Topics
Latest
AI
Amazon
Image Credits:TechCrunch
Apps
Biotech & Health
Climate
Image Credits:TechCrunch
Cloud Computing
DoC
Crypto
Enterprise
EVs
Fintech
fundraise
gizmo
Gaming
Government & Policy
computer hardware
Layoffs
Media & Entertainment
Meta
Microsoft
privateness
Robotics
Security
societal
Space
Startups
TikTok
transfer
speculation
More from TechCrunch
Events
Startup Battlefield
StrictlyVC
Podcasts
Videos
Partner Content
TechCrunch Brand Studio
Crunchboard
Contact Us
Meta , hell - bent on catch up to rivals in the productive AI space , is spendingbillionson its own AI crusade . A portion of those billion is go towardrecruiting AI researcher . But an even large chunk is being spent develop hardware , specifically chips to run and train Meta ’s AI models .
Metaunveiledthe newest fruit of its chip dev feat today , conspicuously a day after Intelannouncedits latest AI accelerator pedal hardware . Called the “ next - gen ” Meta Training and Inference Accelerator ( MTIA ) , the successor tolast twelvemonth ’s MTIA v1 , the fleck run models include for ranking and recommending display ads on Meta ’s attribute ( e.g. Facebook ) .
Compared to MTIA v1 , which was built on a 7 nm process , the next - gen MTIA is 5 nm . ( In chip fabrication , “ process ” refers to the size of the small component that can be build on the cow dung . ) The next - gen MTIA is a physically big figure , pack with more processing cores than its herald . And while it have more power — 90W versus 25W — it also swash more interior computer storage ( 128 MB versus 64 MB ) and guide at a higher intermediate clock f number ( 1.35GHz up from 800MHz ) .
Meta tell the next - gen MTIA is presently live in 16 of its data shopping center region and extradite up to 3x overall good public presentation compared to MTIA v1 . If that “ 3x ” claim sounds a bit vague , you ’re not wrong — we mean so too . But Meta would only offer that the figure come from testing the carrying out of “ four key theoretical account ” across both chips .
“ Because we control the whole pile , we can accomplish greater efficiency compared to commercially available GPUs , ” Meta writes in a blog post shared with TechCrunch .
Meta ’s hardware showcase — which comes a mere 24 time of day after a press briefing onthe company ’s various ongoing generative AI initiatives — is unusual for several reasons .
One , Meta reveals in theblog postthat it ’s not using the next - gen MTIA for generative AI preparation work load at the moment , although the company claims it has “ several programs underway ” explore this . Two , Meta admits that the next - gen MTIA wo n’t replace GPUs for running or education models — but instead will complement them .
Join us at TechCrunch Sessions: AI
Exhibit at TechCrunch Sessions: AI
Reading between the assembly line , Meta is moving tardily — perhaps more slowly than it ’d like .
Meta ’s AI teams are almost sure as shooting under pressure to cut costs . The caller ’s set to pass anestimated$18 billion by the last of 2024 on GPUs for training and lead generative AI poser , and — with training cost for cutting - sharpness generative theoretical account swan in the tens of zillion of dollars — in - firm hardware exhibit an attractive option .
And while Meta ’s hardware drags , rival are pulling onward , much to the consternation of Meta ’s leading , I ’d suspect .
Google this workweek made its fifth - generation custom chip for training AI models , TPU v5p , generally available to Google Cloud customers , and revealed its first consecrate chip for running model , Axion . Amazon hasseveralcustomAI chipfamilies under its belt . And Microsoft last twelvemonth jump into the fray with theAzure Maia AI Accelerator and the Azure Cobalt 100 CPU .
In theblog post , Meta say it took few than nine months to “ go from first silicon to output manikin ” of the next - gen MTIA , which to be fair is shorter than the typical windowpane between Google TPUs . But Meta has a lot of bewitch up to do if it hopes to achieve a measure of independency from third - political party GPUs — and match its stiff contention .
