I-NVIDIA Ikhipha I-Cosmos 3 Edge: Imodeli Yomhlaba Evulekile ye-4B-Parameter Ebangela Izizathu Futhi Ikhiqize Izenzo Zerobhothi Kudivayisi

0 2 3 minutes read

I-NVIDIA Ikhipha I-Cosmos 3 Edge: Imodeli Yomhlaba Evulekile ye-4B-Parameter Ebangela Izizathu Futhi Ikhiqize Izenzo Zerobhothi Kudivayisi

“>

Imodeli yomhlaba yenzani lapha

Imodeli yomhlaba ifunda ukuthi indawo ishintsha kanjani ngokuhamba kwesikhathi. Imele izinto, ukunyakaza, ubudlelwano bendawo, kanye nemiphumela yezenzo.

Cabanga ngerobhothi elifinyelela into ethile. Ukuqaphela into kuyisinyathelo sokuqala kuphela. Irobhothi kufanele futhi lilandelele ukuthi into ikuphi, ukuthi into ebambayo ihamba kanjani, nokuthi kwenzekani lapho ithintwa. Imodeli yomhlaba yezizathu mayelana nalobu budlelwano. Ingakwazi ukubikezela umphumela obonakalayo wesenzo, icabangele isenzo esibangele ushintsho, noma ikhiqize isenzo ukuze kufinyelelwe umgomo.

I-Cosmos 3 Edge iletha lawa makhono kwimodeli eyodwa ekudivayisi. Ukumelwa kwayo okwabiwe kuvumela isistimu ukuthi iqonde isimo somhlaba samanje, ilingise ikusasa elingenzeka, futhi ixhumanise lelo kusasa nezenzo.

Imibhoshongo emibili ye-transformer, ukumelela okukodwa okwabiwe

I-Cosmos 3 isebenzisa i-Architecture ye-Mixture-of-Transformers enemibhoshongo emibili, echazwe embikweni wezobuchwepheshe we-NVIDIA.

I umbhoshongo we-autoregressive ucubungula umbono namathokheni ombhalo ukuze aqonde futhi acabange. I umbhoshongo wokusabalalisa icubungula umbono, umsindo, namathokheni esenzo sokubikezela, ukukhiqiza, nokulingisa kwemizwa.

Le mibhoshongo emibili igcina izendlalelo ezihlukene zokujwayela kanye nama-perceptron angama-multilayer. Babelana ngezendlalelo zokunaka ze-multimodal, eziqondanisa ulwazi ngolimi lonke, ividiyo, umsindo, nesenzo. Lokhu kuvumela imodeli ukuthi icabange ngesigcawu ngaphambi kokuthi ikhiqize okukhiphayo.

Iphethini yokunaka ivumelana nendlela ngayinye. Ulimi lusebenzisa ukunaka okuyimbangela, lapho ithokheni ngayinye inakekela amathokheni angaphambili. Amathokheni okusabalalisa ahambela kabanzi umongo otholakalayo, asekela ukubikezela okuhambisanayo nokukhiqiza. Kuye ngomsebenzi, imodeli ikhipha amathokheni okucabanga embhoshongweni ozenzakalelayo, noma amathokheni evidiyo enomsindo nesenzo asuka kumbhoshongo wokusabalalisa.

I-Cosmos 3 Edge isebenzisa i-2B dense transformer ku-reasoner yayo, futhi ilandela izimiso zemiyalezo ehambisana ne-Qwen3-VL yokufakwayo kwesithombe nevidiyo, ngenqolobane ye-Cosmos GitHub.

'+o.n+' '+o.d+' ';ping();}; dr.appendChild(b);}); dr.firstChild.click(); var EMB=[ {n:’Camera’,dim:’9D’,d:’Camera motion. Used for vision AI agents reasoning across live video.’}, {n:’Autonomous vehicle’,dim:’9D’,d:’Ego pose and movement for road-scene understanding and intent prediction.’}, {n:’Egocentric’,dim:’57D’,d:’First-person embodied motion.’}, {n:’Single-arm robot’,dim:’10D’,d:’End-effector pose plus grasp state (DROID / UR / Fractal / Bridge / UMI).’}, {n:’Dual-arm robot’,dim:’20D’,d:’Two coordinated arms (dual DROID arms).’}, {n:’Humanoid’,dim:’29D’,d:’Whole-body humanoid control (AgiBot).’} ]; var er=document.getElementById('c3eEmb'); EMB.forEach(function(o,i){var b=document.createElement('div');b.className=”c3e-chip”+(i===3?' on':'');b.textContent=on; b.onclick=function(){er.querySelectorAll('.c3e-chip').forEach(function(e){e.classList.remove('on');});b.classList.add('on'); document.getElementById('c3eEmbOut').inner=' '+o.n+' I-'+o.dim+' ye-action vector '+o.d+' ';ping();}; er.appendChild(b);}); er.querySelectorAll('.c3e-chip')[3].chofoza(); // shintsha usayizi ngokuzenzakalela: thumela eyakho i-offsetHeight (+40) kumsebenzi womzali othi ping(){try{var h=document.getElementById('c3e-app').offsetHeight+40; parent.postMessage({c3e:'ubude',ubude:h},'*');}catch(e){}} window.addEventListener('load',ping);window.addEventListener('resize',ping); setTimeout(ping,300); setTimeout(ping,900); })(); </f.length-1>

“>

Ukumelwa kwesenzo esisodwa kuwo wonke ama-embodiments

Amasistimu omzimba achaza izenzo ngendlela ehlukile. Imoto isebenzisa i-ego pose nokunyakaza. Ikhamera isebenzisa ukunyakaza kwekhamera. Ingalo yerobhothi isebenzisa ukuma kwe-effector yayo, futhi i-gripper yengeza isimo sokubamba.

I-Cosmos 3 yenza lokhu kufanekiselwa kube ukumelwa kwesenzo esifanayo. Izenzo zibhalwe ngekhodi njengamavekhtha ejiyomethri ahlangene athwebula ukuhumusha, ukuzungezisa, nesimo sokukhohlisa.

Lokhu kuxhumanisa ukulawula nesakhiwo esibukwayo somhlaba. Imodeli ihlobanisa izinguquko ze-pixel nokunyakaza komzimba nokokufaka kokulawula. Ividiyo ekhiqiziwe ibe ngaphezu kokuqagela. Imelela indlela umhlaba okufanele ushintshe ngayo ekuphenduleni isenzo.

Ubukhulu besenzo esisekelwayo buncike ekufanisweni komuntu. Inqolobane ye-Cosmos GitHub ibala ukunyakaza kwekhamera (9D), imoto ezimele (9D), i-egocentric motion (57D), irobhothi lengalo eyodwa (10D), irobhothi elinezingalo ezimbili (20D), nerobhothi elinobuntu (29D).

Imodi yenqubomgomo isebenza kuzo zombili izinkomba

Njengomgomo, i-Cosmos 3 Edge ibikezela isenzo kanye nomphumela waso olindelekile wokubonwayo. Isimo samanje siyangena; isenzo kanye nomphumela waso obonakalayo ongase uphume.

Isenzo sigeleza kuzo zombili izinkomba. Imodeli ingabikezela umthelela wesenzo, noma inqume isenzo emthelela waso. Lokhu kuxhumanisa ukumodela komhlaba ngqo ekuqeqeshweni nasekuhlolweni kwenqubomgomo yamarobhothi.

I-NVIDIA nayo ikhishwe Inqubomgomo ye-Cosmos 3 Edge (DROID). Kuyinqubomgomo yokukhohlisa irobhothi ngemuva kokuqeqeshwa kudathasethi ye-DROID yemisebenzi yokukhetha nendawo, nemibhalo yangemuva kokuqeqeshwa ifakiwe. Onjiniyela bangakwazi ukushuna kahle iqoqo elincane le-H100 noma Isiteshi se-NVIDIA DGX ngaphambi kokuthunyelwa.

Ingabe Iyasebenziseka

I-Cosmos 3 Edge iletha inkumbulo esebenza kahle kuwo wonke amakhompyutha onqenqema we-NVIDIA. Okuhlosiwe kufaka phakathi i-NVIDIA RTX PRO GPUs, i-NVIDIA DGX, i-GeForce RTX GPUs, ne-NVIDIA Jetson, okuhlanganisa amamojula asanda kumenyezelwa e-Jetson T2000 kanye ne-T3000.

Njengemodeli yesenzo somhlaba esiqeqeshwe ngemuva kokuqeqeshwa (i-WAM), imodeli isebenza ngokulungiswa kwerobhothi kokubonwa okungu-640×360. Ku-NVIDIA Jetson Thor ikhiqiza izenzo ezingama-32 ngokuqonda ngakunye, kuyilapho ifinyelela ukulawula kwesikhathi sangempela ku-15 Hz. Okokugcina, i-Edge tier isekela izinqumo ezingu-256p nezingu-480p, 12–30 fps, kanye nozimele abangu-50–150.

Besebenzisa uhlaka oluvulekile lwe-Cosmos, abathuthukisi bangakwazi ukuqeqesha i-Cosmos 3 Edge ukuze bafane nomuntu othile kanye nenzwa ebekwe cishe usuku. I-NVIDIA ibeka i-GeForce RTX 3070 noma engcono njengendawo yomgwaqo etholakala endaweni ye-prototyping.

<path d="M2 4l4 4 4-4" stroke="%2376B900" stroke-width="1.6" fill="none"/>“); ukuphinda okungemuva:akuphindwa;indawo engemuva:indawo engakwesokudla engu-11px} khetha:gxila{outline:none;border-color:var(–green-line)} /* isinqumo */ .dc-verdict{margin-top:14px;border-radius:11px;padding1pxb solider:13pxb; var(–line); ingemuva:var(–panel2); isibonisi:flex;qondanisa-izinto:center;gap:12px} .dc-icn{flex:0 0 auto;width:34px;height:34px;border-radius:9px;display:flex;align-ify-center:center; usayizi wefonti:18px;isisindo-sefonti:900} .dc-vtext{flex:1} .dc-vtext .v{font-size:14px;font-weight:800;line-height:1.15} .dc-vtext .d{font-size:11.5px-top-top:2) .dc-verdict.ok{border-color:var(–green-line);ingemuva:var(–green-soft)} .dc-verdict.ok .dc-icn{background:var(–green);umbala:#0b0b0b} .dc-verdict.ok .v{umbala:var(–green)} .dc-verdict.tight{border-color:rgba(224,168,58,.45);ingemuva:rgba(224,168,58,.1)} .dc-verdict.tight .dc-icn{background:var(–amber);umbala:#0b0bcight-verdict.dd}. .v{color:var(–amber)} .dc-verdict.no{border-color:rgba(224,106,90,.45);background:rgba(224,106,90,.1)} .dc-verdict.no .dc-icn{background:var(-verb:0-bdnob.0}); .v{color:var(–red)} .dc-meta{display:flex;gap:8px;margin-top:10px;flex-wrap:wrap} .dc-meta .m{flex:1;min-width:120px;background:var(–panel2);border:1px solid var(–line); irediyasi yomngcele:9px;padding:9px 11px} .dc-meta .m .k{font-size:9.5px;letter-spacecing:.05em;text-transform:uppercase;umbala:var(–thulisiwe);isisindo sefonti:700}mc.d. .val{font-size:13px;font-weight:800;margin-top:2px} .dc-meta .m .val b{color:var(–green)} .dc-lic .ok{color:var(–green)} .dc-how{margin-top:12px;i-backp-order:1p-backp:2) var(–line); irediyasi yomngcele:10px;padding:11px 13px} .dc-how .h{font-size:10px;font-weight:800; letter-spacing:.06em;text-transform:uppercase;umbala:var(–green); imajini-ngezansi:7pxwm. ol{margin:0;padding-left:17px;font-size:12px;color:#dcdcdc;line-height:1.55} .dc-how ol code{background:#0d0d0d;border:1px solid var(–line);irediyasi yomngcele:1pxpadding; 5px;usayizi wefonti:11px;umbala:var(–green)} .dc-indlela{color:var(–green); text-decoration:none;font-weight:700} .dc-foot{display:flex;justify-content:space-between;align-items:center;6wp;wrap:gap i-margin-top:12px;padding-top:10px;border-top:1px solid var(–line)} .dc-foot .note{font-size:10px;color:var(–muted);flex:1;min-width:180px} .dc-foot .brand:1-muted.px(5)} .dc-foot .brand b{color:var(–green)} @media(max-width:480px){.dc-title{font-size:15px}} <body> <div id=""dc-w""> <div class=""dc-card""> Sebenzisa isheke4B · imodeli yomhlaba enemibhoshongo emibili Ingabe i-Cosmos 3 Edge iyasebenziseka ekusetheni kwakho? Khetha ukunemba kanye nedivayisi eqondiwe. Isheke lilinganisela ukulingana kwememori kanye nocingo lwelayisensi. Ukunemba Idivayisi eqondiwe <select id=""dcDev""/> <div class=""dc-how""> Indlela yokusebenzisa <ol> <li>Landa izisindo: nvidia/Cosmos3-Edge ku-Hugging Face.</li> <li>Layisha ngohlaka oluvulekile lwe-Cosmos ngokunemba kwakho okukhethile (<code>BF16</code>, <code>FP8</code>noma <code>NVFP4</code>).</li> <li>Isitimela sokuhamba somfanekiso wakho kanye nezinzwa ku-H100 cluster noma i-DGX Station – cishe usuku – bese uphakela ekuqondisweni konqenqema.</li> </ol> </div> </div> </div> <script><![CDATA[ (function(){ var PARAMS=4; // total params in billions (vendor-stated total) var RUNTIME=1.5; // GB runtime/KV at short context var QUANT=[ {id:'bf16',n:'BF16',b:2,note:'Full precision'}, {id:'fp8',n:'FP8',b:1,note:'8-bit'}, {id:'nvfp4',n:'NVFP4',b:0.5,note:'4-bit'} ]; var DEV=[ {n:'Jetson T2000 (16 GB)',mem:16,edge:true}, {n:'Jetson T3000 (32 GB)',mem:32,edge:true}, {n:'Jetson AGX Thor (128 GB)',mem:128,edge:true}, {n:'GeForce RTX 3070 (8 GB)',mem:8,edge:false}, {n:'GeForce RTX 4090 (24 GB)',mem:24,edge:false}, {n:'RTX PRO 6000 (96 GB)',mem:96,edge:false} ]; var qi=2, di=0; // default NVFP4 on the smallest edge module var qWrap=document.getElementById('dcQuant'); QUANT.forEach(function(q,k){var c=document.createElement('div');c.className="dc-chip"+(k===qi?' on':''); c.innerHTML=q.n+''+q.note+''; c.onclick=function(){qi=k;qWrap.querySelectorAll('.dc-chip').forEach(function(e){e.classList.remove('on');});c.classList.add('on');calc();}; qWrap.appendChild(c);}); var sel=document.getElementById('dcDev'); DEV.forEach(function(d,k){var o=document.createElement('option');o.value=k;o.textContent=d.n;sel.appendChild(o);}); sel.value=di; sel.onchange=function(){di=+sel.value;calc();}; function fmt(x){return (Math.round(x*10)/10)+' GB';} function calc(){ var q=QUANT[qi], d=DEV[di]; var weights=PARAMS*q.b; var req=weights+RUNTIME; document.getElementById('dcReq').innerHTML='~'+fmt(req)+''; document.getElementById('dcAvail').textContent=fmt(d.mem); var vd=document.getElementById('dcVerdict'), icn=document.getElementById('dcIcn'), v=document.getElementById('dcV'), dd=document.getElementById('dcD'); vd.classList.remove('ok','tight','no'); if(d.mem>=req+2){ vd.classList.add('ok');icn.innerHTML='✓'; v.textContent="Yes u2014 deployable"; dd.textContent=q.n+' fits with headroom on '+d.n.split(' (')[0]+'.'; }else if(d.mem>=req){ vd.classList.add('tight');icn.innerHTML='⚠'; v.textContent="Yes, but tight"; dd.textContent="Fits at "+q.n+', little headroom for long context. Consider a lower precision.'; }else{ vd.classList.add('no');icn.innerHTML='✗'; v.textContent="Not at this precision"; dd.textContent="Needs ~"+fmt(req)+' but device has '+fmt(d.mem)+'. Try a lower precision or a bigger target.'; } var edgeTxt=d.edge?'On-device inference at 640u00d7360, 15 Hz control loop is the design target for Jetson.':'Good as a local prototyping / post-training on-ramp before deploying to a Jetson module.'; document.getElementById('dcNote').textContent="Memory = weights + ~1.5 GB runtime/KV (short context). "+edgeTxt; ping(); } calc(); // auto-resize using the deploy-check contract (__mtpH) function ping(){try{var h=document.getElementById('dc-w').offsetHeight+24; parent.postMessage({__mtpH:h},'*');}catch(e){}} window.addEventListener('load',ping);window.addEventListener('resize',ping); setTimeout(ping,250);setTimeout(ping,800); })(); ]]></script> </body>

Source link

nimda 3 weeks ago

0 2 3 minutes read