Return to




The exact date is unknown and all we heard is “end of July” which we can take to mean the last week of the month. All of the press that received the invites are under NDA and TR2 was listed as the official reason for the event (that doesn’t mean there can’t be any hidden surprises from the company)


Hoping by 2020-2022 we see a mainstream 16 core 64 thread CPU hit the market. Still half expecting either AMD or Intel to start cramming memory around the CPU. Jut seems like the final frontier for silicon, cluster a bunch of cores and cram ram under the lid.


Wait 4 Threads per Core?!
Please explaine what you mean with ram around the CPU!?


Was just thinking about IBM’s current CPUs are 4-8 threads per core and it seems like some one would get the idea to put extra memory beside the die to have a much larger cache that is still faster than RAM.

Ye ol Edit. IBM currently has a 12 core 96 thread CPU that can use 8TiB of ram and access 230 GiB/s of it. Somewhat expecting AMD and Intel to go off in that direction at some point.


Ahhhh, okay, I completly misread that.

But it seems, at least to me, that those are very niché even for DCs workloads. Most CPUs I’ve seen in the DCs I have “access” to have 98% Xeons und 2% Opteron and the Xeons are middle of the mall Xeons just a frikking bunch of them.

Hey, competition is good and all but I would say that IBM is trying so settle in a niché while Intel and AMD are going for the open market.


That’s really a niche use case though. You really only need that for servers that spawn thousands of threads. SMT only increases performance 1-2% at the most, and that’s only in a high thread count workload.

I don’t see more than 2 SMT happening for a while. It’s just a waste of logic that could otherwise go to actually improving the performance of the CPU.


While I certainly agree that the average core count needs to go up, I’m just as excited at the prospect of incorporating stacked DRAM (e.g. HBM/HMC) into the CPU. Having 4GB or 8GB (or even more in the future) as an L4 cache would do great things for performance.

Greater performance isn’t the only benefit. Ever since seeing how much smaller an R9 Fury was compared to other cards at the time, I’ve wondered when a similar tech would be incorporated into the CPU side. The HBM3 spec allows for up to 64GB in the package. Those kind of capacities make it almost redundant to even have main memory at all (for most systems outside of HEDT/workstation and server uses). Think how much area on the motherboard is required for main memory traces and slots. Think how much smaller a motherboard could be without them.


Indeed exciting but what on the table.

I read 7nm+ as more power management .

More ram and integrated CPUS may be 5nm shit.

I can see the 14nm to 7nm road. Intel will be butt hurt.


I think we’re going to see some really cool solutions coming out of AMD in the near future regarding memory and CPU innovations.

AMD changed the game way back when they implemented the 64 bit instruction set, they did it again when they built an entire ecosystem around a single die.

Next they’re going to start making more changes. Adding ram (either DDR or HBM) on die and moving memory controllers and other bits from the dies with cores over to their own exclusive die.


this is also down to fundamental differences in architecture between x86 and power risc though, IIRC x86 multithreading packs more into a single SIMD path whereas POWER has a much more open queue

not sure you could implement that in a practical way in x86, as upping the threads per core would probably increase the complexity of each core such that it’d waste a ton of die space, or come with extreme IPC compromises


Hmm. Difference between hyperthreading and no hyperthreading on haswell in my experience is WAY more than 1-2%.

Like… order of magnitude more than 1-2%.

Obviously workload dependent, but i definitely noticed the difference between my i5 and an E3 (~i7) based haswell CPU. And it wasn’t just clock speed.


It really is workload dependant. The real benefit of hyper threading is that the CPU spends less time task switching, and that is extremely expensive.


Sure, but even modern games benefit significantly from even hyperthreading thread count increase.

It’s not just server only - i think “Science Studio” on youtube did a bunch of benchmarks on it using an i7 with hyperthreading on/off and limiting core count. 4 cores without hyperthreading vs 4 cores with hyperthreading on even a 2-3 year old game like GTA5 makes a difference. Even more so with hyperthreading on/off with a dual core from memory.


The chip is expected to feature a base clock of 3.4 GHz and a maximum boost clock of 4.0 GHz while the precision boost overdrive clocks are rated at +200 MHz so expect up to 4.20 GHz in single core optimized workloads.

I want a 2990X so so bad…


Holy shit, now let’s wait for the pricing on these. I can’t imagine we’ll see sub $1500 on either of the top two.


is still a steal compared to whatever intel has to compete with it


Probably, but my concern isn’t necessarily comparable price, but if I can afford it LOL

I want, therefore I AMD.


Yep. i wouldnt say no to either one of those. I can just imagine all of the things you can do with all of those cores and PCIe lanes


so much room for activities!

All the transcoding, all the capture cards, all the GPU brute forcing. Sign me up!


Don’t we all mate. :stuck_out_tongue: Clearly you should dumpster dive server parts at work and flip them. /s