+− THE DAILY DIFFdev & AI news
NEEDS REVIEW

Imemori yeNvidia CMP 170HX: i-HBM ebonakalayo, ukucushwa okusebenzayo kanye nokukhubazeka

UKate ubika ukudalula imemori esivele ifakiwe ku-accelerator ye-Nvidia CMP 170HX yezimayini.

Want to see our stories first on Google? Add us to your preferred sources — it helps us grow.

UKate ubika ukudalula imemori esivele ifakiwe ku-accelerator ye-Nvidia CMP 170HX yezimayini. Uhlobo lwesishiyagalombili-gigabyte luqukethe cishe amashumi ayisishiyagalombili-gigabyte ngokwenyama, kanti umthamo osebenzisekayo uncike kuhadiwe olungonakalanga kanye nokucushwa okusebenzayo. Ikhadi elilodwa elinama-gigabyte angamashumi ayisithupha nane kuphela alizange lithole inzuzo ye-framebuffer kunokuvula isofthiwe yangaphambilini.

Funda uhlelo olubhaliwe (isiNgisi) ↗

Lokho okuhlanganiswa yile vidiyo

  • Imemori ebonakalayo, umthamo olungiselelwe ngempumelelo kanye nememori esebenzisekayo ngokuthembekile kwehlukile.
  • Izinguquko eziyisishiyagalombili-gigabyte kanye neziyishumi-gigabyte ze-CMP 170HX ziqukethe amaphakheji ememori ebonakalayo ahlukene.
  • Amarekhodi wokucushwa kwakamuva angashintsha izilungiselelo ezisebenzayo ngenkathi amamaki efuse abonakalayo ehlala engunaphakade.
  • Ukuqhathaniswa kokusebenza okubikiwe kukaKate kukala i-FP16 half2 FMA arithmetic ku-OpenCL, ngaphandle kokuqhathanisa okungazange kubhekwe kokuhlolwa kwejubane.
  • Umonakalo wokukhiqiza uhlala ngemuva kokuvula ukucushwa; ikhadi lomlobi elingamashumi ayisithupha nane gigabyte alizange lithole umthamo oweqile we-framebuffer.

Imibuzo nezimpendulo

Kungani i-GPU yezimayini ingaqukatha imemori eyengeziwe kunalokho izinhlelo zokusebenza ezingazisebenzisa?

Imemori efakiwe nomthamo ovezwe kuhlolwe ngokwehlukana. Ucwaningo lukaKate lubika ukuthi inguqulo ye-CMP 170HX yama-gigabyte ayisishiyagalombili iqukethe cishe amashumi ayisishiyagalombili gigabyte ngokwenyama, kanti imingcele yokucushwa iveza umthamo omncane. Inguqulo yama-gigabyte ayishumi inepakethe ehlukile enamashumi amane gigabyte ngokwenyama. Imemori esebenzisekayo isencike kuhadiwe olungonakalanga kanye nokucushwa umshayeli angakwazi ukukuphatha.

Buka lesi sigaba 0:19 · Kate's blog · NVIDIA Developer

Ama-fuse kagesi ahlala njalo angakuvumela kanjani ukuthi izilungiselelo ezisebenzayo zishintshe?

Amarekhodi wokucushwa kwakamuva angashintsha amanani asebenzayo ahunyushwe emlandweni womzimba ongenakuhlehliswa. UKate uchaza lokhu kuhlukaniswa ocwaningweni lwe-GA100, futhi iphethini ye-Nvidia exhunywe kumthombo inikeza umongo womqondo wamarekhodi efuse angakwazi ukuvalwa. Ifuse yomzimba eshile ihlala ishile. Lesi siqephu sichaza ukuhunyushwa ngaphandle kokuhlinzeka izinqubo zikagesi, amakheli ehadiwe noma iresiphi yokuguqulwa.

Buka lesi sigaba 1:28 · Kate's blog · NVIDIA patent

Ingabe i-benchmark ye-arithmetic ebikiwe isungula ukuthuthuka okusheshayo kwe-AI?

Ukuqhathaniswa okubikiwe kukala umsebenzi othize we-arithmetic wesigamu esinembile. Isibonelo sikaKate esivulwe ngehadiwe senza cishe amaphesenti ayishumi nesishiyagalombili ngaphambili kwe-A100 yokuthengisa ku-ProjectPhysX OpenCL FP16 half2 FMA microbenchmark. Lesi sihloko asihlinzeki ngokuqhathaniswa okungahlelwe ngamathokheni ngesekhondi. Ukufanelana kwemodeli, ukudluliswa kwe-arithmetic kanye nokusebenza kohlelo lokusebenza okuphelele ngakho kudinga ubufakazi obuhlukile, futhi izilinganiso zihlala ziyimiphumela yomlobi.

Buka lesi sigaba 2:50 · Kate's blog · ProjectPhysX

Kungani ikhadi elilodwa elingakhiyiwe lingatholanga imemori eyengeziwe esebenzisekayo?

Lelo khadi elithile lalinama-gigabyte angamashumi ayisithupha nane kuphela aphelele. UKate ubika ukuthi ukukhiya kwehardware akuzange kukhiqize inzuzo ye-framebuffer kunokukhiya isofthiwe yangaphambili, esivele yafinyelela leyo capacity. Ukushintsha umkhawulo kushiya ihadiwe elonakele lonakele. Isibonelo esiphikisayo sikhombisa ukuthi kungani imemori efakwe ngokwenyama kanye nomthamo osebenzisekayo ngokuthembekile kufanele kuhlolwe ngokwehlukana kwekhadi ngalinye.

Buka lesi sigaba 4:19 · Kate's blog · Amogh Munikote

Sasiyini isinqumo se-The Daily Diff mayelana nokuthembela ku-accelerator yezimayini eshintshiwe?

Isinqumo se-The Daily Diff sasiyisi-NEEDS REVIEW. U-Niko uzofuna ukuhlolwa okuzimele komthamo womsebenzi kanye nokuthembeka ngaphambi kokuthembela ekhadini lezimayini eliguquliwe njenge-accelerator. Lesi siqephu sibonga ucwaningo olushicilelwe lukaKate ngenkathi silondoloza imingcele yokulinganisa yomlobi kanye nesibonelo esiphikisayo sekhadi elinesici. Akubiki ukuguqulwa kwehardware okuzimele, ukukhiqizwa kabusha noma i-benchmark yohlelo lokusebenza eyenziwe uhlelo.

Buka lesi sigaba 4:43 · Kate's blog · ProjectPhysX

Umbhalo obhaliwe ohunyushwe

Kuhunyushwe kusuka ekulandiseni kokuqala kwesiNgisi. Umsindo namazwibela atholakalayo alawulwa yi-YouTube.

Imemori eyengeziwe yavelaphi?

0:00 I-GPU engama-gigabyte ayisishiyagalombili izwakala njengama-gigabyte ayisishiyagalombili ehadiwe. Umcwaningi uKate ubika amashumi ayisishiyagalombili akhona ngaphakathi kwekhadi lezimayini le-Nvidia. Kule vidiyo, ibifihlaphi imemori? Kungani ukuyivula kungakusalela mfushane? NgoLwesithathu, Novemba nanye, futhi lokhu The Daily Diff. Futhi elinye lamakhadi kaKate likhombisa ngqo ukuthi kuphuma kuphi ukuthuthukiswa kwamahhala. I-CMP eyodwa namashumi ayisikhombisa i-H X iyi-accelerator ye-Nvidia ye-Ethereum yezimayini,

Yini engaphakathi kulekhadi lezimayini?

0:22 yakhiwe ngokusetshenziswa kwe-GPU efanayo ne-A eyikhulu. UKate ubika ukuthi imingcele yokucushwa ifihla imemori esivele isepakethe. Umthamo osebenzisekayo usencike kuhadiwe elingenalutho kanye nokucushwa umshayeli angakwazi ukuphatha. UKate washicilela ucwaningo ngoNovemba nanye. Namuhla abafundi be-Hacker News baphikisana ngehadiwe efihliwe nokukhiqiza ukungasebenzi kahle. Kokubili ukuhlukaniswa komkhiqizo kanye nesivuno sokukhiqiza kubalulekile lapha. Isigaba samazwana sesivele sizithuthukisile saba sekhwalithi yama-semiconductor

0:46 ukuqinisekisa. Ingxenye ebiza kakhulu isivele ikhona.

Ingahlala kanjani imemori efakiwe ingabonakali?

0:48 Imemori enomkhawulokudonsa ophezulu, noma i-HBM, ihlezi eceleni kwe-GPU die ephaketheni. Incazelo yezakhiwo ze-Nvidia ihlukanisa i-chip egcwele ne- ukucushwa okudayiswa njenge-A eyikhulu. Ihadiwe ehlobene ingaba imikhiqizo ehlukile. UKate uhlonza izinhlobo ezimbili. Ikhadi le-gigabyte ayisishiyagalombili linama-gigabyte acishe abe amashumi ayisishiyagalombili afakwe ngokwenyama. Ikhadi le-gigabyte ayishumi linama-gigabyte acishe abe amane. Lawana amaphakheji ahlukene, ngakho-ke ilebula yomkhiqizo enkulu ingase ihambisane nememori encane

1:10 ngaphansi. Ukuthenga ngamagama emodeli kusanda kuthola umsebenzi ongeziwe. Ukukhiqiza kungashiya izingxenye ze-chip zinesici. UKate uhlukanisa ukucushwa okukhubaza isici nerekhodi elihlonza ihadiwe elinesici. Ezinye izinsiza ezikhubazekile ziyasebenza lapho zidalulwe. Ibhange lemimoriyamu eliphukile ngokomzimba lihlala liphukile, noma ngabe i-GitHub yakho readme iba yinto eshukumisayo kangakanani. Ama-fuse kagesi yizakhiwo ezincane zomzimba ezingunaphakade

Kungani ama-fuse ahlanganisa ngendlela eyodwa evumela ukucushwa okusha?

1:31 irekhodi izilungiselelo. I-fuse eshile ihlala ishile. Izilungiselelo ezisebenzayo zingakodwa ngokusebenzisa amarekhodi okucushwa. UKate uchaza amarekhodi akamuva eshintsha amanani asebenzayo angaphambilini. Umlando womzimba uhlala, ngenkathi ukucushwa okuhunyushwe kuwo kuyashintsha. Cabanga ngedokhumenti ene- umyalelo wokuqala nesichibiyelo sakamuva. Ukwengeza isichibiyelo kushintsha ukuthi yimuphi umyalelo osebenzayo, ngenkathi uyinki wokuqala uhlala.

1:51 Imemori efakiwe ayishintshile. Akekho olungise i-HBM eyonakele ngokukholisa i-fuse ukuthi iphinde ikhule. UKate ukhomba umphumela onzima. Ukushisa ama-fuse amaningi kunganciphisa eminye imingcele. Ngemuva kwalokho, isibuyekezo sokuklama esinengqondo kakhulu sithola ukuthi igama elingunaphakade lidinga ikholomu yesibili kuspredishithi. Ucwaningo lubika ukuphazamiseka okufushane kukagesi okuphazamisa ukuhlolwa kwe-controller futhi kuvumela izinguquko zokucushwa unomphela.

2:13 Sizogcina le ncazelo icatshangelwe. Ukuphindaphinda izinqubo zehardware kungalimaza ikhadi. Ukufunda lo mphumela kukunika imodeli yengqondo; akukuniki iwaranti. Ukukhiya isofthiwe okubizwa nge-CMP unlocker sekuvele kukhona.

Kwakungayenza yini isofthiwe yangaphambilini ebingasebenzi?

2:24 Ukuqhathaniswa kukaKate kubeka inguqulo ye-gigabyte eyisishiyagalombili kumashumi ayisithupha nane ama-gigabyte asebenzisekayo ngaleyo ndlela. Ukukhiya kwehardware kungafinyelela amashumi ayisishiyagalombili uma imemori iphelele. UKate ubika le miphumela yehardware; ukuphinda kusetshenziswe ihadiwe yezimayini kunesikhathi eside Umlando. UKate ubika ukuhlolwa kwememori egcwele ngokubhala idatha kuyo yonke indawo futhi ukuyifunda emuva ngemuva kokusula i-GPU cache. Lesi sihloko sibika nokusebenza kwemisebenzi yokubikezela. Lezo yizivivinyo zomlobi, zinamandla kunokufunda umthamo. Asizange sihlolwe ngokuzimele ikhadi elishintshiwe.

2:49 I-A eyikhulu ye-retail inamaphrosesa ayikhulu nesishiyagalombili okusakaza,

Kulinganiswa yini umphumela osheshayo kune-A100?

2:53 amaqembu enza umsebenzi. UKate ubika ukunika amandla amaqembu amaningi asebenzayo. Ikhadi elilinganisiwe linamashumi ayikhulu namashumi amabili nane. Ukuthi yimaphi amaqembu asebenzayo kuncike ku-die ngayinye. I-microbenchmark ye-arithmetic yesigamu esinembile sikaKate ibeka ikhadi elikhiywe ngehadiwe cishe ngamaphesenti ayishumi nesishiyagalombili phambili kwe-A eyikhulu ye-retail. Ikala umsebenzi ophindaphindwayo othize ku-OpenCL.

3:10 Ukufunda lokho njengesithuthukisi kuyo yonke imodeli ye-AI kunikeza ishadi ukuthuthuka elingakaze lithole. Abafundi be-Hacker News baphakamise ukuhlolwa kwasefektri nokuthembeka. Izinhloso zabo ezincintisanayo aziniki ukuphindaphindwa okuzimele. Ngingakwazi ukwazisa ubunjiniyela futhi ngisafuna izilinganiso zomsebenzi kanye nokugijima okude kokuzinza ngaphambi kokwethemba ihadiwe ngomsebenzi engiwukhathalelayo. Imemori eyengeziwe esebenzisekayo inganquma ukuthi imodeli ingena yini ekhadini.

Yini esalokhu iyihlukanisa ne-accelerator esekelwe?

3:30 Isivinini sincike nasekusebenziseni umsebenzi nokuhambisa idatha ngesistimu. Lesi sihloko asihlinzeki ngokuqhathanisa okungahlelwe ngamathokheni ngesekhondi. Hlola umthamo wemodeli, ukudluliswa kwe-arithmetic kanye nokusebenza kohlelo lokusebenza ngokwehlukana. UKate uthi izici zokufaka ikhodi yezithombe nevidiyo kanye nemodi ye-GPU enezikhala eziningi azange zihlolwe ngokuqondile. Uxhumano kanye nokuxhumana kwe-PCI Express okusheshayo kudinga izinguquko ezengeziwe. Ibhodi, i-firmware kanye nehadiwe elisebenzayo kusazohlola ukuthi yini esebenzisekayo. UKate uthi lezi zinto zathunyelwa eqenjini lezokuphepha lemikhiqizo ye-Nvidia ngaphambi kokushicilelwa.

3:58 Okuthunyelwe kubonga omunye umcwaningi ngokuthola ngokuzimela ubungozi. Asikaze siqinisekise impendulo yomphakathi ye-Nvidia noma ukulungiswa okuhlolwe lapha. Izinguquko zokucushwa okungunaphakade zingakwazi ukuvala ihadiwe unomphela. Ukukhiya okuphumelelayo kusasele ukupholisa, amandla nokuthembeka okumele kuhloliwe. Isihloko esivelele somthengisi yisimangalo ngesampula elilodwa. Isithombe-skrini somthamo sikunika indawo yokuqala yokusihlola.

Kungani ikhadi elilodwa lingatholanga imemori eyengeziwe?

4:19 Manje ikhadi elonakalisa iphupho lokuthuthukiswa kwamahhala. UKate wayenelilodwa elinama-gigabyte angamashumi ayisithupha nane kuphela ephelele. Ukukhiya kwehardware akuzange kulinikeze inzuzo yememori kunokukhiya isofthiwe. Ukuvinjelwa kwashintsha; ihadiwe elonakele lihlala lonakele. Leso yisifundo esibalulekile sokuthenga. Umthamo ofakwe ngokwenyama kanye nomthamo osebenzisekayo ngokuthembekile kungaba izinombolo ezihluke kakhulu. Uma ukhetha ukukufunda lokhu kunokungizwa ngikusho, i-diff ifika ebhokisini lakho le-imeyili njalo ekuseni, mahhala ku-daily diff dot dev,

4:42 isixhumanisi ngezansi. Ngakho, isinqumo sanamuhla, NEEDS REVIEW.

Siyini isinqumo mayelana nokuvula?

4:44 Ngingathanda ukuhlolwa okuzimele komsebenzi kanye nokuthembeka ngaphambi kokusebenzisa ikhadi lezimayini elishintshiwe njenge-accelerator ethembekile. Ucwaningo lukaKate lwenza ihadiwe efihliwe iqondakale. Ukuthenga umphumela othembekile kudinga ubufakazi obuningi. Futhi yilokho okushintshile namuhla. Ngingu-Niko ovela kwa-Axrisi. Hlanganisa ngokuzibophezela.

Imithombo

  1. Using voltage glitching to unleash the A100 in my mining cardKate's blog
  2. NVIDIA Ampere Architecture In-DepthNVIDIA Developer
  3. OpenCL-BenchmarkProjectPhysX
  4. GA100VFI research provenanceKate
  5. Earlier CMPunlocker projectAmogh Munikote
  6. Integrated circuit having an enhanced fuseless fuse structureNVIDIA patent
  7. November 11 discussion of Kate's researchHacker News
  8. Jensen Huang at Stanford, April 30, 2026 — AnderseidesvikWikimedia Commons
  9. Portrait and complete artwork licence: CC BY-SA 4.0Creative Commons
  10. Toy Story process-commentary still sourceImgflip / Pixar-Disney media
  11. Futurama fictional sceptical reaction sourceGiphy / Gifsee
  12. Futurama GIF linked sourceGifsee

Niko from Axrisi

Nikoloz Turazashvili · Vexrail · Axrisi

LinkedIn · dev.to · GitHub · X · Substack

Amavidiyo ahlobene