedAs Apple fully switched to solid-state drives (SSD), SMS's hard drive protection function lost its original meaning, but the accelerometer itself remains on the motherboard.
50%UnverifiedDeepSeek uses a dedicated data cluster equipped with about 100 PB of SSD storage to support its long-duration caching strategy.
50%UnverifiedSMART attributes are a set of health metrics that SSD firmware periodically collects and reports to the operating system, covering multiple dimensions such as temperature, power-on hours, write wear, error rate, and bad sectors.
50%UnverifiedLocal SSDs have microsecond-level random read latency and a fixed hardware cost model, which is entirely different from S3's billing model.
50%UnverifiedThe SSD (State Space Duality) framework proposed by Mamba-2 reveals that SSMs with a specific structure are mathematically equivalent to a masked form of linear attention, allowing training to be parallelized using matrix multiplication.
50%UnverifiedMonitoring the rMB/s and maj_flt growth rate of iostat -xm 1 nvme0n1 can determine whether the SSD has become a bottleneck; if vmtouch reports that the resident ratio has long hovered around 40% and iostat continuously shows large read traffic, it can basically be confirmed that the SSD is the bottleneck.
50%UnverifiedThe MoE architecture allows running large models beyond memory, because each token activates only a small portion of the expert weights, and the remaining expert weights can stay on the SSD and be loaded on demand.
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 streaming MoE experts, each token may hit a different expert, causing weight reads to follow a random pattern, and the actual effective bandwidth may be only a fraction of sequential reads.
50%UnverifiedThe core claim of the Janas project is to run large language models that exceed memory capacity by streaming experts from an SSD.
50%UnverifiedWhen streaming experts from an SSD, if the expert weight blocks are large (a single expert may be several hundred MB), the read granularity can approach sequential, and the performance loss is relatively controllable.
50%UnverifiedIf a model is too large for the memory, the Mac will swap data to the SSD, causing generation speed to drop to less than one word per second and causing severe long-term SSD wear.
50%UnverifiedThe TBW of mainstream consumer-grade SSDs typically ranges from 150 to 600 TB, while enterprise-grade products can reach several thousand TB.
50%UnverifiedThe random write speed of SD cards is typically only 2–5 MB/s, far below the 100 MB/s or more of SSDs.
50%UnverifiedThe Ngram lookup table can be streamed from SSD without seriously slowing down inference, because the memory bandwidth requirements of the table lookup operation are far lower than those of the matrix operations in neural network layers.
50%