Same bits out.
A quarter fewer on the wire.
Lossless GPU compression for AI training and inference. Weights and gradients cross the network, land on disk and sit in memory 25% to 49% smaller, then come back bit for bit at hundreds of GB/s, on under 7% of one GPU.
Smaller everywhere it matters. Bit for bit.
bf16 and fp32, weights and gradients, from production open-source LLMs. Every tensor decompresses to exactly the original bits, so training and inference are untouched: no accuracy loss, no tuning, no convergence risk.
Size reduction by data type. Translucent tops show the range across the models tested.
Faster than the network it feeds.
Compression and decompression run at hundreds of GB/s on a small slice of one GPU, so they work inline: the bytes shrink before they hit the link and expand the moment they arrive.
Blue band: a typical network link per GPU, 50 to 100 GB/s with 400 to 800 Gb/s NICs. Bars: decompression in GB/s on 10 SMs.
| Data | Codec | Smaller | Decompress | Compress | Bit-exact |
|---|---|---|---|---|---|
| bf16 gradients | LX-B1 | 24.5% | 263 GB/s | 223 GB/s | 290 / 290 |
| bf16 weights | LX-B1 | 24.7% | 281 GB/s | 252 GB/s | 218 / 218 |
| fp32 gradients | LX-F2 high ratio | 29 to 49% | 148 to 175 GB/s | 62 to 76 GB/s | 491 / 491 |
| fp32 gradients | LX-F1 high speed | 20 to 25% | 196 to 200 GB/s | 86 to 88 GB/s | 491 / 491 |
| fp32 gradients | LX-F3 fastest† | 13 to 19% | 221 to 232 GB/s | 276 to 303 GB/s | 491 / 491 |
| fp32 weights | LX-F3 fastest† | 14.6% | 253 GB/s | 317 GB/s | 201 / 201 |
† LX-F3 also runs inside a single fused compress, transfer and decompress kernel. On one B200, with the sender and receiver each on their own 10 SMs, it moves 217 to 234 GB/s of compressed data end to end, bit-exact.
Throughput is compressed GB/s on 10 SMs; more SMs scale it close to linearly. Ranges cover the models tested: Qwen2.5-0.5B and TinyLlama-1.1B gradients, SmolLM2-1.7B and TinyLlama-1.1B weights.
Runs on a sliver of one GPU.
Every number on this page was measured on exactly 10 of the B200's 148 streaming multiprocessors, under 7% of the chip. The other 138 keep training. No extra hardware, no new cards, nothing to rack.
Measured, not projected.
- One streamThe whole model file is compressed into a single stream.
- Separate processThe stream is decompressed by a different process, with nothing shared.
- Bit for bitEvery tensor is compared with the original, bit for bit.
- End-to-end timingTiming covers every step over the whole file, on exactly 10 SMs of an NVIDIA B200.
Up to $147M back from the network.
Gradient exchange between nodes runs over the network, typically 50 to 100 GB/s per GPU. The codecs keep pace with that link on a small share of the GPU, so the network carries 25% to 49% fewer bytes and every training step waits less.
| Step time waiting on cross-node traffic | Step time saved, bf16 | Value on $1 billion |
|---|---|---|
| 10% | up to 2.5% | up to $24.5M |
| 20% | up to 4.9% | up to $49M |
| 30% | up to 7.4% | up to $73.5M |
With fp32 gradients at 29% to 49% fewer bytes, the same scenarios are worth $29M to $147M.
Share of step time spent waiting on the network. Solid: bf16 gradients. Translucent: the fp32 ceiling.
Every place a tensor is stored or moved.
Faster multi-node training
25% to 49% fewer bytes in every gradient exchange, compressed faster than the link can carry them.
Smaller, faster checkpoints
Checkpoints 25% smaller for bf16 weights and up to 49% for fp32 gradients. Faster writes and restarts after failures mean less idle GPU time, and storage and egress costs fall in step.
Faster loading and offload
bf16 weights arrive 25% sooner over PCIe or from storage: shorter cold starts, quicker model swaps across inference fleets and leaner CPU-offload pipelines.
More model per GPU
bf16 weights kept compressed in memory free about a quarter of their space, for larger batches, longer contexts or more cached requests.
Put a quarter of your bytes back to work.
We measure it on your models, your parallelism and your network, then pilot on one workload.
Request a pilotSavings scenarios are illustrative upper bounds. They assume communication time scales with bytes sent and that compression overlaps with the transfer. Actual results depend on model, parallelism strategy, network and workload. NVIDIA, B200 and the NVIDIA logo are trademarks of NVIDIA Corporation, used only to identify the hardware tested; weights.zip is not affiliated with or endorsed by NVIDIA.