Home
Creators
Search Recent Random
Posts
Search DMs Popular Hash Lookup Tags Random
Importer
Import FAQ
Register Favorites Keys Review DMs Logout
Community
Telegram ThePornDude
Home Creators Posts Import Register Favorites Logout

Discussing AMD RDNA 4, Nvidia Blackwell, GB10, CDNA w/ KarbinCry (Hot Chips Telegrams) (Patreon)

Published:
2025-08-27 23:49:44
Edited:
2025-08-27 23:51:01
Imported:
2026-06
Tags:

Downloads

Content

That's right - the yearly Hot Chips episode is here! That means Tom & KarbinCry will be discussing everything unveiled at said conference, and answering your corresponding questions regarding the subjects involved! Obviously we expect a lot of AMD and Nvidia stuff...but feel free to ask anything that was part of Hot Chips!

Presentations in PDF form are attached so you can read up, and you have ~24 hours to submit below!

Comments

CompressedAIBlocks

Karbin thanks for the work you do here. What was your favourite part/product of hot chips this year?

Dark Side of the Force

Why does GB10 take so long? What is the problem with it currently? (if you are aware of it). Would there be much of a market for GB10 if it slips to next year? Also not related to Hotchips, do you have any idea why Google went with PowerVR gpu for their G5 Tensor? Are they possibly aware that PowerVR Gpu don't have good compatibility with android games?

CompressedAIBlocks

Karbin, this is above my head but can you explain why an enterprise would choose a system like IBM POWER line up over AMD or even Intel? We usually talk *way more* about those two but I always seem to forget IBM POWER exists. Do they do anything better/more efficient than x86 or ARM does in this space?

CompressedAIBlocks

RDNA 4 looks to be incredibly modular from what they can disable. Do you guys expect this philosophy to continue even when we transition into UDNA? Seems like what AMD does best is hitting the exact segment they want with a variety of solutions.

Timo H

Many years ago Lisa Su said to improve compute performance in electric usage (green goals). Is that visible in those slides? How can that be achieved with AI demanding more and more sram and ram, as those dont scale as well...

Cleansweep

This year's ECAM presentation is understandably focused on AI server applications to attract investment/funding. FabricLabs showed an impressive ability to design cold plates to eliminate hot spots, and their ability to print direct to silicon is going to be a huge game changer in the future. However, in the 2023 presentation, they showed off an impressive ability to generate all sorts of complex geometries. Is there value in using ECAM to make an IHS "core" that gets slotted into a more traditional fin stack to cool something like Magnus or the various PlayStation APUs if they're probably going to have a 10 year lifespan?

Chris Rijk

Many thanks for the Hot Chips coverage. Can you explain what’s going on in AMD’s presentation on RDNA 4 on page 18 where it talks about on-chip compression and there’s points about reducing fabric bandwidth by 25% in some workloads. Does it reduce VRAM bandwidth or memory usage as well?

Chris Rijk

As usual there’s a lot of focus on AI but this year it seemed to be more on the problems of extreme cooling, rack design and networking. What AI related presentations did you find the most interesting? Was there anything that could reasonably be expected to directly benefit consumer designs in the coming years?

AnotherChris

Not exactly a direct hardware question but will we see announcements about Redstone or ROCM updates at the event this year or other software optimizations to the hardware?

David

Is there anything new in the Blackwell neural rendering presentation that highlight gaps for RDNA4 and possibly RDNA5 for future game engines? What about local AI inference use-cases? (e.g. Universal MIG, Workload streaming, Improved SER, etc )

David

To expand on the below question "why an enterprise would choose a system like IBM POWER line up over AMD or even Intel"... or even the IBM z-series?

Prior is the theme 999

Can you clarify/elaborate the RDNA4x presentation under SOC. They seem to be using language to me at least that say they can essentially combine SOC’s especially under page 19. Is that just not chiplets?

AllenC

Anything on ROCm performance for HPC? How about release cadences and how tightly they'll tie that to hardware releases in the future

danm1988

Anything about this Intel IPU would be great (I actually have a 1st gen one I have not worked much with yet, C6000X-PL). Also, anything regarding the new Intel Xeon Clearwater release and if it will have any relevance or relation to what we may see for intel's premium desktop processors and GPUs.