Megabyte per second to Tebibyte per second

Many USB drives use a small SLC cache for initial writes at high MB/s, then slow dramatically once the cache fills and data writes to slower TLC/QLC NAND. A drive that starts at 200 MB/s might drop to 20–30 MB/s after the first few gigabytes. Check sustained write speed reviews, not just peak numbers.

Editing 4K ProRes footage requires about 200–400 MB/s of sustained read speed. 8K RAW can demand 1,000+ MB/s. A SATA SSD (550 MB/s) handles 4K fine, but 8K workflows really need NVMe drives at 3,000+ MB/s. The timeline scrubbing experience directly correlates with MB/s.

Look at the capitalisation: lowercase "b" (Mbps) means megabits, uppercase "B" (MB/s) means megabytes. Most speed test websites (Speedtest by Ookla, fast.com) default to Mbps. If your result seems 8× lower than expected, you are probably reading MB/s where you expected Mbps.

PCIe 5.0 NVMe SSDs hit 12,000–14,000 MB/s sequential read speeds. That is fast enough to load an entire 50 GB game in about 4 seconds. PCIe 6.0 drives, expected soon, will double this again to roughly 25,000 MB/s.

Network transfers add latency, protocol overhead (SMB, NFS), and are limited by the network link speed. A file on a local NVMe SSD reads at 7,000 MB/s, but sharing it over a 1 Gbps network caps throughput at 125 MB/s. Even 10 GbE only gives 1,250 MB/s — a fraction of modern SSD capability.

AMD's MI300X stacks 8 HBM3 memory modules and multiple compute chiplets on a single package using advanced 2.5D packaging with silicon interposers. The short physical distance between compute and memory dies — millimeters instead of centimeters — dramatically reduces signal latency and power per bit. This allows a 5.3 TB/s aggregate bandwidth that would be physically impossible with traditional socketed memory. The trend toward chiplet packaging is how the industry keeps scaling bandwidth despite hitting limits in single-die manufacturing.

Significantly. When provisioning an AI training cluster with hundreds of GPUs, a 10% bandwidth miscalculation cascades through the entire system design — buffer sizes, interconnect capacity, cooling, and power. Getting the units wrong could mean the difference between a training run finishing in 30 days vs 33 days.

Training large language models (100B+ parameters), molecular dynamics simulations, weather modeling, and fluid dynamics at scale. These workloads move enormous matrices through memory billions of times. The TiB/s memory bandwidth of modern GPUs is what makes training models like GPT-4 possible in months rather than decades.

Memory bandwidth dwarfs network bandwidth. Each H100 GPU has 3.35 TiB/s of internal memory bandwidth but connects to the network at only 0.05 TiB/s (400 Gbps InfiniBand). This 60:1 ratio is why AI chip designers obsess over keeping computations local to each GPU and minimising network communication.

Not in the same way. Quantum computers process information through qubits that exist in superposition, so they do not shuttle classical data around at TiB/s. However, the classical control systems that manage quantum processors and process measurement results do need high bandwidth — current quantum-classical interfaces operate at modest Gbps rates.