AI Networking From the interconnect of 10,000 GPUs in a Single Model to

📎 Click on me to view all my attachmentsTheme:

ListAttBBC | v6.12 | 2026-09-10 | Andrew.human & Claude.ai

#	File	Size	Downloads	Info	DL
00	instagram-1792241633384.mp4	5.5 MB	9	ℹ️	⬇️
Attachment ID 7540 Filename (display) instagram-1792241633384.mp4 File extension mp4 MIME type video/mp4 Size 5.5 MB (5731599 bytes) Downloads 9 Dimensions n/a Uploaded by Guest Message ID 49807 Attachment type Normal attachment (0) Storage folder ID 2 Thumbnail attach ID none File hash 192ca2bd998a7384ee9a5c4ad2195a35d73e5df9 Approved Yes Filename (on disk) 7540_192ca2bd998a7384ee9a5c4ad2195a35d73e5df9 Storage directory 2 (/var/www/html/attachments2) File exists on disk? Yes (core) — Media metadata (none) — Media metadata Unavailable - Neither getID3 nor ffprobe is available on this install
Chip — Today at 01:56:33 AM
01	gpu_network_transcript.md	2 KB	0	ℹ️	⬇️
Attachment ID 7547 Filename (display) gpu_network_transcript.md File extension md MIME type Size 2 KB (2013 bytes) Downloads 0 Dimensions n/a Uploaded by Guest Message ID 49807 Attachment type Normal attachment (0) Storage folder ID 2 Thumbnail attach ID none File hash fc77cedf23fddddfd30d3447f14f662972a8da75 Approved Yes Filename (on disk) 7547_fc77cedf23fddddfd30d3447f14f662972a8da75 Storage directory 2 (/var/www/html/attachments2) File exists on disk? Yes (core)
No recorded downloads yet.
02	gpu_network_explained.md	3.3 KB	0	ℹ️	⬇️
Attachment ID 7548 Filename (display) gpu_network_explained.md File extension md MIME type Size 3.3 KB (3380 bytes) Downloads 0 Dimensions n/a Uploaded by Guest Message ID 49807 Attachment type Normal attachment (0) Storage folder ID 2 Thumbnail attach ID none File hash d748927cef720667d92c74f5619a1d9902a1e645 Approved Yes Filename (on disk) 7548_d748927cef720667d92c74f5619a1d9902a1e645 Storage directory 2 (/var/www/html/attachments2) File exists on disk? Yes (core)
No recorded downloads yet.
03	gpu_network_explained_2.md	3.7 KB	0	ℹ️	⬇️
Attachment ID 7549 Filename (display) gpu_network_explained_2.md File extension md MIME type Size 3.7 KB (3760 bytes) Downloads 0 Dimensions n/a Uploaded by Guest Message ID 49807 Attachment type Normal attachment (0) Storage folder ID 2 Thumbnail attach ID none File hash 44b8037112fda7858a9c7ef0589c54d62c100763 Approved Yes Filename (on disk) 7549_44b8037112fda7858a9c7ef0589c54d62c100763 Storage directory 2 (/var/www/html/attachments2) File exists on disk? Yes (core)
No recorded downloads yet.
04	gpu_network_map.md	4.8 KB	0	ℹ️	⬇️
Attachment ID 7550 Filename (display) gpu_network_map.md File extension md MIME type Size 4.8 KB (4913 bytes) Downloads 0 Dimensions n/a Uploaded by Guest Message ID 49807 Attachment type Normal attachment (0) Storage folder ID 2 Thumbnail attach ID none File hash 6e1cad336ea4ae58cada1592136e788e139f3176 Approved Yes Filename (on disk) 7550_6e1cad336ea4ae58cada1592136e788e139f3176 Storage directory 2 (/var/www/html/attachments2) File exists on disk? Yes (core)
No recorded downloads yet.
05	gpu_network_protocols.md	10.3 KB	0	ℹ️	⬇️
Attachment ID 7551 Filename (display) gpu_network_protocols.md File extension md MIME type Size 10.3 KB (10592 bytes) Downloads 0 Dimensions n/a Uploaded by Guest Message ID 49807 Attachment type Normal attachment (0) Storage folder ID 2 Thumbnail attach ID none File hash fac4213d76708520f94931cb8ac6d0a81f33be53 Approved Yes Filename (on disk) 7551_fac4213d76708520f94931cb8ac6d0a81f33be53 Storage directory 2 (/var/www/html/attachments2) File exists on disk? Yes (core)
No recorded downloads yet.
06	gpu_datacentre_units.md	7.5 KB	0	ℹ️	⬇️
Attachment ID 7552 Filename (display) gpu_datacentre_units.md File extension md MIME type Size 7.5 KB (7666 bytes) Downloads 0 Dimensions n/a Uploaded by Guest Message ID 49807 Attachment type Normal attachment (0) Storage folder ID 2 Thumbnail attach ID none File hash 453d3c6e133c72e311f123f14ab83f437e513796 Approved Yes Filename (on disk) 7552_453d3c6e133c72e311f123f14ab83f437e513796 Storage directory 2 (/var/www/html/attachments2) File exists on disk? Yes (core)
No recorded downloads yet.
07	gpu_model_copies_and_training.md	8.3 KB	0	ℹ️	⬇️
Attachment ID 7553 Filename (display) gpu_model_copies_and_training.md File extension md MIME type Size 8.3 KB (8526 bytes) Downloads 0 Dimensions n/a Uploaded by Guest Message ID 49807 Attachment type Normal attachment (0) Storage folder ID 2 Thumbnail attach ID none File hash 1681b27215e6c90fcb0d02eb024199494d308d2b Approved Yes Filename (on disk) 7553_1681b27215e6c90fcb0d02eb024199494d308d2b Storage directory 2 (/var/www/html/attachments2) File exists on disk? Yes (core)
No recorded downloads yet.
08	From AI Networking to the Datacenter.md	49.1 KB	0	ℹ️	⬇️
Attachment ID 7557 Filename (display) From AI Networking to the Datacenter.md File extension md MIME type Size 49.1 KB (50292 bytes) Downloads 0 Dimensions n/a Uploaded by Guest Message ID 49807 Attachment type Normal attachment (0) Storage folder ID 2 Thumbnail attach ID none File hash 154a9b425b36c5e04d8887bbd032e2866e6ca4c7 Approved Yes Filename (on disk) 7557_154a9b425b36c5e04d8887bbd032e2866e6ca4c7 Storage directory 2 (/var/www/html/attachments2) File exists on disk? Yes (core)
No recorded downloads yet.

We transcribe it, we make it coherent, we explore the various networks, their protocols what they connect to, during Training and Inference

dopetalk

News:

dopetalk does not endorse any advertised product nor does it accept any liability for it's use or misuse

Author Topic: AI Networking From the interconnect of 10,000 GPUs in a Single Model to (Read 12 times)

Chip (OP)

What Connects Ten Thousand GPUs Into One AI Model? (Original Transcript)

Why Thousands of AI Chips Need a Very Good Network

The setup: one chip, then thousands

The constant conversation

Where it goes wrong

Why this traffic is unusually hard

Fix 1: take the processor out of the path

Fix 2: a smarter way of sending

The takeaway

Why Thousands of AI Chips Need a Very Good Network

What this is about

The setup: one chip, then thousands

The constant conversation

Where it goes wrong

Why this traffic is unusually hard

Fix 1: take the processor out of the path

Fix 2: a smarter way of sending

The takeaway

The Different "Networks" in AI Training: A Map

The short version

Level by level

1. The neural network: the model

2. The feed-forward network: a part inside the model

3. Links inside one server or rack: the short, fast hop

4. The cluster network: the long hop

5. Transport methods: how the data is sent

6. The internet: not involved

How the pieces fit together

The Named Protocols That Connect GPUs, From the Model Down to the Wire

1. The whole stack at a glance

2. How the model is split (and why that decides the traffic)

3. The group-communication routines

4. Inside the box: connections within one server or rack

5. Between boxes: the cluster network

6. The transcript's mystery name is probably MRC

7. The bridge: GPUDirect RDMA

8. Training versus inference

9. What each network does differently, in one table

Sources

The Functional Units of an AI Datacentre: How Many, Of What, and How They Connect

1. The nesting, from chip to site

2. The main hardware models, and what each one connects

NVIDIA GB200 NVL72 (current generation)

NVIDIA Vera Rubin (shipping from the second half of 2026)

AMD Helios (reference design, volume deliveries expected in the second half of 2026)

Google TPU "Ironwood" (a different design, built around its own chip)

3. How big a site gets

4. The other functional units

5. How the interconnects compare

6. Caveats

Sources

How Many Copies of a Model Run, How They Stay in Sync, and Why You Can't Just Keep Training It

1. One model, many running copies

2. Training: many copies that behave as one

There is no master copy

Why there are no conflicts

What does go wrong

The older alternative

3. Inference: many independent copies serving users

Two special cases

4. Why you can't just keep training the live model

There is no "training mode" switch

What goes wrong in the training itself

What goes wrong around it

The cheaper route for new information

Glossary

Conversation transcript: AI datacentre networking, multi-LLM ideas and the NAT memory model

Operator

Claude

Operator

Claude

1. Transcript, verbatim, with paragraphs

2. Plain-English summary

Operator