AMD’s New EPYC 7H12: A Specially-Binned HPC Processor

The impact of AMD’s second-generation EPYC processors is starting to surface. At its European launch in Rome, Italy, the company highlighted a series of design wins. Among those wins is TSMC which will be leveraging AMD’s processors for next-generation research and leading process technology. Dell is expanding its PowerEdge platform with five new design wins intended to take full advantage of 2nd generation EPYC features such as PCIe 4.0.

EPYC 7H12

Pushing up the stack, AMD also unveiled a new processor – the EPYC 7H12. This processor is similar to the EPYC 7742, having 64 cores and 128 threads and comes with 256 MiB of L3 cache. However, the 7H12 is specifically geared towards HPC customers and workloads. AMD bumped up the TDP from 225 Watts to 280 W. With 280 W, those chips are intended to be liquid-cooled. For 25% more TDP, AMD is pushing the base frequency from 2.25 GHz to 2.6 GHz – or roughly 16%. That’s 350 MHz higher for each of the 64 cores. The maximum turbo frequency is 3.3 GHz.

Intel Has A Problem

The EPYC 7H12 is exactly the kind of SKU AMD needed to take on Intel in the HPC market. The EPYC 7742 was already very competitive in the HPC space, but the 7H12 really takes this a step further. Intel current lineup is unable to compete well. Up until recently, their flagship processor was the Xeon Platinum 8280. This is a $10K processor with 28 cores and a base of 2.7 GHz. Intel has since upgraded their flagship chip to the Xeon Platinum 8284. With the 8284 Intel increased the base frequency to 3 GHz and the TDP to 240 W. For those benefits Intel feels charging an additional $5,000 makes sense. We can almost guarantee the 7H12 won’t cost anywhere near that prices.

The only thing the Intel SKUs have going for them is the single-core performance which still a hair better. And this is very marginal at this point. The max turbo boost on the 8284 is 4 GHz, but that’s only true for just two active cores. With three or more busy cores, the frequency drops to 3.8 GHz and with five or more cores it’s already down to 3.7 GHz. Given how favorably the Zen 2 core compares to Cascade Lake, there is a lot more parity. And yes, there is AVX-512, but since AMD doubled the AVX2 throughput on Zen 2 and since there is no major throttling like Intel experiences with AVX512, this isn’t proving to be an exceptional selling point .

Intel argues that their new Xeon Platinum 9200 series is what we should be comparing against but that’s a really hard buy. The top Cascade Lake AP SKUs are the 9242 and 9282 with 48 and 56 cores respectively. Other than costing well over $20K, their base frequencies barely manage to match the 7H12 at the cost of 70-120 Watt higher TDP. It’s also worth noting that the 7H12 is socket compatible with all other Rome processors so you get the full 128 lanes of PCIe Gen 4.0 lanes (or 160 in a dual-socket configuration). This is unlike those Xeon 9200 series processors which only come in Intel’s prebuilt S9200WK compute module preventing OEMs from designing their own solutions. By the way, the S9200WK only exposes 40 PCIe Gen 3.0 lanes and that’s all you have to work with. We do recognize that the 9200-series parts have 12 memory channels which are a huge plus for HPC workloads, but since AMD supports higher data rates (3200 MT/s vs 2933 MT/s), the 50% more memory channels translate to 37.5% higher peak bandwidth. And that number is split between two dies whereas the 7H12 has unified 8 channels of DDR4 memory. Also, EPYC’s 256 MiB L3 somewhat compensates for this. And then there is the cost, single-solution option, and power which are all really big minuses compared to the 7H12.

Currently, Intel is gearing up Cooper Lake and Ice Lake server parts for next year which might improve their competitiveness. But for now, they have a problem. It’s worth noting that at 280 W, one might ask how high AMD can go with a 300 W SKU?