Meta announces its next-generation Meta Training and Inference Accelerator chips for AI training, and says MTIA v1 and the new chips are both now in production
Our very own Nicolaas Viljoen is featured in this blog post. … Tanmay Zargar : Unveiling what we've been working on for the past few years - the next generation of MTIA Inference platform. We're just getting started. … See also Mediagazer
The Verge Emilia David
Related Coverage
- Our next-generation Meta Training and Inference Accelerator Meta
- Introducing Our Next Generation Infrastructure for AI Meta
- View article CCN.com
- View article Bloomberg
- Meta challenges Nvidia's dominance with new AI chips Interesting Engineering
- Meta unveils more advanced AI chip for faster model training NewsBytes
- Zuckerberg's Meta Launches MTIA Chip To Rival Nvidia's AI Offerings In Cloud Business: Here's What You Need To Know Benzinga
- Meta unveils second-gen AI training and inference chip ZDNet
- Meta unveils the next step in its AI hardware evolution: MTIA v2 MediaNama
- [News] Meta Reportedly Unveils Next-Generation In-House AI Chip, Using TSMC 5nm Process TrendForce Insights
- Meta Announces New MTIA AI Accelerator with Improved Performance to Ease NVIDIA's Grip TechPowerUp
- Meta Unveils Details of its Latest AI Chip, MTIA Analytics India Magazine
- Meta unveils latest AI chip Mobile World Live
- Meta's New Chip Underlines Questions About AI Spending The Information
- Meta unveils its newest custom AI chip as it races to catch up TechCrunch
- Meta's upgraded MTIA AI chips offer 3.5x performance boost DatacenterDynamics
- Meta debuts new generation of AI chip Reuters
- Meta launches new custom AI chip News.ng
- Meta vs. Nvidia - No competition Radio Free Mobile
- Meta Debuts New AI Chip, Aiming to Decrease Reliance on Nvidia Data Center Knowledge
- Meta's new AI chips offer faster model training speed Android Headlines
- Meta's New AI Chips Will Support GenAI Models Cryptopolitan
- We just announced our next-generation training and inference accelerator, the AI-MTIA (Meta Training and Inference Accelerator). … Mikael Strandlund
- Everybody is dipping into chips. For all the buzz around the AI race, the chip race is the one that deserves more attention. … Albert Fong
- Announcing the next generation of our in-house silicon designed for Meta's unique workloads, products and services such as ranking and recommendation models. … Ash Jhaveri
- MTIA v2 is finally here! It's been an incredible multi-year partnership across many orgs to land this 0 to 1 product at Meta scale. … Neeraj Agrawal
- Meta has launched our first AI Accelerator into production! — Our very own Nicolaas Viljoen is featured in this blog post. … Dan Rabinovitsj
- Unveiling what we've been working on for the past few years - the next generation of MTIA Inference platform. We're just getting started. … Tanmay Zargar
Discussion
-
@paul_rietschka
Paul Rietschka
on threads
I think they have the money to throw at things, and the entirety of their ML/AI efforts are pure Zuckerberg vanity projects. This isn't a company with a cloud platform like Google or Microsoft, but it **is** a company that wants to pretend it's in the same league and not a diffe…
-
@paul_rietschka
Paul Rietschka
on threads
Stockholders should be asking why a company like Meta needs to build its own silicon. Because there's no business case. And all this on top of the fact the past tense of “to lead” is “led.”
-
@benbajarin
Ben Bajarin
on x
Assuming the same arch YoY this is a RISC-V based accelerator. “designed for Meta's AI workloads”
-
@iancutress
@iancutress
on x
@SquashBionic @Meta 1.35 MHz up from 800 MHz. 800 MHz is going to be a lot more efficient, and you're moving out of the efficiency window quite a lot, even with the arch/node change
-
@rao_hacker_one
Arun Rao
on x
Meta is not in the chip-selling or chip-renting business, but we are getting more vertically integrated to better serve our app users & businesses, and to make the best open source AI models and tools freely available to the world. Kudos to the MTIA team!
-
@mikeyanderson
Mikey Anderson
on x
It's good that more chips are being created. A diverse ai ecosystem is a healthy one.
-
@aiatmeta
@aiatmeta
on x
This new MTIA chip can deliver 3.5x the dense compute performance & 7x the sparse compute performance of MTIA v1. Its architecture is fundamentally focused on providing the right balance of compute, memory bandwidth & memory capacity for serving ranking & recommendation models. […
-
@iancutress
@iancutress
on x
Looks like @Meta is talking about next-gen MTIA already. ➡️ 90W, TSMC N5 ➡️ 256 MB SRAM, 2.7 TB/sec ➡️ 128GB LPDDR5, 204.8 GB/sec ➡️ 2.35B transistors, 1.35 GHz (up from 800 MHz) ➡️ 354 TF INT8 GEMM ➡️ 2 chips/board, 12 boards/system ➡️ 3x perf vs Gen1 https://ai.meta.com/... [im…
-
@isidentical
Batuhan Taskaya
on x
nice. still so much to go but this is pretty good (and day zero full on torch support is so good)
-
@iamadifuchs
Adi Fuchs
on x
Interesting. MTIAv2 has a small die and 90W TDP (typical training accelerators are ~350-500W / 700-1000W for MCM) and about 1/3 of H100's TFLOPs, so could be an overall win, maybe something like Google's approach of scaling out many small TPUs. Nonetheless, great times for chips!
-
@squashbionic
@squashbionic
on x
Interesting chip, but there's a 3.6x power consumption increase for ~3x performance improvement(25w->90w) ? Am I seeing that correct?
-
@thetechbrother
@thetechbrother
on x
pictured: @ylecun conducting cutting-edge AI research [image]
-
@__tinygrad__
@__tinygrad__
on x
If you want to build your own training accelerator, you must have your own NN framework with adoption. Meta happens to have one, so this chip might work.
-
@aiatmeta
@aiatmeta
on x
Introducing the next generation of the Meta Training and Inference Accelerator (MTIA), the next in our family of custom-made silicon, designed for Meta's AI workloads. Full details ➡️ https://ai.meta.com/... [image]
-
@soumithchintala
Soumith Chintala
on x
Meta announces 2nd-gen inference chip MTIAv2. * 708TF/s Int8 / 353TF/s BF16 * 256MB SRAM, 128GB memory * 90W TDP. 24 chips per node, 3 nodes per rack. * standard PyTorch stack (Dynamo, Inductor, Triton) for flexibility Fabbed on TSMC's 5nm process, its fully programmable via the.…