ResearchPublished All news from that day
Nvidia reports its Reshard call cuts DeepSeek-V3 weight refit from 14.9 to 2.6 seconds in its own tests.
Nvidia's engineers report 2.7x to 5.7x faster transfer of updated weights from training to inference GPUs in reinforcement learning. The figures are Nvidia's own benchmarks.
Check our sources · 7 facts from 1 sourceNvidia's engineers report that a new communication call called Reshard cut the time to move updated DeepSeek-V3 671B weights from training GPUs to inference GPUs from 14.9 seconds to 2.6 seconds on DGX H100 systems. The figures come from NeMo RL benchmarks in an Nvidia developer blog post published on 6 November 2026.
The step is called weight refit in reinforcement learning: after each training update, the new weights have to reach the GPUs that generate the next answers. On DGX GB200 systems the same model went from 6.1 seconds to 2.3 seconds. Across its benchmarks the post reports a speedup of 2.7x to 5.7x over the usual gather-and-broadcast method.
Reshard applies to the feed-forward projection weights, which the post says make up around 95% of the parameters in the latest large language models. Smaller weights still use gather-and-broadcast.
The post says NCCL M2N Reshard, an open source reference implementation, is available through the NVIDIA/nccl-extensions GitHub repository, and that NeMo RL can use it. The post says resilience support is planned for a later update.
Underlined sentences link to the facts we checked. Dotted words open a short explainer.
Fact 6 of 7 · from NVIDIA, 6 Nov 2026
NCCL M2N Reshard, an open source reference implementation of the Reshard collective, is available through the NVIDIA/nccl-extensions GitHub repo.
Your reaction
We count reactions per story and day, never who reacted. The counts help us choose what goes in the monthly issue. If you are signed in, your own page shows yours too.
We count reactions per story and day, never who reacted. They help choose what goes in the monthly issue.
Check our sources
Every sentence above is checked against this source.
1 Accelerating Reinforcement Learning Weight Refit with the Reshard Collective
Open the source archived copy-
Nvidia's developer blog post of 6 November 2026 reports NeMo RL weight-refit latency for DeepSeek v3 671B on NVIDIA DGX H100 of 14.9 s with gather-and-broadcast and 2.6 s with Reshard (PP16 x EP16 to TP32 x DP8). the company says
citeNVIDIA DGX H100 DeepSeek v3 671B PP16 x EP16 → TP32 x DP8 14.9 s 2.6 s
-
The post reports DeepSeek v3 671B refit latency on NVIDIA DGX GB200 of 6.1 s with gather-and-broadcast and 2.3 s with Reshard. the company says
citeNVIDIA DGX GB200 DeepSeek v3 671B PP16 x EP16 → TP32 x DP8 6.1 s 2.3 s
-
The post says that across various benchmarks the reshard refit achieves a speedup of 2.7x to 5.7x over conventional gather-and-broadcast refit. the company says
citeReshard refit achieved a 2.7x to 5.7x speedup over conventional gather-and-broadcast refit
-
The post says Reshard is a new type of collective communication primarily designed to accelerate RL weight refit, which transfers a tensor distributed across M source GPUs into the layout required by N destination GPUs; in RL refit the training GPUs are the source group and the inference GPUs the destination group. the company says
citea tensor distributed across M source GPUs into the layout required by N destination GPUs. In RL refit, the training GPUs form the source group and the inference GPUs form the destination group.
-
The post says feed-forward projection weights account for around 95% of the model's parameters in the latest LLMs, and that NeMo RL's hybrid approach sends them through Reshard while other small weights (for example QKV projection or mamba weights) use conventional gather-and-broadcast. the company says
citethese weights account for around 95% of the model's parameters.
-
The post says NCCL M2N Reshard, an open source reference implementation of the Reshard collective, is available through the NVIDIA/nccl-extensions GitHub repo, and that NeMo RL is an open source framework that can use it. the company says
citeNCCL M2N Reshard, an open source reference implementation of the Reshard collective, is available through the NVIDIA/nccl-extensions GitHub repo.
-
The post says NeMo RL currently uses a communication-group approach for resiliency, and that resiliency support using NCCL shrink and grow APIs is planned for future updates. the company says
citeNeMo RL currently uses the first approach for resiliency support. Resiliency support using NCCL shrink and grow APIs is pla
Every sentence above is checked against this source. Each fact has its own address you can cite.
Source 1 · NVIDIA · 6 Nov 2026
Accelerating Reinforcement Learning Weight Refit with the Reshard Collective
Nvidia's developer blog post of 6 November 2026 reports NeMo RL weight-refit latency for DeepSeek v3 671B on NVIDIA DGX H100 of 14.9 s with gather-and-broadcast and 2.6 s with Reshard (PP16 x EP16 to TP32 x DP8). the company says
NVIDIA DGX H100 DeepSeek v3 671B PP16 x EP16 → TP32 x DP8 14.9 s 2.6 s
The post reports DeepSeek v3 671B refit latency on NVIDIA DGX GB200 of 6.1 s with gather-and-broadcast and 2.3 s with Reshard. the company says
NVIDIA DGX GB200 DeepSeek v3 671B PP16 x EP16 → TP32 x DP8 6.1 s 2.3 s
The post says that across various benchmarks the reshard refit achieves a speedup of 2.7x to 5.7x over conventional gather-and-broadcast refit. the company says
Reshard refit achieved a 2.7x to 5.7x speedup over conventional gather-and-broadcast refit
Show all 7 factsShow fewer facts
The post says Reshard is a new type of collective communication primarily designed to accelerate RL weight refit, which transfers a tensor distributed across M source GPUs into the layout required by N destination GPUs; in RL refit the training GPUs are the source group and the inference GPUs the destination group. the company says
a tensor distributed across M source GPUs into the layout required by N destination GPUs. In RL refit, the training GPUs form the source group and the inference GPUs form the destination group.
The post says feed-forward projection weights account for around 95% of the model's parameters in the latest LLMs, and that NeMo RL's hybrid approach sends them through Reshard while other small weights (for example QKV projection or mamba weights) use conventional gather-and-broadcast. the company says
these weights account for around 95% of the model's parameters.
The post says NCCL M2N Reshard, an open source reference implementation of the Reshard collective, is available through the NVIDIA/nccl-extensions GitHub repo, and that NeMo RL is an open source framework that can use it. the company says
NCCL M2N Reshard, an open source reference implementation of the Reshard collective, is available through the NVIDIA/nccl-extensions GitHub repo.
The post says NeMo RL currently uses a communication-group approach for resiliency, and that resiliency support using NCCL shrink and grow APIs is planned for future updates. the company says
NeMo RL currently uses the first approach for resiliency support. Resiliency support using NCCL shrink and grow APIs is pla
Topics
The morning email
On the mornings we publish: the three top stories and up to four short ones. Free.
We email you a link to confirm. An issue may include one sponsor, always labelled Sponsored · Advertisement. Our emails count opens and clicks, not who made them. Unsubscribe in one click. What we keep