[ad_1]
//php echo do_shortcode(‘[responsivevoice_button voice=”US English Male” buttontext=”Listen to Post”]’) ?>
Cerebras has unveiled a 3rd technology of its wafer-scale chip, providing 125 PFLOPS (at FP16 precision) from a single gadget. Given a single day, a four-chip set up might fine-tune Llama2-70B, whereas the most important installations of two,048 chips would be capable of practice it from scratch in the identical time.
The wafer-scale engine 3 (WSE3) doubles the big language mannequin (LLM) coaching pace of the WSE2, in the identical 15kW energy envelope and on the identical price level, Cerebras CEO Andrew Feldman instructed EE Instances.
Feldman stated that deployment of strategic accomplice G42’s WSE2-based AI supercomputer household, Condor Galaxy, goes properly. Condor Galaxy 2 was stood up on schedule. The third supercomputer within the household, Condor Galaxy 3, will use new third-gen WSE3 {hardware}. G42 has additionally chosen to put in a “vital” quantity of Qualcomm Cloud AI100 inference-only {hardware} as a part of Condor Galaxy3. For this, Cerebras partnered with Qualcomm to regulate the coaching course of in order that ensuing fashions are optimized for inference on Qualcomm Cloud AI100 chips. This work has resulted in a efficiency enchancment of 10× versus non-optimized fashions, Feldman stated.
Wafer-Scale 3
The WSE3 is identical bodily measurement as earlier generations, however packs much more into the house. The brand new chip has moved to TSMC 5 nm (from TSMC 7 nm) and options 900,000 cores in comparison with 850,000 within the earlier technology. The cores are additionally larger. General, the distinction is 4 trillion transistors versus 26 trillion within the earlier gen.
By Conquer Electronics 03.12.2024
By Winbond 03.11.2024
By TME Staff 03.08.2024
“We made a sequence of architectural enhancements born of all of the learnings we’ve had over the previous 5 years of deploying techniques, and went to a barely bigger core,” Feldman stated.
Cerebras’ third-gen wafer-scale engine. (Supply: Cerebras)
The WSE additionally options 42 GB of SRAM with 21 PBytes/s reminiscence bandwidth. SRAM may be prolonged with giant exterior DRAM subsystems equipped by Cerebras, to allow coaching of AI fashions as much as 24 trillion parameters (about 10× the dimensions of immediately’s fashions like GPT-4 and Gemini). Even the most important fashions may be saved in a single logical reminiscence house with out partitioning or refactoring.
The WSE3 is available in a system referred to as the CS3. The utmost cluster measurement has been elevated to 2,048 CS3s (as much as 256 ExaFLOPS of FP16 compute).
“With [2,048 CS3s] you would take Meta’s largest GPU cluster and do the work that’s taking it a month in a day,” he stated. “You possibly can deliver to enterprise customers the dimensions of compute that hyperscalers hold for themselves.”
Requested when a 2,048-CS3 cluster may turn out to be a actuality, Feldman stated decisively: “this yr.”
CS3s put in at Colovore in Santa Clara, California. (Supply: Cerebras)
Software program stack
Cerebras’ software program stack helps Python 2.0 and helps all mannequin varieties, together with the most important multi-modal LLMs, ViT, combination of consultants and diffusion. There’s additionally assist for unstructured sparsity and dynamic sparsity (for zeros that emerge in the course of the coaching course of).
Coaching a 175B model of Megatron utilizing hundreds of GPUs would require greater than 20,000 traces of Python, C++, CUDA and HTML code, he stated. Cerebras requires solely 565 traces of Python—which might be carried out by a single engineer in a day, Feldman added.
Condor Galaxy 3
The third cluster within the Condor Galaxy household of AI supercomputers, to be stood up in Dallas, Texas, can have 64 CS3 techniques for a complete of 8 ExaFLOPS of FP16 compute.
CG3 will be a part of CG1 in Santa Clara, California, and CG2 in Stockton, California, which every characteristic 64 previous-gen CS2s. Condor Galaxy 1 has skilled fashions together with Jais-30B, essentially the most outstanding Arabic-language LLM, and Med42, G42’s scientific LLM.
The full AI compute supplied by CG 1, 2 and three will likely be 16 ExaFLOPS at FP16, however Feldman stated the household of supercomputers will attain 55 ExaFLOPS in complete by the tip of 2024, with the commissioning of Condor Galaxies 4 via 9.
Inside Condor Galaxy 1. (Supply: Cerebras)
Inference {hardware}
G42 can be including inference-only {hardware} to Condor Galaxy 3 for the primary time, at a scale supposed to offer inference the big fashions CG3 will likely be coaching. G42 has chosen Qualcomm CloudAI 100 units for inference workloads.
“Right this moment, Cerebras doesn’t concentrate on inference, apart from some very laborious inference issues in nationwide safety and protection,” Feldman stated. “We don’t have an inference providing, we’re coaching solely.”
Cerebras has been collaborating with Qualcomm to allow shifting from coaching on Cerebras CS3s to inference on Qualcomm Cloud AI 100s with a single click on. The businesses have additionally been collectively engaged on inference-aware coaching, leading to a ten× speedup versus an unoptimized variations of LLMs operating on the identical Qualcomm chips.
Andrew Feldman (Supply: Cerebras)
“Till just lately, nearly all of the compute was in coaching, however as we transfer from an setting the place AI is a interest to AI in manufacturing, coaching doesn’t lower, however inference will increase,” Feldman stated. “This isn’t only a drawback within the summary for the hyperscalers, this can be a drawback for among the largest and most forward-looking firms as they consider shifting into manufacturing.”
This work used 4 most important strategies to tailor Cerebras-trained fashions for inference on Qualcomm CloudAI 100 units:
The full speedup from the combos is as a lot as 10×.
“These are improvements that may actually transfer the market,” Feldman stated, including that this work with Qualcomm “offers the attain, capability and engineering horsepower to remodel your complete market.”
Cerebras stated it has a “sizeable” backlog of orders for CS3s, which span enterprise, authorities and worldwide clouds.
[ad_2]
Supply hyperlink
