AMD announced today the release of three new 2nd-gen EPYC server CPUs. The primary focus of these new CPUs are to push base frequencies even higher, yielding significant gains in single-threaded performance, even claiming highest per-core record in SPECrate®2017_fp_base using the new EPYC 7F32 8-core CPU, beating their own EPYC 7262 by just over 10%. These CPUs are compatible with existing 2nd-gen EPYC server solutions and should see rapid availability in vendor solutions.
The New CPUs
CPU Model # | 1Ku Pricing (USD) | Cores | Threads | Base Freq | Max Boost | Default TDP | L3 Cache |
EPYC 7F32 | $2100 | 8 | 16 | 3.7 GHz | ~3.9 GHz | 180W | 128MB |
EPYC 7F52 | $3100 | 16 | 32 | 3.5 GHz | ~3.9 GHz | 240W | 256MB |
EPYC 7F72 | $2450 | 24 | 48 | 3.2 GHz | ~3.7 GHz | 240W | 192MB |
The EPYC 7F32 looks to directly upgrade the EPYC 7262 by raising the base frequency from 3.2 GHz to 3.7 GHz while also adding 500 MHz to the max boost as well. It consequentially raises the TDP from 155W to 180W, while maintaining the same 128MB of cache on presumably 4 chiplets. This gives 16MB of L3 cache per core. AMD asks quite the premium on the CPU, listing 1Ku pricing at $2100 instead of the lesser model’s $575.
Next, the EPYC 7F52 is the CPU I was expecting to see since the initial lineup was revealed. It’s a full 8 chiplets brining 256MB of cache, but with (once again, presumably) only one core per CCX (thus 2 cores per CCD) left enabled. This core configuration is the same as seen in the EYPC 7F32, but this time with all 8 chiplets rather than just 4. Naturally, the TDP was raised to 240W, which is higher than the previous max EPYC TDP of 225W. This allows for a 3.5 GHz base frequency and a still receives a ~3.9 GHz boost rating. Asking $3100 for this 16-core CPU is a fair jump from the $978 for the previous top 16-core, the EPYC 7302.
Lastly, the EPYC 7F72 basically halves the core-count of the EPYC 7552, brining (presumably) 6 chiplets to achieve the 192MB of L3 cache. This time, two cores per CCX, and thus four per CCD, leaving only 8MB of L3 cache per core this time. This replaces the 16-core EPYC 7302 as the highest core count EPYC with 8MB of L3 cache per core. The EPYC 7402 is the nearest equivalent to the 7F72, and at $1783, the 7F72 is only a moderate cost bump at $2450 to gain the extra L3 cache (up from 128MB), base frequency (3.2 GHz instead of 2.8 GHz), and max boost (~3.7 GHz up from 3.35 GHz) at the cost of an extra 60W.
The 8- and 16-core CPUs are great options for databases and HPC that demand real-time compute that higher clockspeeds can provide. This paired with the great IPC and now a huge per-core L3 cache, these EPYC CPUs will be powerhouses for these specialized workloads.
I see the EPYC 7F72 being a particularly useful all-around performer for similar workloads, but also hyperconverged systems that still demand more L3 cache per core, but also more cores overall and increased clockspeeds due to more single-threaded IPC needs.
Ecosystem Support
Dell Technologies is supporting all three CPUs across its entire EPYC-based Dell EMC PowerEdge server lineup. HPE will also support all three CPUs in their Apollo 2000 Gen10 Plus, HPE ProLiant DL385 Gen10 Plus, and Proliant DX server ranges. Lenovo is also supporting them in their ThinkSystem SR635 and SR655 platforms. Supermicro announced immediate support for the EYPC 7Fx2 CPUs across all Supermicro A+ platforms: including Ultra, GPU, WIO, Twin and Mainstream systems.
On the software side of things, IBM Cloud announced as the first cloud provider to offer the new CPUs to clients in their bare metal offering, specifically only the EPYC 7F72. Microsoft is offering all the 7Fx2 CPU options for their data platform customers, claiming 17% higher SQL Server TPM per core performance. Nutanix, in conjunction with HPE, announced expected Nutanix HCI support by May. VMware has, critically, also added support for these new EPYC 7Fx2 CPUs.
The full press release can be viewed on AMD’s website.