In an interview with Tom's Hardware (adwalled), AMD's CTO Mark Papermaster has hinted that the 16 cores of the "mainstream" Ryzen 9 3950X CPU is not a stopping point for the company. Zen 3, Zen 4, or Zen 5 based Ryzen CPUs could feature up to 24 or 32 cores:
There are a lot of interesting details that Mark has mentioned in the interview in particular to the next-generation technologies that would be featured on their processor lineup ranging from Ryzen and EPYC CPUs. The most significant detail and the one I would start this article is with the fact that AMD isn't stopping at just 16 cores. According to AMD, there are now many applications that can scale across multiple cores and threads. The addition of cores is entirely relative to the number of applications that can take advantage of those cores so as long as this balance exists, there would not be a saturation point of cores on next-generation CPUs, whether these be mainstream or the HPC server parts.
[...] In the coming Zen iterations, Mark has stated that Infinity Fabric would continue to evolve to keep up with higher-bandwidth interfaces such as DDR5 and PCIe 5.0 (already confirmed for 2021) that would be featured on AMD's lineup around 2021-2022.
[...] AMD is also looking into integrating BFloat 16 on their next-gen EPYC lineup much like Intel's 14nm Cooper Lake CPUs which are expected to launch around mid of 2020. As for SMT4, it all boils down to whether there's enough demand or workloads that can take advantage of it.
SMT4 = simultaneous multithreading with 4 threads per core. The feature has been rumored to appear on Zen 3 or Zen 4 CPUs, although it may be included with Threadripper or Epyc instead of Ryzen CPUs.
A rumor based on a China Times report suggests that TSMC's "5nm" node is ahead of schedule and that AMD Zen 4 CPUs based on the process node could appear in "early 2021" instead of late 2021 or early 2022. The increased transistor density of the "5nm" node should allow for at least 50% higher core counts.
Related Stories
The AMD Ryzen 9 3950X Review: 16 Cores on 7nm with PCIe 4.0
Earlier this year AMD pushed again, this time putting 12 cores in the market for the same price as 8, or what had been the 4-core price point only three years prior. In three years we had triple the cores for the same price, and these cores also have more raw performance. The frequency wasn't as high as the competition, but this was offset by that raw clock-for-clock throughput and ultimately where the competition now offered eight cores, AMD offered 12 at a much lower power consumption to boot.
Today is round 2 part 2: taking that same 12-core processor, and adding four more cores (for a 50% increase in price), and not only going after the best consumer processor Intel has to offer, but even the best high-end desktop processor. This is AMD squeezing Intel's product portfolio like never before. What exactly is mainstream, anyway?
AMD's new Ryzen 9 3950X has a suggested retail price of $749. For that AMD is advertising sixteen of its latest Zen 2 cores built on TSMC's 7nm process, running at a 3.5 GHz base frequency and a 4.7 GHz single-core turbo frequency. The TDP of the chip is rated at 105 watts and it has 24 PCIe 4.0 lanes as well as dual memory channels that support up to 128 GB of DDR4-3200.
(Score: 0) by Anonymous Coward on Friday December 06 2019, @10:47PM (6 children)
AMD is going to keep on eating your lunch, year after year, until your only hope for survival is to remove IME.
(Score: 3, Insightful) by takyon on Friday December 06 2019, @11:11PM (5 children)
Intel Revenue $70.8 billion (2018)
AMD Revenue $6.48 billion (2018)
Intel has a problem of not being able to make enough CPUs.
They still dominate in laptops, and could be able to counter AMD's desktop CPUs by 2021.
If all else fails, they can resort to using their "financial horsepower" [wccftech.com].
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 0) by Anonymous Coward on Friday December 06 2019, @11:16PM
Not necessarily. If total cost of ownership is mostly fixing security holes and power, it could become impossible for intel to give away their cpus for free.
(Score: 2) by Mojibake Tengu on Saturday December 07 2019, @08:36AM (1 child)
2018 is so history...
https://wccftech.com/eight-of-amazons-top-10-best-selling-processors-are-amd-cpus/ [wccftech.com]
Respect Authorities. Know your social status. Woke responsibly.
(Score: 2) by takyon on Saturday December 07 2019, @10:32AM
"Intel has a problem of not being able to make enough CPUs" is still the current reality. Of course, AMD has its own shortages.
It's going to take something more to impact Intel's bottom line. Maybe Intel will waste their fabs on unwanted discrete graphics cards while TSMC will invest heavily into capacity for its newer nodes, allowing AMD to make more chips.
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 2) by Freeman on Monday December 09 2019, @05:20PM (1 child)
The reality is that even though AMD's revenue is 10% that of Intel, they're the best/only competitor. Without AMD, we'd be in a whole lot of hurt. AT&T get's too big for it's britches, you can can break it up easy peasy. If AMD goes under, what will we have as an alternative, ARM? AMD-Intel? AMD-Facebook . . . okay, yeah, nightmares fueled.
Joshua 1:9 "Be strong and of a good courage; be not afraid, neither be thou dismayed: for the Lord thy God is with thee"
(Score: 2) by takyon on Monday December 09 2019, @05:53PM
AMD's doing a lot better financially and they broke out of most/all of a lame contract with GlobalFoundries.
If AMD hadn't spun off GlobalFoundries ("real men have fabs!), they might have been destroyed.
Now I'm wondering why some company hasn't bought AMD. They missed some good chances to do it.
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 5, Informative) by fyngyrz on Friday December 06 2019, @10:56PM (7 children)
From TFS:
I wrote the ray tracer in my image processor to be able to launch as many threads as you have cores (or less, if you so choose.) The tracer hands off each scan line to be traced to the next available core, that way all the permitted cores remain max busy until pretty much the last few lines of the trace.
It's down to memory bandwidth / hardware caching strategies. At some point, if you're doing a lot of near-memory access, more cores won't help much, because only so many cores can get to the same memory at once, as I understand it anyway.
But at 24 cores (all I have), I'm still seeing a fairly linear decrease in processing time as I go from 1 to 24, so I'd definitely welcome more.
--
Hypocrisy is the Vaseline of political intercourse.
(Score: 2) by takyon on Friday December 06 2019, @11:25PM
24-32 cores maximum with "5nm" Zen 4 Ryzen. With a new AM5 socket, it might support quad-channel DDR5 SDRAM.
Sometime after that, there's "3nm" with new gate-all-around transistors.
Assuming Threadripper and Epyc have up to quadruple the Ryzen core count, that's 96-128 cores. Possibly with support for SMT4.
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 3, Interesting) by opinionated_science on Friday December 06 2019, @11:25PM (3 children)
molecular wobbler here. Give me 1024 threads, I'll chew them up!
How about a new marketing trope "Thread Monster".
In any case, there is a lot you can do with 256 threads and 2TB of memory , so here's hoping they push the envelope.
(Score: 0) by Anonymous Coward on Friday December 06 2019, @11:31PM (2 children)
Don't you ever think there must be something fundamentally wrong with what you are doing if it takes a supercomputer to simulate a single molecule for a few seconds? I mean it isn't like nature has a supercomputer running for every molecule.
(Score: 4, Funny) by takyon on Friday December 06 2019, @11:37PM
On the other hand, a lot of good chemists have been burnt by mother nature.
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 4, Funny) by The Mighty Buzzard on Saturday December 07 2019, @12:17AM
Yeah, the universe simulates a molecule with the actual molecule. It cheats.
My rights don't end where your fear begins.
(Score: 0) by Anonymous Coward on Monday December 09 2019, @08:52AM (1 child)
Wouldn't using GPUs be better bang for the buck for such tasks?
Are high core general purpose CPUs really better than using GPUs for rendering, encoding and other specialized computation intensive tasks?
Even for high density VM hosts, if you have many active VMs you might start to run out of I/O or RAM before you run out of CPU. So you might be better off having many lower density VM hosts instead.
(Score: 2) by fyngyrz on Tuesday December 10 2019, @02:52PM
Absolutely they would. However:
...so I'm not a fan of the GPU approach.
Based on the above issues, yes, they can be. Faster, no, and not likely to be for some time yet.
There's no VM involved in this at all, unless you are actually running this in your own VM. This is a desktop application.
--
The eyes are the windows to the soul.
Sunglasses are the window-shades.
(Score: 2) by bradley13 on Saturday December 07 2019, @06:49AM (3 children)
Beyond a certain point, adding more cores can actually slow down processing, due to the extra coordination required by the operating system.
For anything except very special workloads, I suspect 8cores is where diminishing returns set in, and 16 will simply be useless. More than 16 only gives begging rights, not performance.
Everyone is somebody else's weirdo.
(Score: 0) by Anonymous Coward on Saturday December 07 2019, @07:44AM
Agreed. 16 cores ought to be enough for everybody.
(Score: 2) by takyon on Saturday December 07 2019, @10:29AM (1 child)
I actually doubt this is the case in any way that you will notice, if done properly.
Basically, it depends on the scheduler and the way chiplets are used. AMD has put 8 cores and L3 cache into chiplets. Let's say your next-gen Ryzen has 24 cores across 4 chiplets. Each chiplet has 2 disabled, unusable cores. If you have a program that needs to use 2 cores, these should be located on the same chiplet and use the same pool of L3 cache. Using 1 core each from 2 chiplets could induce a small performance penalty. But this should almost never be the case since you have so many cores to work with.
If your program needs 8+ cores, then it is presumably embarrassingly parallel and it should be speeding up despite any small penalties imposed by inter-chiplet communication over Infinity Fabric.
Your use of a bragging rights core count should not complicate things too much. It may result in entire 8-core chiplets sitting idle most of the time, so maybe you paid too much for your CPU. It remains to be seen whether or not software will exploit all of the cores. Games are definitely going to use 8 cores since that is going to be the console baseline. Other software is going to trend towards using more cores if possible, since mainstream core counts have quadrupled from 4 to 16 and are likely climbing higher.
Threadripper 3000's massively improved performance over Threadripper 2000 suggests that high core counts are not necessarily a problem. Threadripper 2990WX starved 2 of the dies of memory bandwidth [pcworld.com], leading to scenarios with big slowdowns. Comparing 32-core 2990WX with 32-core 3970X in benchmarks shows a lot of improvements and more consistent results between the 3970X and mainstream desktop chips.
One possible benefit of the high core count CPUs could be the greater amount of usable silicon. The 16-core Ryzen 9 3950X has the highest turbo clock in the lineup because AMD has used a slightly better binning for the chiplets, but the 12-core Ryzen 9 3900X also has a high turbo clock. Because there are 2 chiplets instead of the 1 chiplet that some of the lineup has, it could run a highly clocked thread on each chiplet instead of two on the same one, spreading the heat.
Maybe this is where the operating system complication comes in. After all, some amateur coder was supposedly able to make Ryzen run ~250 MHz faster just by using a custom power plan [soylentnews.org] that more effectively matches high performance threads with the best cores.
We are on Zen 2 right now. A couple changes are coming in the future. One confirmed change is that Zen 3 will unify the L3 cache for all 8 cores on the chiplet. Previously, the chiplet was comprised of two 4-core CCXs, each with their own pool of L3 cache. And they apparently had to connect over Infinity Fabric to communicate. This change should speed things up.
Zen 4 is rumored to include L4 cache stacked on the I/O die. This would be a larger amount than the L3 cache, likely at least 1 GB and maybe 8 GB at most, and would be used by all cores in the CPU. It could reduce trips to off-CPU DRAM and speed things up some more.
Zen 4 is likely to increase Infinity Fabric speeds and memory bandwidth. Sources of latency may be reduced. This should help communication between chiplets.
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 2) by takyon on Saturday December 07 2019, @03:22PM
AMD can also choose to increase the number of cores on each chiplet, e.g. 12 or 16, or keep it at 8 cores but shrink it (aside from adding additional L3 cache).
I think they will keep it at 8 cores and enjoy better yields because of it.
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 3, Insightful) by fadrian on Saturday December 07 2019, @10:26PM (1 child)
There are several types of loads that will benefit from this - music and video production, simulations, machine learning, and the list goes on. People who think that this power is useless are deluding themselves.
That is all.
(Score: 2) by takyon on Saturday December 07 2019, @11:34PM
Some of these use cases you mention would warrant going directly to Threadripper to get even more cores, better I/O, more RAM, etc. Not that it wouldn't be nice for users to get more cores for less money, allowing a climb down from Threadripper in some cases.
Push Ryzen towards 24 and 32 cores, and new software and games will need to be written for more casual users to see any benefit. I am however, pretty confident that more parallelized software will materialize, allowing cool new things to be done rather than just bloaty things.
I'm not sure that these CPU cores are the best thing for machine learning. Although:
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 2) by turgid on Sunday December 08 2019, @11:19AM
I'm waiting for itanic to come out. AMD will never be able to compete.
I refuse to engage in a battle of wits with an unarmed opponent [wikipedia.org].