AMD在2022年第四季度财报电话会议上，再次确认了基于Zen 4c架构、代号Bergamo的EPYC服务器处理器将会在2023年上半年推出，不过Instinct MI300要等到2023年下半年。
Bergamo针对计算密集应用进行了优化，删除了某些不需要的功能并提高了密度，最多拥有128个核心，竞争对手除了英特尔的同类产品，还有各种针对云服务的Arm处理器。相比于普通的EPYC服务器处理器，Bergamo有着更高的电源效率和每插槽性能，并与Genoa/Genoa-X采用相同的插座，也会支持PCIe 5.0、DDR5和CXL 1.1。
var tableid = settableid();
document.write('');
Table().init({
title: "AMD EPYC服务器处理器系列参数",
header: ["型号","EPYC 7001系列","EPYC 7002系列","EPYC 7003系列","EPYC 9000系列","EPYC 9000系列?","TBC"],
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","Naples","Rome","Milan/-X","Genoa/-X","Bergamo","Turin"],
["插座","SP3(LGA4094)","SP3(LGA4094)","SP3(LGA4094)","SP5(LGA6096)SP6","SP5(LGA6096)SP6","SP5(LGA6096)SP6"],
["架构","Zen","Zen 2","Zen 3","Zen 4","Zen 4c","Zen 5"],
["制造工艺","14nm GF","7nm TSMC","7nm TSMC","5nm TSMC","5nm TSMC","TBC"],
["最大L3缓存","64MB","128MB","256MB / 768MB","384MB / TBC","TBC","TBC"],
["芯片结构","4 CCD","8 CCD+1 IOD","8 CCD+1 IOD","12 CCD+1 IOD","8 CCD+1 IOD?","TBC"],
["内存类型","DDR4 ECC","DDR4 ECC","DDR4 ECC","DDR5 ECC","DDR5 ECC","DDR5 ECC"],
["内存通道数","8","8","8","12","12","12"],
["内存频率","2666MHz","3200MHz","3200MHz","5200MHz","5200MHz","6000MHz?"],
["PCIe版本","3.0","4.0","4.0","5.0","5.0","TBC"],
["PCIe通道数","128","128","128","128","TBC","TBC"],
["最大核心/线程数","32C/64T","64C/128T","64C/128T","96C/192T","128C/256T","256C/512T?"],
["最大功耗","200W","280W","280W","400W(SP5)225W(SP6)","400W(SP5)225W(SP6)","600W(SP5)?225W(SP6)?"],
["推出时间","2017年","2019年","2021年","2022年","2023年?","2023/2024年?"],
]
});Instinct MI300是AMD面向数据中心的下一代APU加速卡产品，采用了Chiplet设计，拥有13个小芯片，通过3D堆叠的方式，包括了24个Zen 4架构CPU内核、基于CDNA 3架构的GPU模块、以及128GB的HBM3显存，共1460亿个晶体管。其中的小芯片以5nm和6nm工艺制造，其中5nm的有9块，6nm的有4块。据推测，Instinct MI300的9个计算模块里，有3个属于CPU，6个属于GPU。这是AMD迄今为止最大的芯片，超过了英特尔1000亿个晶体管的Ponte Vecchio。
AMD在CES 2023上并没有透露太多Instinct MI300的细节，只是简单地与Instinct MI250X做了部分性能比较，AI性能是后者的8倍，每瓦性能则是后者的5倍。美国劳伦斯利弗莫尔国家实验室（LLNL）的ExaFLOP级El Capitan超级计算机计划在2023年末部署，将会采用Instinct MI300。
预览var tableid = settableid();
document.write('');
Table().init({
title: "AMD Instinct系列计算卡规格参数对比",
header: ["型号", "Instinct MI60", "Instinct MI100", "Instinct MI250X", "Instinct MI300"],
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: [
["GPU", "Vega 20", "Arcturus", "Aldebaran(MCM)", "?(MCM)"],
["架构", "GCN5", "CDNA", "CDNA2", "Zen 4+CDNA3"],
["工艺", "7nm", "7nm", "6nm", "5nm+6nm"],
["计算模块数量", "1", "1", "2", "3 CPU+6 GPU"],
["CU数量", "64", "120", "110x2", "-"],
["GPU频率", "1800 MHz", "1500 MHz", "1700 MHz", "-"],
["FP16算力", "29.5 TFLOPS", "185 TFLOPS", "383 TFLOPs", "-"],
["FP32算力", "14.7 TFLOPS", "23.1 TFLOPS", "47.9 TFLOPs", "-"],
["FP64算力", "7.4 TFLOPS", "11.5 TFLOPS", "47.9 TFLOPs", "-"],
["显存容量", "32GB", "32GB", "128GB", "128GB"],
["显存类型", "HBM2", "HBM2", "HBM2e", "HBM3"],
["显存速率", "2.0 Gbps", "2.4 Gbps", "3.2 Gbps", "-"],
["显存位宽", "4096-bit", "4096-bit", "8192-bit", "8192-bit"],
["显存带宽", "1 TB/s", "1.23 TB/s", "3.2 TB/s", "-"],
["板卡规格", "双槽/全尺寸", "双槽/全尺寸", "OAM", "-"],
["TDP", "300W", "300W", "560W", "600W+?"]
]
});
