Originally Posted by magnek
Let me rephrase that: with the introduction of a $750 16-core part into the "mainstream" platform, AMD has effectively made the HEDT platform all but unnecessary except for those actually focused on "productivity" and "content creation", however you want to define those nebulous terms. It just feels... weird I guess? Almost as if AMD is saying "nonono please don't buy X599 unless you actually have a real need for MOAR COARZ (or just want to overcompensate)
, we already have a 16-core 3950X so please just focus on that please".
But still see my comment above.
I think you misunderstand what AMD has done. They haven't made their products unnecessary for HEDT, they are making Intel's products for HEDT unnecessary at all. If you want a professional HEDT product, with the upcoming pricing and core counts of AMD with extra PCIe lanes over Intel's products. Other than that, the 16-core AM4 3950X will likely make Intel's 18-core HEDT useless in many circumstances. What they are doing is bringing their approach to more cores for cheaper in the HEDT segment down to the mainstream segment, which absolutely destroyed sales. It's pretty simple. This has NOTHING to do with telling consumers don't buy the HEDT platform and everything to do with saying Intel's HEDT platform is useless, don't buy it.
Originally Posted by guttheslayer
You are missing the point, having a $329 12C means AMD will push the lineup down by a stack, it means a 8C/16T will "downgrade" to a R5 and that is what we are hoping for, a $200+ 8C/16T that is clocked to 4.7G.
To have a gaming edge over current Intel offering, Zen 3 need to refine its latency between IF by another notch, and that is something i dont think we will see next year, aka refreshes.
And with regards to LISA comment, I think she might mean the next TR. Either way i think AMD will follow the footstep they have for TR1 and TR2, rather than go for double the cores this fast. We shall see.
That is bullspit. Yes, eventually 8-core will be moved to R5, but the frequency thing is exactly the problem. They don't need the frequency, they need software optimization and IPC. Even early samples of Intel's Ice Lake show under 4GHz, but still performing like their 5GHz products. Instead, IPC is needed and latency reductions, which would take another redesign of the IF to further reduce latency.
With TR core count, why wait? Intel has a 28-core $3000 CPU. AMD charges $1700 for the 32-core, meaning that they can put out the 64-core chips for $3000-3400, thereby neutering the Xeon 3175, which means there is NO PRICE POINT with the redesigned Zen 2 Threadripper where Intel chips will be recommended. Seeing the 3900X in flight, using multi-die and with only dual channel memory and not getting the hits like previously, there is a good chance the 24, 32, 48, and 64 chips will not suffer the same issues as first gen and second gen, making Intel's HEDT lineup meaningless except primarily for benchmarks and specific AVX-512 workloads (which are a very small percentage of actual workloads).
This is AMD dropping the hammer on Intel's HEDT lineup.
Originally Posted by GHADthc
Yeah SMT4 is going to be a game changer, and negate the need for insane core counts (Though I am sure they will still push higher than 64 cores for Epyc no doubt).
I envision AMD bridging the market segments of Mainstream Desktop and HEDT into just one socket eventually, especially with SMT4 making it so they can have insane thread counts, there wont be a need for such physically huge sockets/CPU packages in the mainstream market, and we will probably max out somewhere around 64 cores with triple or quadruple the threads...I mean in just the latest generation jump from X470 to X570, they've increased lane count, and X570 boards can take up to 128GB of ram, socket AM5 and beyond will essentially be HEDT for all intents and purposes at the rate things are going.
That isn't how threads work. Threads do not work the same as cores, nor do they give equivalent performance. For each additional thread per core, you get diminishing returns. Four threads, at most, will give 50% more performance than a non-SMT core. So 1.5X is the most you will get from it, although it increases heat at the same time, thereby decreasing frequency to some extent. Also, they will need the insane packages moving forward. The market is expecting over 100 cores in the next couple years, while AMD's timeline to reach 128 cores is ahead of the industry predictions.
Also, no, AM5 WILL NOT be HEDT for all intents and purposes. What AMD has done with more cores is forced Intel's back against a wall for a couple years. Without the extra I/O and memory bandwidth, though, and until software optimizations better utilize multi-core counts that are scaling faster than the software, we will have future proof chips without ways to stretch their legs.
Instead, these high core count products have created a situation where mainstream users are now exploring content creation and uses they normally never considered, along with enhanced multi-tasking on computers in a significant way. This just gives users more.
Originally Posted by tyvar
I don't think next year is going to be a refresh. I'm pretty sure we are going to get Zen 3 next year, and that its got quite substantive changes. For example the TAGES branch predictor was one of the Zen 3 teams babies, and they finished it up in time and it was easily to plug in and replace the old branch predictor, which is why we got it this year. This means they were already working on serious overhauls. vs Zen 1 and even what was planned originally for Zen 2.
I think the 4000 series will bring more changes in terms of IPC then the 2000 series did. Now how much IPC that actually might be I'm not sure. Also hoping for more clocks, even if its just 150mhz. They give us 150 now and then another 150 with Zen 4 in 2021, they will have negated Intel's realistic (golden chips aside) clock speed advantage.
You are correct. They are focusing on increasing IPC. Now, the reason the TAGE branch predictor was moved forward is they assumed the frequency would be much lower than it currently is, so they needed more IPC increase this gen to make up for it. Then TSMC figured out how to get the frequency up, so we had both great IPC AND frequency, thereby making Zen 2 pretty awesome.
For Zen 3, they are keeping up on the IPC front, while TSMC already has reached the same or better yields with the 4-layer EUV on 7nm+. Now, 6nm uses 5-layer EUV, while 5nm is all layer EUV. I foresee AMD going for 5nm in 2021, especially on the condition Intel may have 7nm products out late that year (which 7nm Intel density would be close to 3nm EUV density from TSMC, although Intel doesn't tell you that with high performance libraries, the densities are DRASTICALLY lower, meaning that TSMC HP and Intel HP, comparing 7nm to 10nm, are right in line with each other, meaning the uarch matters.
As to frequency, it depends on a lot of things. As you scale down on nodes, you increase heat density. So there should be no expectation of extra speed. That means that IPC increase is where the focus now is for performance, not just transistor density nor frequency.
Originally Posted by ozlay
It would be nice if they could stack the IO die. Put it under/over the chiplets. Then we could have 32c/64t AM5?
They already have plans for use of an active interposer, along with a patent for TEC (peltier) in between logic and HBM. I do not think AM5 will get 32C though. That, by 2021, may be absurd ab initio.
But, I do see them using Hi-lo HBM or HBM3 stacked on the I/O chip with the TEC, thereby giving 8-16GB on chip for mainstream, with a bandwidth of over 250GBps, thereby meaning you likely don't need to buy DDR4/5 for supporting ram on the platform (although you still could and use that to feed it, or use Intel/Micron Xpoint memory for faster storage feeding the HBM).
For the discussion of Active Interposers, see response to next comment.
Originally Posted by tyvar
I think people are putting to much stock in 3d stacking. Heat dissipation problems with 3d stacking are even worse then that of die shrinks reducing area to shed heat.
At most you will see Interposer with a CPU and GPU on top plus maybe memory on top of that, but I'm skeptical if we will even see that in any wide spread use. I think everything will just be laid out 2.5d style on a interposer and thats the most we will see.
And SMT-4 isn't happening. It actually incurs a rather substantial area penalty, were talking 20% or so area of the chip to do it right, at least on both Sun and IBMs SMT-4 chips.
I agree. 3d stacking is good in theory, but has not and will not be seen outside of low power silicon for a LONG while. Even AMD's patent of using a TEC in between the logic and the HBM seems more apt to be stuck on the I/O die rather than the core die, just because the core die puts out so much heat, if you try to overclock, the TEC would likely destroy the HBM in the process. Instead, if put on the I/O die, you still have IF latency, but you would easily be able to keep the I/O die cool with the HBM on top of it.
But, 2.5D on active interpose with HBM having its own stack and having the I/O die components integrated into the active interposer makes a lot more sense. By doing such, you get rid of the need for IF, replacing it with TSV and faster pathways that have lower latency. You then just stack out what you need.
So I think you are on the right track.
Meanwhile, 7nm+ theoretically gives an extra 18% size reduction or so, meaning doing SMT-4 may be possible, but I'm dubious. Software already has an issue on scaling to the core counts and thread counts of 22+. Server workloads do scale to it, but mainstream and even HEDT workloads do not yet. As such, I find the purpose and need a bit shaky, especially considering many schedulers had issues just dealing with Zen and Zen+.
Hopefully Zen 2 allows AMD to expand the software engineering division to help companies scale it out with thread count, with latency awareness built into the scheduler.