Glossary.
The terms we actually use when we talk about GPU clouds — in plain language, with the boring parts left in where they matter.
HBM
High Bandwidth Memory — stacked DRAM placed next to the compute die. The reason modern accelerators move terabytes per second instead of gigabytes.
FP8
An 8-bit floating-point format used heavily in AI training and inference. Roughly twice the throughput of FP16 for workloads that tolerate reduced precision.
FSDP
Fully Sharded Data Parallel — a technique that shards model parameters, gradients, and optimizer state across devices to train models larger than any single GPU's memory.
RDMA
Remote Direct Memory Access — letting one machine read another machine's memory over the network, bypassing both CPUs. The foundation of fast multi-node training.
Fabric
The high-speed interconnect linking GPUs across nodes. Fabric bandwidth, not GPU count, usually decides whether a cluster trains at full speed.
Checkpoint
A snapshot of training state — weights, optimizer state, RNG state — used to resume a run after a failure. Elastic checkpoints can restart from any shard.
Tensor Parallel
Splitting a single layer's operations across multiple GPUs. Requires very high interconnect bandwidth — it's why fabric specs matter.
P99
The latency below which 99% of requests fall. For inference, P99 tells the truth about user experience that average latency hides.
Confidential Compute
Execution with memory encrypted against the host itself, plus hardware attestation — so even the platform can't read your weights.
LoRA
Low-Rank Adaptation — fine-tuning only a small set of extra parameters instead of the whole model. Cheap, fast, and easy to swap between tasks.
MFU
Model FLOPs Utilization — the fraction of theoretical compute your training run actually uses. Great for comparing runs, dangerous for obsessing over.
Single-Tenant
A node dedicated to one customer. No neighbors on the silicon, no shared caches — the isolation profile serious training runs expect.
Ready to push the frontier?
Provision bare-metal GPUs in minutes, scale capacity on demand, and get a real engineer on call — not a ticket queue.