fosstodon.org is one of the many independent Mastodon servers you can use to participate in the fediverse.
Fosstodon is an invite only Mastodon instance that is open to those who are interested in technology; particularly free & open source software. If you wish to join, contact us for an invite.

Administered by:

Server stats:

9.8K
active users

#mi300

0 posts0 participants0 posts today

even with constraints amd will have steppings and respins of the silicon plus driver improvements - memory speeds should continue to go up, zen6 will probably jump to ddr6 and pci-e v6 to continue to boost i/o and make compute faster - the real gains are for enterprise data chips where inference is going to go through the roof #mi300 #mi350 #mi400

#AMD bets demand for its #MI300 accelerator will balance dips across other product lines
CEO #LisaSu boasted AMD now predicts #datacenter GPUs – this includes the #MI300A – to see consistent sequential growth over the next several quarters, driving revenues of $3.5 billion in 2024 alone. Tthe MI300 represents AMD’s first compelling contender in the #AI arena, with its previous-gen #MI200 parts primarily designed for #highperformancecomputing (#hpc) deployments
theregister.com/2024/01/31/amd #GPU #APU

The Register · AMD bets demand for its MI300 accelerator will balance dips across other product linesBy Tobias Mann

#AMD #MI300 – Taming Hype – #AI Performance, Volume Ramp, Customers, Cost, IO, Networking, Software
Amazing engineering, but what of the path to market?
AMD delivered #HPC #GPU silicon in 2021 for world's first #ExaFLop #supercomputer, Frontier. Now, everyone is looking forward MI300 which will be delivered later this year to El Capitan, 2nd Exascale supercomputer win. There hasn’t been this much buzz for a #datacenter chip since #Nvidia’s #Volta GPU and AMD’s Rome CPU.
semianalysis.com/p/amd-mi300-t

Cross post from birdsite: Confirmed by Nicholas Malaya of #AMD, #MI300, the new/upcoming datacentre #APU will present itself as a *single* GPU device in ROCm!

Unlike MI250 which used 2 compute dies over a high speed interface, presenting as 2 devices, MI300 will display all of its #CDNA chiplets as a single device, similar to how the newest #RDNA3 consumer cards are based around a multi die approach (even if those are a single graphics die and multiple memory die)

twitter.com/nicholasmalaya/sta

TwitterNicholas Malaya on Twitter“@never_released @FelixCLC_ @bcorni7 @filippospiga It will appear as one GPU!”