GLM-5.3 Optimizations, Part 2: PCP, DCP, and Hybrid HiSparse on 16 H200s
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
vLLM's second GLM-5.3 write-up serves long agentic contexts on 16 H200s split into prefill and decode nodes. Prefill combines token-wise context parallelism (PCP8) with KV sharding (DCP8), and decode uses Hybrid HiSparse host-memory offload. On AgentX, the authors report 1.7–3.2× the per-user speed of data-parallel prefill at equal throughput, median time to first token of 0.8–0.9 s, and the only tested setup serving 1M-token contexts. Results are self-reported, from one benchmark with fixed speculative-decoding acceptance, and DCP adds 6–15% per-step overhead.
Score
0.73total under rubric v2.0
- Significance0.64
- Reader interest0.82
- Audience breadth0.45
- Practical relevance0.79
- Substantiated0.91
- Technical depth0.98
- New information0.92
- Counts against: Promotional0.10
- Counts against: Overclaim0.18
Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).
| Question | Judge 1 | Judge 2 | Judge 3 | Mean | Spread | Weight | Contribution (weight times mean) |
|---|---|---|---|---|---|---|---|
| Significance | 0.64 | 0.50 | 0.78 | 0.64 | 0.28 | +0.25 | +0.160 |
| Reader interest | 0.76 | 0.80 | 0.91 | 0.82 | 0.15 | +0.20 | +0.165 |
| Audience breadth | 0.36 | 0.40 | 0.58 | 0.45 | 0.22 | +0.10 | +0.045 |
| Practical relevance | 0.80 | 0.70 | 0.86 | 0.79 | 0.16 | +0.15 | +0.118 |
| Substantiated | 0.96 | 0.85 | 0.92 | 0.91 | 0.11 | +0.10 | +0.091 |
| Technical depth | 1.00 | 0.95 | 0.98 | 0.98 | 0.05 | +0.10 | +0.098 |
| New information | 0.93 | 0.85 | 0.99 | 0.92 | 0.14 | +0.10 | +0.092 |
| Promotional | 0.05 | 0.20 | 0.04 | 0.10 | 0.16 | −0.15 | −0.015 |
| Overclaim | 0.26 | 0.20 | 0.08 | 0.18 | 0.18 | −0.15 | −0.027 |
| Sum of contributions | +0.726 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.726 | ||||||
Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.
Claim check
Claim check: 8 no source given. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.
No source givenBenchmark result
On AgentX, PCP8 × DCP8 prefill with Hybrid HiSparse decode gives 1.7–3.2× the per-user speed of a data-parallel deployment at the same throughput.
The article gives no link or citation for it.
The article says “PCP8 × DCP8 prefill with Hybrid HiSparse decode gives 1.7–3.2× the per-user speed of a data-parallel deployment at the same throughput”
The source says No source to compare with.
No source givenFirst, only or fastest
PCP8 × DCP8 with Hybrid HiSparse is the only configuration the authors tested that serves 1M-token contexts.
The article gives no link or citation for it.
The article says “it is the only configuration we tested that serves 1M-token contexts”
The source says No source to compare with.
No source givenNumber
PCP8 × DCP8 prefill cuts median TTFT from 1.5–5.2 s (DP8) to 0.8–0.9 s.
The article gives no link or citation for it.
The article says “cuts median TTFT from 1.5–5.2 s to 0.8–0.9 s”
The source says No source to compare with.
No source givenBenchmark result
At c64, HiSparse decode nearly doubles throughput to 8,340 logical tok/s/GPU at 29.6 tok/s/user, while every other configuration drops to 1–2 tok/s/user.
The article gives no link or citation for it.
The article says “At c64 it nearly doubles throughput, to 8,340 logical tok/s/GPU at 29.6 tok/s/user, while every other configuration drops to 1–2 tok/s/user”
The source says No source to compare with.
No source givenNumber
A 1M-token GLM-5.3 request needs 53.4 GiB of FP8 KV cache, while each GPU has room for only about 15 GiB; DCP8 reduces this to 6.7 GiB per GPU.
The article gives no link or citation for it.
The article says “A 1M-token GLM-5.3 request needs 53.4 GiB of FP8 KV,3 but each GPU has room for only about 15 GiB”
The source says No source to compare with.
No source givenNumber
DCP8 adds about 6% per step overhead up to 64K tokens of context, rising to 15% at 248K.
The article gives no link or citation for it.
The article says “DCP8 adds about 6% per step up to 64K tokens of context, rising to 15% at 248K”
The source says No source to compare with.
No source givenNumber
PCP raises sparse MLA's arithmetic intensity from about 26 to about 191 FLOPs per byte moved.
The article gives no link or citation for it.
The article says “raising sparse MLA from about 26 to about 191 FLOPs per byte moved”
The source says No source to compare with.
No source givenNumber
The indexer overtakes sparse MLA under TP8 at about 32K tokens, while under PCP8 the crossover moves out to about 256K.
The article gives no link or citation for it.
The article says “the indexer overtakes sparse MLA under TP8 once the prompt reaches about 32K tokens; under PCP8, where no GPU replicates it, the crossover moves out to about 256K”
The source says No source to compare with.
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN —
Recorded for context only, checked 24 Oct 2026, 08:53 UTC. It played no part in the ranking.