上周商汤的最新刚靠软银融资的新闻,给国内的AI创业公司们重新找回了点“关注度”。但就在前几天早上,引领AI走向现实世界的英伟达又发布新产品了,这对于很多喊着要做芯片 deaichuangyegongsidoushiyigezuihuaidexiaoxi。
yuanyinhenjiandan,zaibenzhousijubandegtc(gpujishufenghui)2018deribenzhanshangyingweidagongbulequanxinyidaideyunduangaoxiaolvban“aituiduanjiasuqi”——nvidia tesla t4。
tesla t4youduoqiang?
xianlaishuoshuotesla t4dedingwei,tashijishiyikuanyunduanzhuanyongdegpujiasuka,getouhenxiao(bangaopcieka),danshixingnenghenqiang。
genjuyingweidaguanfanggongbudeshuzi:tadedanfudian(fp32)yunsuannenglinengdadao8.1tflops,hunhejingduyunsuannenglinenggoudadao65 fp16 tflops,int8jingduyunsuannenglinenggoudadao130tops,int4jingduyunsuannenglinenggoudadao260tops。tongshi,juchuanyingweidayejiangzaizheikuanchanpinshangchangshiquanxindeint1yunsuanmoshi,genjuint4yunsuannenglituisuan,int1deyunsuannenglikenenghuidadaoshiwuqianlide1080tops。
jinxingnengruciqiangjing,dantesla t4zhenzhengzhuiqiudeshizhengtixitongxiaolverbushijueduixingneng。
这也符合英伟达“英文字母+4”这一命名规则的历史其他产品(例如M4和P4)。相比单卡更加强劲的TESLA家族其他型号,这几款产品的思路在明显不过——在最高效简洁的供电条件上,搭配最高效的核心配置,从而最终实现全服务器系统最高的工作效率。
以已经被证实在AI图像“推断”处理中更高效的INT8运算能力来说,T4达到了130TOPs,“上一代(Volta架构无该定位产品)”的P4只有区区22TOPs。这意味着,在不到两年的时间里,性能和能效比都增长了6倍。
zheigesudushizuihuihuangshiqide“moerdinglv”(18geyuefanyibei)doubunengbinide。
性能之外,“4”系列GPU加速卡还有3个必定遵循的设计原则:最高功耗始终不超过PCI-E供电最高的75W;始终采用更加紧凑的半高单卡槽设计(便于密集部署);始终采用无风扇的全被动散热方案。三个原则最终进一步加强了产品的实际效能。
dangran,dangejisuandanyuanxingnengjiaoruodewentizaishijiyingyongzhongyeyousuoyingxiang,jiunagpujiasukazuichangyongdeailingyuzhongde“moxingxunlian(tongguodaliangshujuheheshideshenjingwangluorangjisuanjixuehuiguilv)”laishuo,gengdadejisuandanyuanhecunchunengli(rongliang、duqusudu),jihebeishujiyouyuxiaode。
danxunlianbuxing,aihaixuyao“tuiduan(genjutiqianshengchengdeshenjingwangluoguilv,shuchujieguo)”,zheiyibufencaozuobingbuxuyaotebiedadejisuandanyuanhecunchunengli,erqiezaiyileiyingyongdeaisuanfa、shenjingwangluoxunliandaoyidingchengdudejichushang,shijishangxuyaodedoushi“tuiduan”nengli。
niyekeyixingxiangdelijiewei,“xunlian”shiyigebuduankaihuangheyouhuazhongzhijishudeguocheng,danzuizhongdenongtianguanlihenongzuowudeguanli,shijishangdoushi“tuiduan”laifuzede。
yimuqianaiyingyonglingyu“zouzaizuiqian”deanfanglingyulaishuo,guoneideyizhongjutouqiye,ruhaikangweishi、dahuadenghenzaodoucaiyongleyingweidadejishufangan,qizhongjiubaokuodalianganzhuangyoutesla p4deyunduanchuliqi。
yeyoubushaoaichuangyegongsi,yezhengzaiyanjiujinjunanfanglingyudekenengxing。chuleyunshichang,yingweidahaizaiduanshichangbushulejetsonxiliechanpin,xiangxinbujiudejianglai“tuling”jiagouyebiranhuigengxindaojetsonpingtaishang,zuizhongduijuedaduoshu、xiangyaozaoxinpiandeaichuangyegongsi,xingchengzhimingweixie。
nvidiazhengzaijiasushougeaixinpian?
在很多AI公司的宣传资料、券商、调研机构的报告中,对于AI芯片市场,总会有一句类似这样的判断——英伟达是“AI训练”领域的霸主,相比之下“AI推断”市场还有更多的机会。
danshiqingzhendeshizheiyangma?youqishizaikanwantesla t4zhihou,nihairenweizheishigekeyixinlaideshuofama?
zhishaowobuzheiyangrenwei,youqishilaohuangzaisiggraph(jisuanjituxingxuehuiyi)shangfabulequanxinde“tuling”jiagouzhihou,womenshijishangkeyiguanzhudaoyigehenzhongyaodequshi——benshenpianxiangtongyongchuliqidegpu,zhengzaiyufa“zhuanyonghua”。
yizuixindertx 2080tiweili,fenbieduiyingaichulidetensor core、duiyingguangxianzhuizongchulidert core,zongshujiudadaole612ge。zheiyekeyizuoweiyingweidayingdui“moerdinglv”jijiangshixiaodeyigeshouduan,jiranzhichenghejichujishuwufadailaitisheng,womenjiucongxinpiandejiagouhexiaolvshangxunzhaokongjian。
zai“qiao”buguodeshi,zuijinjiniangpuzhongyujianghuanxiangleshushinianderengongzhinengbianchenglexianshi,bingqiezaiduofangdegongtongnulixia,yufakaizhisanye。
jinguanyingweidazijidequeconglaimeiyoushuoguo“longduanshichang”、“weiwoduzun”zhileidehuayu。baokuozhumingde“piyijiaozhu”,haishiyirujiwangdeheaikeqin。danyingweidadejishuyanfashiliyijiuzaineili,bufagengshicongweitingxie,buduanxiangaijieshuchuquanxindelinian。
yingweidaguanfangdeyanjiubiaoming,int8yunsuanmoshinenggoudafutigaoaituxiangchulidesulv
jiulirucongtesla p4 gpujiasukakaishizhichideint8yunsuan,tashijishangshiyizhonggengdijingdudeyunsuanfangshi,dangenjuyingweidaguanfangzijideyanjiu,zaihenduoyingyongchangjingzhong,zheizhongjingdudexiajiangbingbuhuimingxianyingxiangjieguo,fanernenggourangxinpianxingnengbeishushangsheng(xiangduiyujingdugenggaodefp16、fp32)yunsuan。
zheizhongquqiaodeluxianhenkuaiyeyingxianglezhenggeyejiederenshi,yushihu,aichuangyegongsimenyekaishizaishoukuanhuozheshixiayidaichanpinzhongjiaruint8deyunsuannengli。
kejieguoshi,zheitouaichuangyegongsimende“zizhuxinpian”haimeiyoukandaoshenmeyingzi,neitouyingweidayougengxinle,erqiehaidailaileint4yunsuannengli,tongshihaizhichishiyanxingdeint1yunsuannengli。houliangzhemuqiandeyanjiuhaibushen,danxiangxinyingweidabukeneng“kongxuelaifeng”。
yidanzhengmingleint4naizhiint1gongzuoxiaolvgenggao,neimehenduoaichuangyegongsihualedaqiancaizaochulaidexinpianhechanpin,gangchushengjiuyijingluohoule。
yingweidadepingtaituozhannengliyezhideguanzhu,yizheicigtc japanweili,yingweidachulefabulequanxindetesla t4 gpujiasuqizhiwai,tongshihaigongbulelianggequanxindekaifapingtai——agx、clara。qianzheduiyingdeshizidongjixie,houzheduiyingdeshiyiliaoyingyong。
buguoyingweidaduiyuzheilianggelingyubingbumosheng,yingweidadechanpinshijishangyezaoyijingyingyongdaolezheilianggelingyudangzhong,danchuangjianpingtaizheiyijudong,huoxuyiweizheyingweidazai2019nianbazheilianggedangzuozhuyaofazhanfangxiangle。
zongdelaikan,yingweidazheigebangzhuaijueqi、bandaotijishushenhou、buduantuozhanshichangdeguigujutou,zhengzaibuduanchixiaaishichangdehongli,zheiyeshigongsiyinglihechuangzaojiazhidejichumubiaosuozhiyindefangxiang。danchuangyegongsizhendebixujiasule,zaibupaokuaidian,neijiuzhendeyaosizaijutoude“jiaozhang”xiale。