ABOUT
What this is, and how it's measured
A private lab journal made public. Every number on this site is a measured run on the hardware below, no vendor claims, no cherry-picked windows, and the traps, dead ends, and problems get logged right alongside the wins.
HARDWARE
GPU × 2
Intel Arc Pro B70 32 GB
card 1 · Intel reference, card 2 · ASRock Creator. One GPU per run; all current runs used card 2.
TEST PROTOCOL
fixed ~65k-token prompt (Moby-Dick, ch. 1–~44) + fixed question
temperature = 0 · max_tokens = 2048
request #1 = cold (empty prefix cache) · request #2 = warm (APC hit)
sequential single request, 1 concurrent, no parallel load during measurement
tokens counted from stream usage, not SSE chunks
temperature = 0 · max_tokens = 2048
request #1 = cold (empty prefix cache) · request #2 = warm (APC hit)
sequential single request, 1 concurrent, no parallel load during measurement
tokens counted from stream usage, not SSE chunks
HONESTY RULES
- Numbers are only as good as their protocol. The protocol is fixed and published, every run uses it.
- Speculation (MTP, speculative decoding) is labeled, never hidden, "105.8 t/s with MTP speculation" is not "105.8 t/s raw".
- When a lever doesn't help (batched-tokens 4096→8192, ubatch 4096), the negative result is published too.
- Confounded comparisons are labeled confounded, model, quant, and MTP depth all differ across engines here.
- Client timing beats server averages for TTFT; usage-based counting beats chunk counting for tokens.
CONTACT
· rss