Zesty tool aims to stretch existing Nvidia GPUs for AI workloads
Zesty introduced Automated Fractional GPUs on October 15, a platform feature that uses Nvidia's Run:ai technology to automate sharing of existing GPU capacity. The company says this can help customers handle more AI work before needing to buy additional accelerators, though it did not provide contract or independently verified savings figures. TECHi noted Nvidia shares were down about 2% in a regular-session observation, while cautioning that the move cannot be tied to the launch.
Zesty's October 15 release, Automated Fractional GPUs, is delivered through its platform and relies on Nvidia's Run:ai software, which Nvidia made fully available two days earlier. The tool automates GPU sharing decisions, adjusting memory, compute and placement as needs shift. It works with CUDA-native fractioning, time-slicing, Nvidia MPS and Kubernetes schedulers including KAI, according to the company.
Zesty says applications and models need no code changes, but it disclosed no contract value, purchase commitment or independently verified savings. A simple utilization example shows how moving from 50% to 80% could allow 60% more work before extra hardware is needed. TECHi's regular-session quote at 11:26 a.m. EDT showed Nvidia down 2.02%, though the article says this cannot be linked to the launch.
Count first para: Zesty's(1) October(2) 15(3) release,(4) Automated(5) Fractional(6) GPUs,(7) is(8) delivered(9) through(10) its(11) platform(12) and(13) relies(14) on(15) Nvidia's(16) Run:ai(17) software,(18) which(19) Nvidia(20) made(21) fully(22) available(23) two(24) days(25) earlier.(26) The(27) tool(28) automates(29) GPU(30) sharing(31) decisions,(32) adjusting(33) memory,(34) compute(35) and(36) placement(37) as(38) needs(39) shift.(40) It(41) works(42) with(43) CUDA-native(44) fractioning,(45) time-slicing,(46) Nvidia(47) MPS(48) and(49) Kubernetes(50) schedulers(51) including(52) KAI,(53) according(54) to(55) the(56) company.(57) Good.
Second: Zesty(1) says(2) appli
... No extra. Need maybe no bullet. Good. Double-check "under 120 words total" for expanded. We have 48+60=108. Good. "2 short paragraphs" yes. "additional factual detail and background drawn ONLY from material above" We included all from material. "entirely original wording, never copying source phrasing." Some phrases: "CUDA-native fractioning", "time-slicing", "Nvidia MPS", "Kubernetes schedulers including KAI" are technical names. "independently verified savings" maybe source phrase? Summary says "independently verified savings figures"; article says "independently measured savings". We used "independently verified savings" maybe okay but could be considered source phrasing? It's a common phrase. "purchase commitment" source says "purchase commitment to Nvidia". We used "purchase commitment" okay. "contract value" source. Fine. Need maybe mention "Nvidia's former proprietary fractioning implementation