AMD launches Milan-X with 3D V-Cache
AMD has announced at this year’s Taipei Computer Show that it will launch the Zen 3 Ryzen processor with 3D V-Cache. The CCD used by AMD’s EPYC processor is exactly the same as the CCD on the Ryzen processor. Therefore, it is not surprising that AMD introduced 3D V-Cache into the server field. At the meeting this morning, AMD released the Milan-X processor with 3D V-Cache.
The CCD used by Milan-X is the same as the previously announced Ryzen processor, that is, a layer of 7nm SRAM is stacked on the original CCD of Zen 3. In this way, the three-level cache capacity of each CCD will increase from the original 32MB to 96MB, but it is different from the Ryzen processor which has at most two CCDs. EPYC Milan has up to 8 CCDs, so its three-level cache capacity reaches 768MB, and the total three-level cache capacity on a two-way system is an astonishing 1.5GB.
Milan-X will retain the same CPU socket as the current third-generation EPYC processor. The maximum number of cores is also 64 cores. It supports 8-channel DDR4-3200 memory and provides 128 PCI-E 4.0 channels. AMD says it will launch Milan-X processors in the first quarter of next year.
Of course, AMD did not disclose the specific specifications of the Milan-X, if there are no accidents, the additional L3 cache will bring more power consumption. If the processor’s power limit is maintained at the limit of 280W, this will inevitably come at the cost of lowering the frequency.
In addition, the cost will be higher than that of existing processors. The CCD chip area of Zen 3 is 80.7mm, the area of the 64MB SRAM added is 36mm2. This means that the wafer area has increased by 45%, and the additional cost of 3D packaging has not been accounted for, so the price of Milan-X will definitely be higher than that of Milan.
In terms of performance, after increasing the three-level cache capacity by 3 times, it brings a higher hit rate, which can effectively reduce the bandwidth pressure of the memory. In the same 16-core Milan and Milan-X performance comparison, Synopsys VCS workload efficiency increased by 66%. There is a more detailed performance preview on the Azure HBv3 VM document based on Milan-X released by Microsoft, and the performance of CFD workload is improved by up to 80%. EDA RTL simulation load performance is improved by up to 60%, and performance in explicit finite element analysis workloads is improved by up to 50%. For specific tests, please refer to Microsoft’s documentation.
