FCLChere we go!
AI counter within 2 sentences? 2
LMAO, freudian slip:
AI Error instead of Era
Ok so software devs are important to colab with: can haz software roadmap pweaze? š»
Mi300X is "launching"
TF32 FP8, sparsity support on top of fp/bf16 and ofc int8
Dozen chiplets (including HBM stacks?)
4 io dies with vcache built in (!!!!)
8 GPU tiles
17TB/s mesh, 5.3 TB/s memory BW
ehhh properly calling the functions "kernels", big <3 to Lisa Sue and/or the script writers
Some AI inference benchmarks, kind of cool but not super shocking.
comparison with H100 (not surprising) but back of the envelop puts Mi300 as within a few dozen percentage points of H200 when both in an SPR host machine with 8x GPGPU
building the foundation of AI for "x" years? sure.
We've been waiting/working on them since the fifth gen language discussions of the 80 and 90's
DAT DENSITY
OCP compliant is pretty pog NGL
NGL a little surprised to see oracle at an AI thing with AMD of all folks
"Open, proven and ready" is one way to put ROCm š¬
Love you guys and galls, but I hard disagree with the claim
You can't claim it's FOSS and that you accept work from FOSS when you won't accept patches from anyone :/
ROCm 6: adds support for FP8 (which ones?)
Graph compiler is cool but not that new.
still good to see as long as it gets *delivered*
Software partners is good, running natively* with pytorch and requiring no code is also good.
native support for Mi300 in trident from OpenAI
Meh, how many times has "software" 2.0 been brought up?
first with high level langs, then run time langs etc.
will "Ai" change some parts of software developement, probably?
"reshape" the enterprise, maybe? but every "new" thing has changed some parts of the enterprise, question is always how much
Back to the real presentation
ROCm 6 is "SoonTM" (next year? next quarter? etc.)
Ehhhhhhh dell getting Mi300! (knew it was coming, still good to see)
8xMi300X
LMAO, this guy from SM is a hoot xD
holy shit, 100 KW in a single Rack?!?!?!?! š³
Can the term "Ai PC" die already? It's beyond cringe
ok so at this point looks like Mi300X maxes out at 8x GPUs
New plan for smart doorbells:
use different voice clips of VPs, principal engineers, CEOs etc. for saying "good morning"
opening up of infinity fabric is super cool.
I assume this is GPGPU infinity fabric, not chiplet to chiplet IF?
( @philparkbot ?)
@philparkbot ok, starting to think this is just Ultra Ethernet
Yeah, this is just ultra eth from the looks of it
#HPC MENTIONED LETS GOOOOOOOOOOOOOOOOO
Phil Park@fclc this is the first time Iām hearing of this. Unlikely that people will get to see inside the design and coherence protocols but maybe some port connection into the IF (UCIe? xGMI?)
@philparkbot Yeah, seems like a branding of UltraEthernet when coupled with AMD devices specifically
Domenico Lamberti@fclc looksl ike all the liquid cooling companies at SC23 were on to something
@Mobile_dom I just don't know who has a DC that can support that level of density outside of NYC, London or maybe SF?
*maybe* HongKong and Signapore?
@fclc theres that new google server thats using a novel geothermal idea in nevada
@Mobile_dom Yeah, but then why do you need that much density in a single rack? costs go exponential per rack once you go above ~40KW
@fclc hey man i just wrote googles checks i dont need to worry where the money comes from :P
Brandon Biggs@fclc HPEs Cray Ex 400 "cabinets" are basically two racks and 400KW. Bonkers.
Tom Hayward
JP Lehr