近日，一款名为Xeon W5-3433的新款工作站处理器出现在SiSoftware数据库中，配备了16个核心和32线程，具有32MB的L2缓存，以及45MB的L3缓存。其运行的频率为1.99 GHz，应该是一个工程样品。页面上显示，其属于Fishhawk Falls平台产品。

按照去年曝光的信息，英特尔在新一代HEDT平台上，Sapphire Rapids HEDT处理器或许会有新的后缀，不再是“X”而是“W”，最多会配备56核112线程，采用LGA 4677插座，支持PCIe 5.0和DDR5内存。配套的芯片组同样不再使用“X”前缀，转而变成“W”，新款处理器对应的是W790芯片组。

图：早前曝光的Sapphire Rapids和8核Coffee Lake芯片对比

这意味着英特尔很可能完全放弃Core-X系列的命名方式，全部归类到支持Xeon的型号上。按照英特尔的路线图，新的HEDT平台可能会与第13代酷睿系列处理器的平台，也就是Raptor Lake及配套的700系列芯片组同步推出。不过近期传出用于数据中心的Sapphire Rapids-SP延期的消息，相关的HEDT平台产品很可能也会受到影响。

预览var tableid = settableid(); document.write(''); Table().init({ title: "英特尔HEDT处理器规格表", header: ["规格","Sapphire Rapids-X?","Cascade Lake-X","Skylake-X","Skylake-X","Broadwell-E","Haswell-E","Ivy Bridge-E","Sandy Bridge-E","Gulftown"], width: '92%', columnWidth: 'MAX', //表格单元宽度，"MAX"最大内容显示，""AVG"平均，"AUTO"自动，指定[15,16,20] id: tableid, data: [ ["发布时间","2022Q3?","2019Q4","2018Q4","2017Q3","2016Q2","2014Q3","2013Q3","2011Q4","2010Q1"], ["顶级型号","TBA","Core i9-10980XE","Core i9-9980XE","Core i9-7980XE","Core i7-6950X","Core i7-5960X","Core i7-4960X","Core i7-3960X","Core i7-980X"], ["工艺节点","10nm ESF","14nm++","14nm++","14nm+","14nm","22nm","22nm","32nm","32nm"], ["最大核心/线程数","56/112?","18/36","18/36","18/36","10/20","8/16","6/12","6/12","6/12"], ["L3缓存","TBA","24.75MB","24.75MB","24.75MB","25MB","20MB","15MB","15MB","12MB"], ["插座","LGA 4677?","LGA 2066","LGA 2066","LGA 2066","LGA 2011-3","LGA 2011-3","LGA 2011","LGA 2011","LGA 1366"], ["内存支持","DDR5-4800?","DDR4-2933","DDR4-2800","DDR4-2666","DDR4-2400","DDR4-2133","DDR3-1866","DDR3-1600","DDR3-1066"], ["PCIe版本","PCIe 5.0?","PCIe 3.0","PCIe 3.0","PCIe 3.0","PCIe 3.0","PCIe 3.0","PCIe 3.0","PCIe 2.0","PCIe 2.0"], ["PCIe通道数","64?","44","44","44","40","40","40","40","32"], ["主板芯片组","W790?","X299","X299","X299","X99","X99","X79","X79","X58"], ["最高TDP","TBA","165W","165W","165W","140W","140W","130W","130W","130W"] ] });

虽然目前英特尔的Intel 7工艺已日趋成熟，不过用在Sapphire Rapids上似乎有点不太适应。据ComputerBase报道，在美国银行证券全球技术会议上，英特尔数据中心和人工智能事业部总经理Sandra Rivera表示，Sapphire Rapids的产量提升并未按计划进行，比英特尔预期的时间更晚，目前英特尔的合作伙伴需要建立更多的平台并花费更多时间进行验证。

Sapphire Rapids原定于2021年发货，随后英特尔推迟到2022年初，后来又将批量生产的时间推迟到2022年中期。在2022年第一季度，英特尔向特定客户发送了Sapphire Rapids，之后又再次将大规模供货的时间延期到了下半年。此次Sandra Rivera的表态，意味着Sapphire Rapids又一次推迟了。

图：英特尔至强服务器处理器路线图

Sapphire Rapids的继任者是Emerald Rapids，按计划会在2023年末出现，两者采用相同的插槽。随着Sapphire Rapids不断延期，不少人担心Emerald Rapids也会出现同样的状况，不过暂时进度似乎没有受到影响。

英特尔早在2021年第三季度首次对外展示了Meteor Lake测试芯片，其采用了Tile设计，至少会有三个不同的模块，分别是计算模块、SOC-LP模块（负责I/O）和GPU模块。这些模块可以搭配不同制程节点的模块进行堆叠，包括使用其他晶圆厂制造的模块，再使用EMIB技术互联。通过Foveros封装技术，直接将不同IP、不同工艺的各种成熟方案封装在一起。英特尔会在Meteor Lake首次采用自家的Intel 4工艺，加入了EUV光刻技术。

Meteor Lake的GPU模块最低配置96个EU，最高可配置192个EU，相比Alder Lake和Raptor Lake有大幅度的提升。可能这是英特尔首款使用台积电（TSMC）制造的模块，采用3nm工艺制造，SOC-LP模块则采用台积电的4nm或5nm工艺制造，剩下的计算模块才是英特尔自己的Intel 4工艺。

此外，帕特-基尔辛格还确认了在2022年第一季度向特定的合作伙伴提供了Sapphire Rapids，更多的型号将会在今年内陆续推出。按照之前内存厂商宇瞻（Apacer）有关DDR5内存的介绍资料，Sapphire Rapids将会在2022年第三季度或第四季度正式到来。

有消息指出，英特尔新的HEDT产品将分为Sapphire Rapids-X和Alder Lake-X，前者为Cascade Lake-X的继任者，将专注于高端工作站市场，两者之间的定位可能类似于2017年时候的Skylake-X和Kaby Lake-X。据悉，Alder Lake-X有可能仍是LGA 1700平台，同样基于8个Golden Cove架构的性能核（Performance Core），以及基于8个Gracemont架构的能效核（Efficient Core）。

不过有人对此提出质疑，认为近期推出的酷睿i9-12900KS已经将目前8+8架构的性能榨干，同样的配置很难再有提升空间，所以另外一种说法是Alder Lake-X将是Sapphire Rapids-X的减配版本，两者都属于Fishhawk平台的一部分，不过会取消对ECC内存的支持，并减少PCIe通道数量。那么Alder Lake-X就不会有Gracemont架构的能效核，仅有Golden Cove架构的性能核，但性能核的数量大于8个，同时支持AVX-512指令集，面向极致发烧级游戏和主流工作站市场。

近期Ryzen Threadripper系列供应紧张，有人认为缺乏竞争使得AMD降低了其供应的优先级，将可用的芯片集中在EPYC处理器上。如果英特尔重新切入HEDT市场，或许会刺激AMD加大Ryzen Threadripper系列处理器的供应量。

近日，TheNextPlatform汇总了英特尔未来多款可扩展至强服务器处理器的信息。即将到来的Sapphire Rapids将采用小芯片设计，使得比Ice Lake-SP有更高的效率。其最高会配备56个核心，使用Golden Cove架构，采用10nm Enhanced SuperFin工艺制造，TDP为350W。新平台还支持PCIe Gen5、CXL 1.1（Compute Express Link）和八通道DDR5内存，同时会延续英特尔的内置AI加速策略，支持英特尔高级矩阵扩展（AMX）。据称，Sapphire Rapids基本频率是2.3 GHz。

此外，Sapphire Rapids还会推出两种版本。其中HBM版本会搭载四组HBM2e，每组容量为16GB，总有64GB内存，峰值带宽介乎于1.432 TB/s到1.640 TB/s之间，并会与普通版的Sapphire Rapids处理器共享插座（LGA 4677）。不过相比竞争对手AMD代号Genoa的EPYC处理器，Sapphire Rapids仍然有所欠缺。

Sapphire Rapids之后是Emerald Rapids，仍采用10nm Enhanced SuperFin工艺，也就是Intel 7工艺，整体效率会更高一些。预计Emerald Rapids将使用Raptor Cove架构核心，这是Golden Cove架构的优化版本，预计会有5%到10%的IPC提升。Emerald Rapids可以提供64核心和128线程，基本频率提高到2.6 GHz，配备120MB的L3缓存，支持DDR5-5600内存（最高容量4TB），TDP增加到375W。Emerald Rapids会与Sapphire Rapids保持兼容，仍使用Eagle Stream平台。

接替Emerald Rapids的将是Granite Rapids，英特尔会做出一些较大的改变。Granite Rapids将基于Intel 4工艺制造（原来的7nm EUV工艺），推出时间暂时不清楚，可能会在2023年到2024年之间出现，应该属于一个过渡性的解决方案。Granite Rapids将使用Redwood Cove架构核心，并会增加核心和线程数量，预计将达到120核心240线程，L3缓存为240MB。其基本频率为2.5 GHz，提供128条PCIe 6.0通道以及12通道DDR5内存。此前英特尔曾透露，Granite Rapids会在单个SoC中包含多个小芯片，通过EMIB封装，也会有HBM和Rambo cache芯片。

从Granite Rapids开始，英特尔可扩展至强服务器处理器将采用最新的AVX-1024/FMA3，以提高各种工作负载的性能，不过有可能意味着功率的增加。Granite Rapids会使用新的Mountain Stream平台，并于下一代产品保持兼容性。

自从AMD在2017年推出第一代EPYC处理器以后，在服务器市场上就逐渐对英特尔形成了性能上的压制。英特尔将翻盘的希望寄托在Diamond Rapids上，预计会在2025年与AMD代号Turin、基于Zen 5架构的EPYC处理器竞争。英特尔至今都没有太多透露Diamond Rapids的信息，预计会采用Intel 3工艺（原5nm工艺），以及使用增强型的Lion Cove架构核心，可提供144核心288线程，L3缓存为288MB，基本频率提高到2.7 GHz，并会引入DDR6-7200内存。

Diamond Rapids可能还会有代号为Sierra Forest的变体，有可能也是配备HBM内存的版本，针对特定客户提供。时间会晚于Diamond Rapids，应该要到2026年才会出现。

据了解，AMD计划将EPYC处理器的定价提高10%到30%，这是芯片短缺导致未来出货量下降的现实情况，AMD目前对此事并没有发表评论。定价提高的幅度和具体客户有关，大型的云端或数据中心客户提高的幅度会较小，至于供货情况暂时还不清楚。

AMD代号Milan和Milan-X的Zen 3架构EPYC处理器有着非常好的综合性能，其每瓦性能也占据着竞争优势，有助于降低运营成本。在晶圆和封装产能短缺时期，涨价似乎并不意外，而且有其合理性。毕竟现阶段半导体供应链中，几乎每个环节的制造成本都在上升，适当地将成本转嫁给客户有时候也是无奈之举。

报告指出，英特尔已经提高了Ice Lake的产量，预计2022年的供应量可能会同比增长50%。与AMD不同的是，为了保住市场份额，英特尔不会提高Ice Lake的价格，以削减AMD产品的竞争力。不过英特尔寄予厚望的Sapphire Rapids很可能会延期，全面量产将推迟到2022年第三季度。投资机构并不认为Sapphire Rapids能扭转形势，预测其定价会提高，英特尔的市场份额最终仍然会下降。

Raptor Lake仍采用big.LITTLE混合架构、Intel 7制程、以及LGA 1700底座，与Alder Lake-S使用的主板相兼容，仍然会支持DDR4内存。不同的是，其性能核（Performance Core）将由Golden Cove架构改为Raptor Cove架构，能效核（Efficient Core）仍沿用Gracemont架构，同时会优化缓存架构。英特尔也会推出对应的700系列芯片组，但不确定是否会引入DLVR技术。

预览var tableid = settableid(); document.write(''); Table().init({ title: "英特尔酷睿桌面处理器参数表", header: ["规格","Coffee Lake-S","Comet Lake-S","Rocket Lake-S","Alder Lake-S","Raptor Lake-S","Meteor Lake-S","Lunar Lake-S"], width: '92%', columnWidth: 'MAX', //表格单元宽度，"MAX"最大内容显示，""AVG"平均，"AUTO"自动，指定[15,16,20] id: tableid, data: [ ["发布年份","2018","2020","2021Q1","2021Q4","2022","2023?","2024?"], ["工艺节点","14nm","14nm","14nm","Intel 7","Intel 7?","Intel 4","Intel 4?"], ["核心架构","SkyLake","SkyLake","Cypress Cove","Golden Cove+Gracemont","Raptor Cove+Gracemont?","Redwood Cove+Gracemont?","Lion Cove+Skymont?"], ["核显","Gen9","Gen9.5","Gen12","Gen12","Gen12?","Gen 12?","Gen 13"], ["最大核心数","8","10","8","16 (8+8)","16 (8+16)?","TBC","TBC"], ["插座","LGA1151","LGA1200","LGA1200","LGA1700","LGA1700","LGA1700","TBC"], ["内存支持","DDR4","DDR4","DDR4","DDR4/5","DDR4/DDR5","DDR5","DDR5"], ["PCIe版本","PCIe 3.0","PCIe 3.0","PCIe 4.0","PCIe 5.0","PCIe 5.0","PCIe 5.0","PCIe 5.0"], ["主板芯片组","Intel 300系列","Intel 400系列","Intel 500系列","Intel 600系列","Intel 700系列?","TBC","TBC"] ] });今天有网友透露了更为细致的信息，英特尔700系列芯片组仅有Z790和B760两款，没有H710。此外，目前Alder Lake的Golden Cove架构内核并非“满血”，做了某些屏蔽，不过在Raptor Lake上可能会完全打开，预计会有不小的性能提升。据称，英特尔可能会在2022年第三季度中旬发布Raptor Lake。

英特尔明年应该也会推出Sapphire Rapids-X处理器的Fishhawk Falls平台，这是久违的HEDT平台更新，采用LGA 4677插座，搭配W790芯片组。该平台同样支持PCIe 5.0和DDR5内存，发布时间可能会在2022年第二季度末或第三季度初。

预览var tableid = settableid(); document.write(''); Table().init({ title: "英特尔HEDT处理器规格表", header: ["规格","Sapphire Rapids-X?","Cascade Lake-X","Skylake-X","Skylake-X","Broadwell-E","Haswell-E","Ivy Bridge-E","Sandy Bridge-E","Gulftown"], width: '92%', columnWidth: 'MAX', //表格单元宽度，"MAX"最大内容显示，""AVG"平均，"AUTO"自动，指定[15,16,20] id: tableid, data: [ ["发布时间","2022Q3?","2019Q4","2018Q4","2017Q3","2016Q2","2014Q3","2013Q3","2011Q4","2010Q1"], ["顶级型号","TBA","Core i9-10980XE","Core i9-9980XE","Core i9-7980XE","Core i7-6950X","Core i7-5960X","Core i7-4960X","Core i7-3960X","Core i7-980X"], ["工艺节点","10nm ESF","14nm++","14nm++","14nm+","14nm","22nm","22nm","32nm","32nm"], ["最大核心/线程数","56/112?","18/36","18/36","18/36","10/20","8/16","6/12","6/12","6/12"], ["L3缓存","TBA","24.75MB","24.75MB","24.75MB","25MB","20MB","15MB","15MB","12MB"], ["插座","LGA 4677?","LGA 2066","LGA 2066","LGA 2066","LGA 2011-3","LGA 2011-3","LGA 2011","LGA 2011","LGA 1366"], ["内存支持","DDR5-4800?","DDR4-2933","DDR4-2800","DDR4-2666","DDR4-2400","DDR4-2133","DDR3-1866","DDR3-1600","DDR3-1066"], ["PCIe版本","PCIe 5.0?","PCIe 3.0","PCIe 3.0","PCIe 3.0","PCIe 3.0","PCIe 3.0","PCIe 3.0","PCIe 2.0","PCIe 2.0"], ["PCIe通道数","64?","44","44","44","40","40","40","40","32"], ["主板芯片组","W790?","X299","X299","X299","X99","X99","X79","X79","X58"], ["最高TDP","TBA","165W","165W","165W","140W","140W","130W","130W","130W"] ] });AMD代号Raphael的处理器将基于Zen 4架构，或许会称为Ryzen 7000系列，使用全新的AM5插座，采用台积电5nm工艺制造，IOD则是6nm或7nm工艺，集成RDNA 2架构核显，支持PCIe 5.0以及双通道DDR5-5200内存，TDP介乎于105W-120W之间，并提供5GHz左右的频率。除了架构的改进和性能的提升外，传闻AMD也将对温度和电源管理做进一步优化，使得更加容易读取温度信息，并改进电源管理技术。

这位网友表示，随之上市的还有X670芯片组，B650芯片组会稍晚一个月左右，不过AM5平台无法兼容DDR4内存。发布时间可能会在2022年第二季度末或第三季度初，比之前传言的时间早了大概一个季度。

预览var tableid = settableid(); document.write(''); Table().init({ title: "AMD桌面平台Ryzen处理器参数", header: ["型号","Ryzen 1000系列","Ryzen 2000系列","Ryzen 3000系列","Ryzen 5000系列","-","Ryzen 7000系列"], width: '92%', columnWidth: 'MAX', //表格单元宽度，"MAX"最大内容显示，""AVG"平均，"AUTO"自动，指定[15,16,20] id: tableid, data: [ ["代号","Summit Ridge","Pinnacle Ridge","Matisse","Vermeer","Warhol(已取消)","Raphael"], ["插座","AM4","AM4","AM4","AM4","AM4","AM5"], ["架构","Zen","Zen+","Zen 2","Zen 3","Zen 3+","Zen 4"], ["GPU架构","-","-","-","-","-","RDNA 2"], ["制造工艺","14nm","12nm","7nm","7nm","6nm","5nm"], ["内存类型","DDR4","DDR4","DDR4","DDR4","DDR4","DDR5"], ["PCIe版本","3.0","3.0","4.0","4.0","4.0","5.0"], ["最大核心/线程数","8C/16T","8C/16T","16C/32T","16C/32T","16C/32T","16C/32T"], ["推出时间","2017年","2018年","2019年","2020年","2021年","2022年Q3?"], ] });这次流传的消息里，比较让人感到意外的是，AMD明年可能会在AM4平台推出Renior X。这是一款没有核显的产品，规格与Ryzen 4000G系列相当，或许不会有R7级别型号，竞争目标是英特尔酷睿i3处理器，具体上市时间仍未确定。

AMD在明年初的CES 2022上，还会发布采用3D垂直缓存（3D V-Cache）技术的Zen 3架构桌面处理器（Vermeer S），不过至今也没有确认这属于Ryzen 5000还是6000系列。这有可能是AM4平台最后一款真正意义上的新产品，为沿用数年的AM4插座划上完美句号。

此外，AMD明年应该也会更新HEDT平台，推出代号“Chagall(Pro)”的Ryzen Threadripper(Pro)5000系列，这是基于Zen 3架构的产品。不过也消息指，可能基于Zen 3+架构，而不是Zen 3架构。该系列发布时间已一再推迟，估计与HEDT平台暂时缺乏竞争对手有关，传闻也会在CES 2022亮相。

预览var tableid = settableid(); document.write(''); Table().init({ title: "AMD Ryzen Threadripper系列处理器参数", header: ["型号","Threadripper 2000系列","Threadripper 3000系列","Threadripper 5000系列","Threadripper Pro 3000系列","Threadripper Pro 5000系列"], width: '92%', columnWidth: 'MAX', //表格单元宽度，"MAX"最大内容显示，""AVG"平均，"AUTO"自动，指定[15,16,20] id: tableid, data: [ ["代号","Coflax","Castle Peak","Chagall","-","Chagall Pro"], ["插座","sTR4","sTRX4","sTRX4","sWRX8","sWRX8"], ["架构","Zen","Zen 2","Zen 3","Zen 2","Zen 3"], ["制造工艺","14nm GF","7nm TSMC","7nm TSMC","7nm TSMC","7nm TSMC"], ["最大L3缓存","64MB","256MB","256MB","256MB","256MB"], ["芯片结构","4 CCD","8 CCD+1 IOD","8 CCD+1 IOD","8 CCD+1 IOD","8 CCD+1 IOD"], ["内存类型","DDR4","DDR4","DDR4","DDR4","DDR4"], ["内存通道数","4","4","4","8","8"], ["内存频率","2933MHz","3200MHz","3200MHz","3200MHz","3200MHz"], ["PCIe版本","3.0","4.0","4.0","4.0","4.0"], ["PCIe通道数","64","64","64","128","128"], ["最大核心/线程数","32C/64T","64C/128T","64C/128T","64C/128T","64C/128T"], ["最大功耗","200W","280W","280W","280W","280W"], ["推出时间","2018年","2019年","2022?年","2021年","2022年?"], ] });

据推特用户@kopite7kimi透露，英特尔明年在HEDT平台上可能会卷土重来，推出Sapphire Rapids-X处理器的Fishhawk Falls平台，时间会在2022年第三季度。Raptor Lake很可能也会在同一个季度内发布，这意味着英特尔在桌面市场的下一次更新大概相隔一年。

传闻Sapphire Rapids-X最多会配备56核112线程，采用10nm Enhanced SuperFin工艺制造，使用LGA 4677插座，提供最多64个PCIe 5.0通道，支持8通道DDR5内存，显然这样的配置是针对专业工作站和极致发烧友。配套的可能是W790芯片组，不过近期有消息指或许仍会有X699芯片组，以针对不同的使用人群。

预览var tableid = settableid(); document.write(''); Table().init({ title: "英特尔HEDT处理器规格表", header: ["规格","Sapphire Rapids-X?","Cascade Lake-X","Skylake-X","Skylake-X","Broadwell-E","Haswell-E","Ivy Bridge-E","Sandy Bridge-E","Gulftown"], width: '92%', columnWidth: 'MAX', //表格单元宽度，"MAX"最大内容显示，""AVG"平均，"AUTO"自动，指定[15,16,20] id: tableid, data: [ ["发布时间","2022Q3?","2019Q4","2018Q4","2017Q3","2016Q2","2014Q3","2013Q3","2011Q4","2010Q1"], ["顶级型号","TBA","Core i9-10980XE","Core i9-9980XE","Core i9-7980XE","Core i7-6950X","Core i7-5960X","Core i7-4960X","Core i7-3960X","Core i7-980X"], ["工艺节点","10nm ESF","14nm++","14nm++","14nm+","14nm","22nm","22nm","32nm","32nm"], ["最大核心/线程数","56/112?","18/36","18/36","18/36","10/20","8/16","6/12","6/12","6/12"], ["L3缓存","TBA","24.75MB","24.75MB","24.75MB","25MB","20MB","15MB","15MB","12MB"], ["插座","LGA 4677?","LGA 2066","LGA 2066","LGA 2066","LGA 2011-3","LGA 2011-3","LGA 2011","LGA 2011","LGA 1366"], ["内存支持","DDR5-4800?","DDR4-2933","DDR4-2800","DDR4-2666","DDR4-2400","DDR4-2133","DDR3-1866","DDR3-1600","DDR3-1066"], ["PCIe版本","PCIe 5.0?","PCIe 3.0","PCIe 3.0","PCIe 3.0","PCIe 3.0","PCIe 3.0","PCIe 3.0","PCIe 2.0","PCIe 2.0"], ["PCIe通道数","64?","44","44","44","40","40","40","40","32"], ["主板芯片组","W790?","X299","X299","X299","X99","X99","X79","X79","X58"], ["最高TDP","TBA","165W","165W","165W","140W","140W","130W","130W","130W"] ] });

根据英特尔的介绍，这种采用HBM的Sapphire Rapids处理器既可以选择不使用DDR5内存，将HBM内存作为主内存（就像富士通A64FX处理器），也可以选择DDR5内存与HBM内存混用的方法，这时候HBM内存更像是一个大型缓存，处于处理器和DDR5内存之间。

从这次展示采用HBM的Sapphire Rapids处理器可以看到，Sapphire Rapids有四个小芯片，每个对应两个HBM内存堆栈，内存位宽为2048位。目前，JEDEC的HBM2E规范里，最高数据传输速率为3.2 GT/s，不过SK海力士去年已开始批量生产数据传输速率为3.6 GT/s的版本，意味着采用HBM的Sapphire Rapids处理器最高可以有3.68 TB/s的内存带宽，但内存容量仅限于128GB。此外，这款采用HBM的Sapphire Rapids处理器使用的BGA封装，由于LGA 4677空间已经很有限，没办法提供足够空间容纳HBM堆栈。

采用HBM内存可以带来高性能的内存子系统，不过也意味着会有较高的频率和TDP，功耗和发热量是一个需要面对的问题。按照英特尔的说法，这种采用HBM的Sapphire Rapids处理器只会提供给特定的客户，主要针对超级计算机领域。此外，采用HBM的Sapphire Rapids处理器的小芯片，可能与普通的Sapphire Rapids处理器有所区别。