Case Studies
Small, self-contained experiments run directly on the DGX Sparks. Each one isolates a single variable — a quantization scheme, a decode setting, a context length — and reports what actually moved, with the numbers.
Does speculative decode still pay at 128K context?
The MTP head reads the whole KV cache on every draft step, and a paper says that tax eventually turns deep drafts net-negative. We measured it at 2K, 32K and 131K on one Spark: the tax is real, and it is one tenth the size of the target's own.
How much does quantization hurt quality?
FP8 vs NVFP4 Laguna-S-2.1 on a fixed eval set — what 4-bit costs in correctness.
Can you move the speculative drafter onto another machine?
Measuring what a remote drafter would cost and free — the wire, the draft time, and the reason the experiment stalled.
Can you quantize the drafter instead of moving it?
The dspark speculator is a weight read. Shrinking it should buy most of a remote drafter's win, with no network involved.
Does quantization actually buy you speed?
Decode, prefill, and power across precision levels on one DGX Spark.