EPYC 7003 As much as 64 Cores and 768 MB L3 Cache

There may be been a large number of focal point on how each Intel and AMD are making plans for the longer term in packaging their dies to extend general efficiency and mitigate upper production prices. For AMD, that subsequent step has been V-cache, an extra L3 cache (SRAM) chiplet that is designed to be three-D die stacked on most sensible of an present Zen 3 chiplet, tripling the overall about of L3 cache to be had. As of late, AMD’s V-cache generation is in the end to be had to the broader marketplace, as AMD is saying that their EPYC 7003X “Milan-X” server CPUs have now reached normal availability.

As first introduced past due closing 12 months, AMD is bringing its three-D V-Cache generation to the endeavor marketplace thru Milan-X, a complicated variant of its current-generation third Gen Milan-based EPYC 7003 processors. AMD is launching 4 new processors starting from 16-cores to 64-cores, they all with Zen 3 cores and 768 MB L3 cache by way of three-D stacked V-Cache.

AMD’s Milan-X processors are an upgraded model of its recent third technology Milan-based processors, EPYC 7003. Including to its preexisting Milan-based EPYC 7003 line-up, which we reviewed again in June closing 12 months, essentially the most vital development from Milan-X is thru its massive 768 MB of L3 cache the use of AMD’s three-D V-Cache stacking generation. The AMD three-D V-Cache makes use of TSMC’s N7 procedure node – the similar node Milan’s Zen 3 chiplets are constructed upon – and it measures at 36 mm², with a 64 MiB chip on most sensible of the prevailing 32 MiB discovered at the Zen 3 chiplets. 

That specialize in the important thing specs and applied sciences, the most recent Milan-X AMD EPYC 7003-X processors have 128 to be had PCIe 4.0 lanes that can be used thru full-length PCIe 4.0 slots and controllers variety. That is dependent on how motherboard and server distributors wish to use them. There also are 4 reminiscence controllers which can be in a position to supporting two DIMMs in keeping with controller which permits using eight-channel DDR4 reminiscence.

The entire chip configuration for Milan-X is a big, 9 chiplet MCM, with 8 CCD dies and a big I/O die, and this is going for all the Milan-X SKUs. Significantly, AMD has opted to equip all in their new V-cache EPYC chips with the utmost 768 MB of L3 cache, which in flip method all 8 CCDs will have to be provide, from the highest SKU (EPYC 7773X) to the ground SKU (EPYC 7373X). As a substitute, AMD will likely be various the collection of CPU cores enabled in every CCD. Drilling down, every CCD contains 32 MB of L3 cache, with an extra 64 MB of three-D V-Cache layered on most sensible for a complete of 96 MB of L3 cache in keeping with CCD (8 x 96 = 768).

In relation to reminiscence compatibility, not anything has modified from the former Milan chips. Every EPYC 7003-X chip helps 8 DDR4-3200 reminiscence modules in keeping with socket, with capacities of as much as 4 TB in keeping with chip and eight TB throughout a 2P device. It is value noting that the brand new Milan-X EPYC 7003-X chips proportion the similar SP3 socket as the prevailing line-up and, as such, have compatibility with recent LGA 4094 motherboards thru a firmware replace.

Taking a look on the new EPYC 7003 stack with three-D V-Cache generation, the highest SKU is the EPYC 7773X. It options 64 Zen3 cores with 128 threads has a base frequency of two.2 GHz and a most spice up frequency of three.5 GHz. The EPYC 7573X has 32-cores and 64 threads, with a better base frequency of two.8 GHz and a spice up frequency of as much as 3.6 GHz. Each the EPYC 7773X and 7573X have a base TDP of 280 W, even if AMD specifies that each one 4 EPYC 7003-X chips have a configurable TDP of between 225 and 280 W.

The bottom spec chip within the new line-up is the EPYC 7373X, which has 16 cores with 32 threads, a base frequency of three.05 GHz, and a spice up frequency of three.8 GHz. Shifting up the stack, it additionally has a 24c/48t choice with a base frequency of two.8 GHz and a spice up frequency of as much as 3.7 GHz. Each come with a TDP of 240 W, however like the larger portions, AMD has showed that each 16-core and 24-core fashions can have a configurable TDP of between 225 W and 280 W.

Notable, all of those new Milan-X chips have some more or less clockspeed regression over their common Milan (max core efficiency) opposite numbers. When it comes to the 7773X, that is the bottom clockspeed, whilst the opposite SKUs all drop just a little on each base and spice up clockspeeds. The drop is necessitated by means of the V-cache, which at about 26 billion further transistors for a complete Milan-X configuration, eats into the chips’ energy price range. So with AMD opting to stay TDPs constant, clockspeeds had been dialed down just a little to compensate. As all the time, AMD’s CPUs will run as rapid as warmth and TDP headroom lets in, however the V-cache supplied chips are going to achieve the ones limits just a little quicker.

AMD’s audience for the brand new Milan-X chips is consumers who want to maximize per-core efficiency; particularly, the subset of workloads that take pleasure in the additional cache. That is why the Milan-X chips don’t seem to be changing the EPYC 70F3 chips solely, as now not all workloads are going to answer the additional cache. So each lineups will likely be sharing the highest spot as AMD’s fastest-per-core EPYC SKUs.

For his or her section, AMD is especially pitching the brand new chips on the CAD/CAM marketplace, for duties similar to finite part research and digital design automation. Consistent with the corporate, they have noticed upwards of a 66% building up in RTL verification speeds on Synopsys’ VCS verification device in an apples-to-apples comparability between Milan processors with and with out V-cache. As with different chips that incorporate greater caches, the best advantages are going to be present in workloads that spill out of contemporary-sized caches, however will well have compatibility into the bigger cache. Minimizing dear journeys to primary reminiscence signifies that the CPU cores can stay operating that a lot more continuously.

Microsoft discovered one thing an identical closing 12 months, once they unveiled a public preview of its Azure HBv3 digital machines again in November. On the time, the corporate printed some efficiency figures from its in-house trying out, basically on workloads related to HPC. Evaluating Milan-X at once to Milan, Microsoft used information from each EPYC 7003 and EPYC 7003-X within its HBv3 VM platforms. It is also value noting that the trying out was once carried out on dual-socket techniques, as all the EPYC 7003-X processors introduced lately may well be utilized in each 1P and 2P deployments.

Efficiency information printed by means of Microsoft Azure is encouraging and the use of its in-house trying out, it appears as although the additional L3 cache is enjoying a large section. In Computational Fluid Dynamics, it was once famous that there was once a greater accelerate with fewer parts, in order that must be considered. Microsoft said that with its recent HBv3 sequence, its consumers can be expecting most positive factors of as much as 80% efficiency in Computational Fluid Dynamics in comparison to the former HBv3 VM techniques with Milan.

Wrapping issues up, AMD’s EPYC 7003-X processors are actually typically to be had to the general public. With costs indexed on a 1K unit order foundation, AMD says the EPYC 7773X with 64C/128T will likely be to be had for round $8800, whilst the 32C/64T type, the EPYC 7573X, will price about $5590. Shifting down, the EPYC 7473X with 24C/48T will price $3900, and the access EPYC 7373X with 16C/32T will price relatively extra with a price of $4185.

Given the huge order sizes required, the entire retail worth might be relatively upper for one unit. Despite the fact that with nearly all of AMD’s consumers being server and cloud suppliers, unquestionably AMD can have some consumers purchasing in bulk. A lot of AMD’s primary server OEM companions also are slated to start providing techniques the use of the brand new chips, together with Dell, Supermicro, Lenovo, and HPE.

In any case, customers gets their very own likelihood to get their arms on some AMD V-cache enabled CPUs subsequent month, when AMD’s 2nd V-cache product, the Ryzen 7 5800X3D, is launched. The desktop processor is founded round a unmarried CCD with a whopping 96 MB of L3 cache to be had, all of which contrasts effectively with the a lot larger EPYC chips.