After the world's first 7nm mobile phone processor, the CPU and GPU products are also officially released based on the 7nm process. On November 7, AMD released the world's first 7nm code "Rome" (Roman) of the second generation of the second generation of the "Rome" (Rome), a second-generation EPYC Dragon CPU processor, and the GPU calculation card of Radeon Instinct MI60 / MI50.
7nm Rome second generation epyc 龙 processor
First 7nm CPU and GPU
The AMD code is a second-generation EPYC Waylong processor based on the zen2 architecture. Rome Epyc uses a special design architecture, and each processor has 8 CPU dies, and 8 physical cores are integrated in each die. 64 . At the same time, in order to better coordinate such a multi-CPU core synergy, it also designed an I / O die, placed in the central location, specifically responsible for input and output control, but I / O dies are manufactured with 14nm process. The ROME processor has 8-channel DDR4 memory controller, and each slot supports up to 4TB DRAM.
Rome Epyc is also the first server-level CPU that supports PCIe 4.0 technology, supports 128 PCIe 4.0, and the number of bandwidth channels, which can improve accelerator performance, and match the new acceleration card of PCIe 4.0 technology, AMD represents can bring Unprecedented acceleration performance. Of course, Rome Epyc is also the world's first data center processor using a 7nm process. At present, the more detailed parameters of Rome Epyc have not been announced, and the AMD predicts that Rome EPYC performance will double due to the increase in the number of cores, and floating-point performance will also double the improvement of the IPC architecture and the number of cores.
In terms of compatibility, Rome Epyc is compatible with existing NAPLES, and next-generation MILANs, which simplifies server development based on AMD processor, and allows the server to use existing designs in future servers. In terms of progress, AMD is providing Rome Epyc samples to server manufacturers, plans to launch products in 2019, but the specific time has not been announced. Matt Garman, vice president of Amazon AWS, said that R5 and M5 examples based on AMD Dragon chips can be used today, T3 instances will be online in the next few weeks, and focus on helping customers cost the most common application costs. Features .
GPU, AMD announced the 7nm Radeon Vega GPU on ComputEx 2018 this year, but the first 7nm GPU product, which is officially released is not Radeon RX game graphics card, but the Radeon Instinct Calculation Card. AMD also released Radeon Instinct Mi60 / Mi50, all based on the 7nm process VEGA architecture, benefited from the 7nm process, the Radeon Instinct compute card integrates 13.2 billion transistors, which increased by 6.4% more than 12.5 billion transistors of 14nm VEGA, core area It is 331 square millimeters, which is 31.6% compared to the current 484 square millimeter. It is said that under the same power consumption, the new core performance has increased by more than 25%, and the new core power consumption is reduced by 50% under the same frequency.
Radeon instinct memory is matched with up to 32GB HBM2, with a bandwidth of 1TB / s, built-in ECC error correction. Similarly, 7nm Vega is also the world's first GPU supporting PCIe 4.0 technology, bidirectional bandwidth 64GB / s, with a bandwidth up to 100GB / s bus, support 4 channels of GPU parallel, double performance increase 99%, 4 road ratio The road is upgraded 298%, and the 8th road increased by 664%.
AMD 7nm Radeon Instinct Calculation Card
On acceleration, the 7nm VEGA computing card is different from VEGA10 in several key fields, and 7 nm VEGA supports semi-rate double precision while supporting new low-precision data types. Specifically, the MI60 FP64 double precision floating point is 7.4TFLOPS, the FP32 single-precision is up to 14.7TFLOPS, and the int integer is up to 118TOPS, which is 8.8 times compared to the same VEGA 10 core MI25 professional performance.
It is also worth noting that the new GPU also includes a pair of external INFINITY FABRIC links, allowing the Radeon Instinct card to connect directly to each other through a coherent link. Since each GPU is only 2 links, the Topology option of AMD is limited to changes in the ring, which means that the 4-way configuration GPU cannot communicate directly. In addition, AMD still insists that the PCIe card is used, and there is no custom laminated card like NVIDIA, so these cards need to be connected through the top bridge.
In order to make up for the shortcomings of AMD in software, AMD announced an open source computing platform ROCM 2.0 and continued to fully open source. Scales, Radeon Instinct Mi50 will be listed in the fourth quarter of this year, and Mi60 will be listed in the first quarter of next year.
7nm victory?
7nm undoubtedly became the focus of industry concerns, but AMD President and Chief Executive Officer Su Zifeng had previously discussed the 7nm progress of AMD many times. At this next horizon meeting, AMD CTO Mark Papermaster said that AMD gambled in 7nm, first of all, believe that 7nm can be made, followed by 7nm will become an industrial critical node and last for a long time. AMD initial goal is to compete with its own 7nm and Intel 10nm, at least to set up, but I didn't expect AMD to make 7nm first, and will become the first company in the new node.
In addition to publishing a 7nm architecture and product, AMD also said that the sharp dragon, Xiaolong will turn the entire line to the Zen2 architecture of the 7nm process, which will be a new architecture of the 7nm + upgrade process. ZEN3. More than this, AMD also disclosed the Zen4 architecture in design. Papermaster also emphasizes the importance of AMD roadmaps and the ability and determination of the road map. He said that the Zen2 architecture chip is being flowing, ZEN3 is expected to use 7nm + process in 2020, Zen4 development is close to the end. Does ZEN4 have the opportunity to use 5nm?
Beyond Intel and British Weida in one fell swoop?
When the NEXT HORZON meeting was released by 7nm Epyc, AMD rendered 1080p / quadHD / 4K image comparison with two Skylake Xeon 8180M (56 core), and the results were 28.1 seconds and 30.2 seconds, respectively. AMD is more winning. GPU, AMD has a 1TB / S full-memory bandwidth, and even more than NVIDIA flagship products GV100 GPU. Does this mean that AMD surpasses Intel and British Weida with 7nm?
A foreign friend MockingBird broke the newspaper in the foreign forum HardoCP, the RTG (Radeon Technology Group) has received the first ZEN 2 architecture 7nm processor sample, but the basic / Rui frequency frequency of test samples is 4.0 / 4.5GHz, compared to current Ryzen 7 2700X (3.7 / 4.1gHz) is high, but there is a gap between the Rui frequency frequencies with Intel 8700K 4.7GHz, less than the next 9800k / 9900K. Although the AMD continues to polish the frequency space, it is not large from the frequency of the initial Ryzen test sample and the formal release product.
In addition, hardware and 7nm do not mean the success of the processor, and the instruction set is also critical. It is understood that AMD first has a SSE5 instruction set, and Intel launches the AVX instruction set, then AMD is equivalent to AVX, but retains the FMA4, CVT16 and XOP instructions in SSE5; then Intel follows the AMD to add CVT16, but renamed F16C, then Add FMA4 to the FMA3 and AVX2.0 instructions after the FMA4; AMD has to be with wind, add FMA3, and recognize F16C, and add AVX2.0 in the future. It can be seen that AMD is very competitive in the instruction set.
As for the GPU field, the 7nm computing card from AMD can see that it is very important to the server market, but in this market, the British Weida GPU has absolute advantages, like the competition in the CPU, like Intel's competition, the leader of Ying Weida has the ability to The architecture and instruction sets are dominant. Once the 7nm process is used, how long the AMD can be kept in some performance advantages. We can also see from the architecture. In August this year, Ying Weida released the largest Zuling architecture in 12 years, but the RTX2080 GPU of the Tuling architecture has inherited a lot of Fermi architecture, which means that AMD is not It is too likely to abandon the new architecture of the GCN architecture, and the AMD's R & D center is placed on the Zen architecture. The GPU improves weakness. Based on the current situation, AMD can achieve the anti-Overda of Yingda with new processes?
We believe that AMD is very rare as a company with CPU and GPU products in high performance calculations, but unfortunately, it is difficult to surpass Intel in the CPU field, and it is difficult to surpass your Britain in the GPU. This has technical factors, and there are also commercial aspects. reason. Of course, for consumers and cloud manufacturers, only competition will bring better products, so we still remain expected for AMD's 7nm CPU and GPU.
Original address: https://www.eeboard.com/news/amd-7nm-cpu-gpu/
Search for the panel network, pay attention, daily update development board, intelligent hardware, open source hardware, activity and other information can make you master. Recommended attention!
[WeChat scanning picture can be directly paid]
Our other product:
