AMD Prepping 400W Zen 4c Server CPUs with 128 Cores/256 Threads for 2023, 256 Core/512 Core Zen 5 “Turin” in Q4 2023 [Report]

AMD Prepping 400W Zen 4c Server CPUs with 128 Cores/256 Threads for 2023, 256 Core/512 Core Zen 5 "Turin" in Q4 2023 [Report]

AMD is doubling its efforts in the server segment with three product launches planned for the next three quarters. Tom from MLID has made a detailed video on this. We already know about the Zen 4-powered Genoa lineup which will rock up to 96 cores across 12 CCDs in the final quarter of 2022. This will be followed by Bergamo which is a low-power, high-core count version of Genoa. It’ll feature Zen 4c cores which are a cut-down variant of Zen 4 with reduced L3 cache and lower boost clocks.

Bergamo will come with 16-core CCDs with 16MB of L3 cache for a total of 128 cores and 256 threads. Unlike Intel’s low-power design SMT will be supported. AVX-512, on the other hand, will largely be limited to Genoa. The TDP will approach the 400W mark, a heavy increase over Milan’s 280W ceiling. Bergamo is slated for an early 2023 launch. Both the server architectures will feature SDCI (Smart Data Cache Injection) on the I/O die to improve the cache hits from external storage devices. SDXI (Smart Data Acceleration Interface) will be another addition to the I/O die to reduce the overhead on the CPU cores by taking over inter-device copy/move operations.

We’ll have to wait just a bit over a year for the next iteration of the Zen architecture. According to Tom, Zen 5 is expected as early as Q4 2023 with Turin. These Epyc CPUs will cram up to 256 cores (256 threads) along with multiple accelerator chiplets such as CDNA, Xilinx FPGA, cache, etc. The DMA Engine will be upgraded and the increased core counts will push the TDP to 600W. We can easily expect HBM variants which along with the newly unveiled 3D V-Cache technology offer monumental performance gains in certain workloads. Overall, the future of computing will be one ruled by heterogeneous designs with advanced packing designs rather than process nodes.

Leave a Reply

Your email address will not be published.

Main Menu