TP16 can turn Kimi K3 parallelism into repeated storage reads
By OnTheRice · Sources and limitations · Report a correction
TP16 can turn GPU parallelism into repeat storage reads instead of a faster Kimi K3 load. TensorRT-LLM's issue says each TP16×EP1 rank reads about 1.45 TB of expert weights but keeps roughly 90 GB; four ranks per node can request 5.8 TB. The reporter measured 1.2 TB read per node after 11 minutes on 16 GB200 GPUs across four nodes, with loading unfinished. TP4×EP4 readers split the expert quarters and read about 1.45 TB per node. This is one checkpoint/configuration, not a general trend; next compare bytes read and load time after a sharing fix.
Source confidenceNot assessed
What’s changing?
With lazy loading at TP16 x EP1, every rank reads all experts then retains only its slice; page dropping can force node-local rereads. TP4 x EP4 ranks read disjoint expert quarters.
Key points
- TensorRT-LLM issue #19849 reports a Kimi K3 MXFP4 checkpoint with 92 MoE layers and 896 experts per layer. At TP16 x EP1, each rank reads about 1.45 TB of routed experts while keeping about 90 GB; four ranks per node request about 5.8 TB. The issue reporter observed 1.2 TB per node read after 11 minutes on 16 GB200 GPUs across four nodes, with load unfinished. TP4 x EP4 reads disjoint expert quarters and about 1.45 TB per node. One comment asks for engineer approval/direction on an external contribution; no patch/approval is recorded.
What could change this view?
- One issue-reporter benchmark on one Kimi K3 MXFP4 checkpoint and network-storage run; not independent. The issue suggests fixes; a single external contribution still awaits engineer approval and no patch is linked.
Bottom line
This can increase checkpoint storage traffic and delay startup for this Kimi K3 layout; TP4 x EP4 is the issue’s comparison and reads disjoint expert quarters.
This is based on the information recorded with this signal.
How likely is this call? · Three frameworks
This signal has no investible bullish or bearish call. A market-direction percentage does not apply.
Evidence behind this view
Checks whether separate pieces of evidence support the same view.
Recorded connection
With lazy loading at TP16 x EP1, every rank reads all experts then retains only its slice; page dropping can force node-local rereads. TP4 x EP4 ranks read disjoint expert quarters.
We have not established a separate count of independent evidence clusters for this card. Several reports may rely on the same event or data provider.
What past price moves show
Compares today’s price moves with recorded past patterns.
This card has no recorded investible call. A price-pattern comparison is not applicable unless a specific asset is added with evidence.
Evidence against this view
Shows evidence and risks that could challenge this view.
Recorded opposing evidence and risks
One issue-reporter benchmark on one Kimi K3 MXFP4 checkpoint and network-storage run; not independent. The issue suggests fixes; a single external contribution still awaits engineer approval and no patch is linked.
Why the direction changed
Explains recorded changes from expecting a rise to expecting a fall, or the reverse.
No documented bullish-to-bearish or bearish-to-bullish change is available for this card. We keep the original call and its time unchanged.
What people say vs the facts
Compares recorded public discussion with checked market facts.
No separate Reddit or public-discussion comparison with verified market data is recorded for this card. We cannot treat social attention as proof of a market move.
How past forecasts performed
Shows recorded forecast results and what they can teach us.
This card has no recorded investible forecast to score.
No tested mapping from evidence combinations to future success is recorded for this card. Confidence in sources and past outcomes do not establish a calibrated probability.
Sources, original wording and price checks
Published
First published through Sites 05 Oct 2026, 01:54:01 SGT
No explicit investible BULLISH or BEARISH call on this card.
What changed
Recorded public historyTP16 can turn GPU parallelism into repeat storage reads instead of a faster Kimi K3 load. TensorRT-LLM's issue says each TP16×EP1 rank…
This is the first recorded public view. Further public updates will appear here.
- TP16 can turn Kimi K3 parallelism into repeated storage reads
TP16 can turn GPU parallelism into repeat storage reads instead of a faster Kimi K3 load. TensorRT-LLM's issue says each TP16×EP1 rank reads about 1.45 TB of expert weights but keeps roughly 90 GB; four ranks per node can request 5.8 TB. The reporter measured 1.2 TB read per node after 11 minutes on 16 GB200 GPUs across four nodes, with loading unfinished. TP4×EP4 readers split the expert quarters and read about 1.45 TB per node. This is one checkpoint/configuration, not a general trend; next compare bytes read and load time after a sharing fix.
First published by OTR: 5 Oct, 01:54 SGTPublished by OTR: 5 Oct 2026, 01:54 SGT
Why the view changed
This is the first recorded view. Future source-backed updates will appear here.
Evidence, memory & review 1 linked source
Directional takeaway: N/A: these configuration-specific software reports do not support an asset-specific listed-security conclusion.
OTR reasoning: With lazy loading at TP16 x EP1, every rank reads all experts then retains only its slice; page dropping can force node-local rereads. TP4 x EP4 ranks read disjoint expert quarters.
Source-grounded facts (paraphrased)
- Time not supplied: TensorRT-LLM issue #19849 reports a Kimi K3 MXFP4 checkpoint with 92 MoE layers and 896 experts per layer. At TP16 x EP1, each rank reads about 1.45 TB of routed experts while keeping about 90 GB; four ranks per node request about 5.8 TB. The issue reporter observed 1.2 TB per node read after 11 minutes on 16 GB200 GPUs across four nodes, with load unfinished. TP4 x EP4 reads disjoint expert quarters and about 1.45 TB per node. One comment asks for engineer approval/direction on an external contribution; no patch/approval is recorded.
Source trail
- [Performance]: Kimi K3 expert weight load reads the full experts on every rank at MoE TP16 x EP1 · Issue #19849 ↗NVIDIA TensorRT-LLM / GitHub2026-10-04T06:25:31Z · published (UTC)
Review status: SUPPORTED
No score or probability is presented.