AMD introduced the sixth generation EPYC 9006 processors, that is, the EPYC Venice family, with Zen 6 and Zen 6c architectures. A single processor has up to 256 cores and memory bandwidth reaches 1.6 terabytes per second. However, the 27 percent IPC increase claimed by the company does not mean that the single core is accelerated by 27 percent.
Difference between Zen 6 and Zen 6c at EPYC Venice
First of all, AMD offers two designs. Standard Zen 6 focuses on high performance and operating frequency. Zen 6c fits more cores into a smaller space. Moreover, they both use the same instruction set and can run two threads per core.
In standard Zen 6, a CCD (the unit that brings the cores together) carries 12 cores instead of 8. Additionally, the shared L3 cache increases from 32 MB to 48 MB. Thus, the number of cores and cache memory increases by 50 percent compared to the previous generation.
Zen 6c has 32 cores per CCD; In Zen 5c this number was 16. L3 cache increases from 32 MB to 128 MB. Therefore, there is 4 MB L3 per core in both designs. In addition, each core has 1 MB L2 cache.
How was the 27 percent IPC increase measured?
IPC means the amount of processing done by a core in a single cycle. Accordingly, Zen 6c gives an average of 27 percent higher IPC than Zen 5 in certain server workloads. Testing was done between 16 Zen 6c cores at the same frequency and 16 Zen 5 cores spread across two CCDs.
The increase was 20 percent in enterprise and cloud workloads and 31 percent in artificial intelligence workloads. As a matter of fact, 27 percent is the approximate average of these two results.
For example, the reorder buffer increased from 448 to 640 entries.
Save units increased from 240 to 352.
Instruction transmission increased from 12 to 18 per clock cycle.
The upload queue grew from 64 to 80 and the write queue grew from 96 to 128.
On the other hand, AVX-512 FP16 support comes on the command side. Improved Bit Matrix Multiply commands and VNNI-INT8 units; In fact, INT8 pipelines increased from two to four.
SP7 and SP8 platforms
EPYC Venice comes on two platforms, SP7 and SP8. SP7 initially includes support for up to eight CCDs. Standard Zen 6 models reach 96 cores and 384 MB L3, while Zen 6c models reach 256 cores and 1 GB L3. However, the frequency increases to 5.3 GHz in the standard Zen 6 and 4.1 GHz in Zen 6c.
On the other hand, in the more compact SP8, Zen 6c models go up to 128 cores with four CCDs, and Zen 6 models go up to 96 cores with eight CCDs. At the same time, SP7 uses 16 memory channels, SP8 uses eight. Power limit is 600 W on SP7, 400 W on SP8.
Memory and connectivity
The processors support DDR5 MRDIMM Gen2, DDR5 RDIMM, LPDDR5X SOCAMM2 and CXL 3.1. Especially with MRDIMM Gen2, a speed of 12,800 MT/s is achieved; In SP7, bandwidth is 1.6 TB per second. The platform also has 3D V-Cache support.
On the other hand, standard RDIMM speeds up to 8,000 MT/s. SP7 supports 1 TB/s bandwidth and up to 8 TB memory per socket. CXL 3.1 provides additional capacity beyond memory directly attached to the processor. Moreover, AMD says this memory tier can reduce the cost per GB by up to 87 percent compared to DDR5.
On the other hand, on the PCIe 6.0 side, SP7 offers 128 lines in a two-processor system and 96 lines in a single processor. In SP8 these numbers are 192 and 128.
AMD's performance claims
Accordingly, the sixth generation EPYC is up to 20 times faster in general-purpose computing, up to 18.6 times faster in scientific computing, and up to 11 times faster in high-performance computing compared to the first generation. At the same time, memory bandwidth increased by 9.6 times in six generations.
The 256-core EPYC Venice 9996 outperforms Intel's 128-core Xeon 6980P processor by an average of 2.2x in some tests, and 3.2x at best. However, the 96-core 9686F provides a 1.6x advantage in certain workloads. The average difference in models with similar core numbers is 1.9 times.
In AI tests, Venice is up to 4.4x faster than Intel Granite Rapids. On the other hand, a 2.9x advantage in threads per watt and a 2x advantage in performance per watt is claimed over NVIDIA Vera.
For example, the three-day weather forecast calculation takes 2.66 hours on the 256-core Venice; It takes 5.22 hours on the 128-core Turin. In fact, a 100 nanosecond drug discovery simulation takes 57 days in Venice and 182 days in Granite Rapids.
However, most of these numbers are based on specific workloads and AMD's own testing. EPYC Venice is now generally available. Finally, when Zen 6 will come to desktop Ryzen processors and its performance values have not been announced yet.
