AMD is an organization that’s actually no stranger to constructing semi-custom processors. Certainly, the semi-custom unit of the corporate principally saved it afloat for a few years earlier than the launch of Ryzen by supplying SoCs to Sony and Microsoft for his or her Eighth-generation recreation consoles. We have not seen a lot semi-custom work within the datacenter, although—at the least, not till now.
This announcement comes not from AMD, however in actual fact from Microsoft, who’s rather more fascinated by telling you about its Azure HBv5 Digital Machines—cloud computing you may hire with some fairly distinctive capabilities. We, in fact, are rather more within the {hardware} behind these capabilities: particularly, what Microsoft describes as a {custom} 4th-generation EPYC chip with high-bandwidth reminiscence (HBM).
In case you did not know, HBM is not merely a categorical descriptor, however slightly a selected reminiscence expertise. It achieves extraordinarily excessive reminiscence switch charges by utilizing an absurdly-wide reminiscence interface. Traditionally, HBM ran comparatively low switch charges, however trendy HBM3e has pushed the switch charge as much as be aggressive with desktop DDR5 reminiscence. Contemplating that the interface is usually at the least eight occasions wider than a regualr desktop CPU’s reminiscence interface, you are actually speaking about some excessive bandwidth.
Microsoft says that these chips are based mostly on AMD’s Zen 4 structure and run with SMT disabled for optimum single-threaded throughput. They clock as excessive as 4 GHz, which is not spectacular to desktop gamer varieties, however not dangerous for an EPYC machine with as much as 352 CPU cores per server. These cores are served by “400-450 GB” of HBM3 reminiscence, and Microsoft says these chips additionally sport double the Infinity Cloth bandwidth between CPUs in comparison with typical EPYC servers.
The height bandwidth achieved on this configuration is a really good 6.9 TB/second, as measured within the STREAM Triad benchmark. This sort of reminiscence bandwidth is sort of unprecedented, at the least for CPUs; folks within the know will recall that Intel additionally shipped server CPUs with HBM onboard as a part of its Knowledge Heart MAX Xeon line.
Reminiscence bandwidth per core has been a sore spot for high-performance computing for some time. We have reported on this matter a number of occasions earlier than, together with when JEDEC proposed its MRDIMM customary for Multi-Rank Buffered DIMMs that promise double the reminiscence bandwidth of normal DIMMs. Which will assist ease issues as core counts proceed to rise, however bandwidth-per-core continues to be comparatively low in such a configuration. Making use of HBM3 neatly places a bow on that drawback.
It is seemingly that, simply because the Steam Deck’s “Aerith” processor was seemingly not created with the Steam Deck in thoughts, these EPYC chips in use by Microsoft had been most likely not fabbed with cloud-based HPC in thoughts, both. There have been rumors up to now that AMD would introduce Intuition accelerators that dropped the CDNA GPU for Zen CPU CCDs, and that is precisely what this seems to be. These merchandise by no means got here to common availability, however it appears they did discover their technique to market a method or one other.
