AMD Unveils EPYC 'Milan' 7003 CPUs, Zen 3 Comes to 64-Core Server Chips
AMD unveiled its EPYC 7003 'Milan' processors today, claiming that the chips, which bring the company's powerful Zen 3 architecture to the server market for the first time, take the lead as the world's fastest server processor with its flagship 64-core 128-thread EPYC 7763. Like the rest of the Milan lineup, this chip comes fabbed on the 7nm process and is drop-in compatible with existing servers. AMD claims it brings up to twice the performance of Intel's competing Xeon Cascade Lake Refresh chips in HPC, Cloud, and enterprise workloads, all while offering a vastly better price-to-performance ratio.
Milan's agility lies in the Zen 3 architecture and its chiplet-based design. This microarchitecture brings many of the same benefits that we've seen with AMD's Ryzen 5000 series chips that dominate the desktop PC market, like a 19% increase in IPC and a larger unified L3 cache. Those attributes, among others, help improve AMD's standing against Intel's venerable Xeon lineup in key areas, like single-threaded work, and offer a more refined performance profile across a broader spate of applications.
One interesting new SKU is the EPYC 7663, a 56-core, 112-thread CPU with 7 working cores on each of the 8-core chiplets. There is also a 28-core EPYC 7453.
Next up, Zen 4 "Genoa".
Also at AnandTech, The Next Platform, Phoronix, and Ars Technica.
See also: The Tour of Italy with EPYC Milan: Interview with AMD's Forrest Norrod
AMD video announcement (51m4s) and recap (10m43s)
Related Stories
AMD has announced its "Milan-X" Epyc CPUs, which reuse the same Zen 3 chiplets found in "Milan" Epyc CPUs with up to 64 cores, but with triple the L3 cache using stacked "3D V-Cache" technology designed in partnership with TSMC. This means that some Epyc CPUs will go from having 256 MiB of L3 cache to a whopping 768 MiB (804 MiB of cache when including L1 and L2 cache). 2-socket servers using Milan-X can have over 1.5 gigabytes of L3 cache. The huge amount of additional cache results in average performance gains in "targeted workloads" of around 50% according to AMD. Microsoft found an 80% improvement in some workloads (e.g. computational fluid dynamics) due to the increase in effective memory bandwidth.
AMD's next-generation of Instinct high-performance computing GPUs will use a multi-chip module (MCM) design, essentially chiplets for GPUs. The Instinct MI250X includes two "CDNA 2" dies for a total of 220 compute units, compared to 120 compute units for the previous MI100 monolithic GPU. Performance is roughly doubled (FP32 Vector/Matrix, FP16 Matrix, INT8 Matrix), quadrupled (FP64 Vector), or octupled (FP64 Matrix). VRAM has been quadrupled to 128 GB of High Bandwidth Memory. Power consumption of the world's first MCM GPU will be high, as it has a 560 Watt TDP.
The Frontier exascale supercomputer will use both Epyc CPUs and Instinct MI200 GPUs.
AMD officially confirmed that upcoming Zen 4 "Genoa" Epyc CPUs made on a TSMC "5nm" node will have up to 96 cores. AMD also announced "Bergamo", a 128-core "Zen 4c" Epyc variant, with the 'c' indicating "cloud-optimized". This is a denser, more power-efficient version of Zen 4 with a smaller cache. According to a recent leak, Zen 4c chiplets will have 16 cores instead of 8, will retain hyperthreading, and will be used in future Zen 5 Ryzen desktop CPUs as AMD's answer to Intel's Alder Lake heterogeneous ("big.LITTLE") x86 microarchitecture.
Also at Tom's Hardware (Milan-X).
Previously: AMD Reveals 'Instinct' for Machine Intelligence
AMD Launches "Milan" Epyc Server CPUs, with Zen 3 and up to 64 Cores
AMD at Computex 2021: 5000G APUs, 6000M Mobile GPUs, FidelityFX Super Resolution, and 3D Chiplets
AMD Unveils New Ryzen V-Cache Details at HotChips 33
AMD Aims to Increase Energy Efficiency of Epyc CPUs and Instinct AI Accelerators 30x by 2025
(Score: 1, Funny) by Anonymous Coward on Monday March 15 2021, @11:58PM (2 children)
Epic, eh.
(Score: 0) by Anonymous Coward on Tuesday March 16 2021, @01:56AM
Sheesh, what a beatdown. Intel are screwed !
(Score: 0) by Anonymous Coward on Tuesday March 16 2021, @03:58AM
And likely the last post of the "Ides of March".
(Score: 1, Informative) by Anonymous Coward on Tuesday March 16 2021, @12:10AM (1 child)
Like helping your wallet by not being available months after launch? Lisa Su = Queen of Vapor(hard)ware
(Score: 0) by Anonymous Coward on Tuesday March 16 2021, @02:16AM
Be interesting to see the yield / binning stats for some of these APU's.
(Score: -1, Offtopic) by Anonymous Coward on Tuesday March 16 2021, @01:47AM
Fuck the lot of yous.
(Score: 0) by Anonymous Coward on Tuesday March 16 2021, @04:57AM (2 children)
Linus Tech Tips posted a video the other day speculating that Threadripper might be going away in favor of just making workstation-oriented Epyc boards, but that would be a bummer since enthusiasts and workstation users don't normally need so many memory channels or PCIe lanes but do need higher clocks. And you'd have to pay twice as much for one of these Epyc CPUs that could match the performance of a 5950x consumer CPU even in a highly threaded workstation task (and you couldn't match or even come close to Ryzen for gaming). Hopefully the speculation is wrong and a Zen 3 Threadripper will be coming out in a month or so.
Still, this is good news for people who buy CPUs for datacenters - these are much cheaper and much better performing than Intel's offerings. Intel is risking losing the server market to ARM, just like Sunacle, HP, and IBM lost it to Intel. Server programs don't need PC compatibility all that much. Most of them are writing applications that run in some kind of VM (whether that's Java or Python or whatever) and increasingly the cloud customers don't even know what operating system or hardware they're even using as deployment is increasingly toward interpreted languages in provider-managed containers. AMD is at least making a serious effort toward competing on value.
(Score: 0) by Anonymous Coward on Tuesday March 16 2021, @07:43AM
As you add cores, ram becomes more and more of a bottleneck. I have a 2990wx (32c/64t) and 128 GB. I run out of ram all the time and can rarely use all the cores since each thread can only use 2 GB each.
(Score: 2) by takyon on Tuesday March 16 2021, @09:13AM
This guy? [reddit.com]
Threadripper is interesting. AMD is whipping Intel in the HEDT space (with Threadripper as well as the "prosumer" 5950X). TR is similar to Epyc, but makes AMD less money than Epyc. So it's less of a priority.
I think AMD might have promised to do more than one release on the sTRX4 socket [wikipedia.org], which would support a Zen 3 TR. "Genesis Peak" is the rumored codename and it may include the re-introduction of a 16-core part [notebookcheck.net].
You're *likely* going to see a Zen 4 successor to the 5950X with 24 cores and a different memory situation due to the switch to DDR5 (partial ECC enabled by default on-die, possibly more memory channels). That is not a replacement for Threadripper's capabilities but might make more sense for some people.
If AMD did decide to replace TR with Epyc, they could probably disable some features and price it accordingly (the $3990 64-core TR 3990X was already more than 50% the price of a 64-core Epyc, and the TR PRO 3995WX is even worse at $5490).
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 0) by Anonymous Coward on Tuesday March 16 2021, @05:12AM (1 child)
Looks like 7nm is getting mature enough that AMD is shipping a lot of silicon with seven working cores as six-core CPUs. It makes sense that AMD would want to refine their binning. The 5600 was available in quantity before the 5800 was, and I've seen 5900s in stock (once) at Microcenter but have never seen a 5950. If seven-core silicon is coming out of the factory, AMD is leaving money on the table selling it with only six enabled.
(Score: 2) by takyon on Tuesday March 16 2021, @09:38AM
I think the move to the unified 8-core core complex (CCX) is what makes the 7-core chiplet possible, or at least make more sense. They didn't want irregular amounts of cores on Zen 2 CCXs, or it was not possible to disable a core on a CCX without disabling a core on the other one. Each of the 7 cores can access the full 32 MB of L3 cache.
TSMC "7nm" was already mature to the point that many perfectly good 8-core chiplets have been neutered into 6-cores. There is already slight variance in performance of chips due to the silicon lottery, so a chiplet used for a 6-core might be tolerated in an 8-core.
Zen 4 is going to be one to watch for the binning. It's assumed that it will increase core counts by 50% on at least some product lines (no "mainstream" 24-core would be a disappointment). They can do that by adding more 8-core chiplets or switching to a 12-core chiplet. A recent leak points to 8-core chiplets [notebookcheck.net] again, but 12-core would probably be better for consumers since it could force 6-core out of the picture, improve latency across 12 cores on a single chiplet (for workloads or games that can use 12 cores), enable a cheap single chiplet 10-core, etc.
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 3, Interesting) by bzipitidoo on Tuesday March 16 2021, @05:13PM (4 children)
Dang, I'm still running 2 core, 2 threads per core hardware with 8G or even just 4G RAM. Has the gap between high end server stuff and low end consumer laptop/tablet grown larger? My experience is that a 5 year old server is about equivalent in performance to a mid-range new consumer grade personal computer, while using twice as much or more power.
(Score: 2) by takyon on Tuesday March 16 2021, @06:10PM (3 children)
Workstations and server owners can actually benefit from extremely high core counts and extra memory channels, or at least, they can benefit before consumers will.
Broadly speaking, there's single-threaded and multi-threaded performance. A consumer-grade 5950X is going to clock up to around 4.9 GHz, higher than most if not all of these Epyc CPUs, and at 16 cores, it will handle everything most consumers throw at it. Compare it to a 64-core Epyc with lower clock speeds, and the user experience would actually worsen with the 64-core system.
With Zen 4, we should see 16 or 24 cores for consumers and 96 cores for workstations/servers.
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 0) by Anonymous Coward on Wednesday March 17 2021, @01:52AM (2 children)
Why do servers have so many cores, memory channels and pci lanes, but lower clock speeds? IOW, why are server CPUs and boards designed the way they are in comparison to desktop boards and CPUs?
(Score: 2) by takyon on Wednesday March 17 2021, @05:56AM (1 child)
Here are some potential reasons:
There are CPUs in the lineup that clock higher, like the 32-core EPYC 75F3, or the EPYC 72F3, an expensive 8-core with only 1 working core per chiplet.
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 0) by Anonymous Coward on Wednesday March 17 2021, @09:07PM
cool, thanks