AMD Launches 2nd Gen Ryzen Pro & Athlon Pro APUs

40 Comments

View All Comments

• SaberKOG91 - Wednesday, April 10, 2019 - link

  1) Old Process != cheap. Zen2 chiplets do not have the same power management features as the mobile parts. Mobile Ryzen doesn't have enough PCI-E lanes for an IF link to an external I/O die.
  2) Not really. High-end parts are low volume and would not chip in high enough quantities to offset the tooling and marketing costs, for AMD or OEMs.
  3) Again, Zen2 chiplets of today do not have the power-saving tech found in the mobile parts.
  4) Fine. Time will tell.
  5) An interposer or other integration tech is 100% required to connect the packages together. FP5 is just a normal BGA package where the traces connect the die contacts to the BGA pads. You would still need another layer to connect the packages together or you would have to route through to the motherboard and back which adds latency, power consumption, and design cost. No one would do this. Even if you were stupid enough to do so, the GPU/IO die you have proposed on 14nm would barely fit on the package and only be as fast as last gen. You underestimate how much of the APU die is GPU (hint: over half). The reason that matisse won't have a GPU chiplet version is far simpler: limited memory channels. The size of GPU that you could fit on matisse would require 2 additional memory channels just to keep the GPU fed. The current APUs are already bottle-necked by having to share two channel with a GPU on a 4C/8T system. This is why increasing memory clockspeed (bandwidth) has so much uplift on gaming benchmarks. Trying to do that with 8C/16T and a GPU that's at least as powerful as a 560 and the same number of memory channels is not remotely feasible. I can guarantee you that when the 16C/32T matisse parts get reviewed we will see two memory channels become a bottleneck and that is why AMD made the L3 cache so much larger (to reduce the impact).

• neblogai - Thursday, April 11, 2019 - link

  1) Old process, and a lot of reused blocks (Vega, not Navi; same display, multimedia engines, etc) = way, way cheaper and could be done for this summer, compared to trying to build everything new on 7nm for next year.
  2) Still needed if AMD aims to enter the high performance mobile/embedded market. Entering with an expensive design is even worse, because sales will not jump up anyway, and will only ramp up slowly, over many years.
  3) With an energy efficient I/O and iGPU, Zen2 7nm chiplet with behavior tuned for efficiency would be good enough for long battery life, radically better than current desktop chips + dGPUs in certain Asus/Acer laptops.
  5) Interpozer is not needed, same like it is not needed for Epyc/Threadripper, or Matisse. So, it is cheap. And you are thinking wrong about the GPU for it. For such mobile chip, mobile iGPU would be needed for extreme energy efficiency at idle, multimedia playback and video outputs, not for rendering games. So, 1-3CU in total, and no iGPU memory bandwidth issues. For Embedded and HPE- that is enough, and for gaming machines- proper dGPU would be added and switched on while gaming.

• SaberKOG91 - Thursday, April 11, 2019 - link

  1) You are neglecting tooling costs. Small runs on a 1000+ step process are not cheap, especially for something they won't sell more that a few tens of thousands of.
  2) They don't want to or they would have by now. What you have proposed IS an expensive design that no OEM would buy. Even if they could, it takes time to roll something like that out which AMD is no doubt planning on doing this fall with OEMS for next spring. It's too late in the year for something like that and we've had no leaks whatsoever.
  3) Not really. Even if you can dramatically reduce dynamic power consumption by running at lower TDPs and idle clocks, you still have to properly manage static power dissipation. Desktop parts typically have an order of magnitude higher static dissipation than their mobile counterparts. This is precisely why Ryzen Mobile 3XXX is not using the same die as the desktop version. That fact doesn't change for your imagined 7nm chiplet. When TSMC is talking about 0.5X power consumption they are talking about dynamic, not static.
  5) Epyc, Threadripper, and Matisse all use interposers. Not sure where you are getting your information from, but it's flawed. Infinity Fabric links are routed on the interposer between the chiplets directly in Zen1 and now to the I/O die in Zen2, not through the package substrate or the motherboard. Only the socket-to-socket IF links actually reach the LGA pads and motherboard.
  
  1-3 CU? I thought we were talking about a supposed high-end mobile part -_- . Now you're talking about embedded and HPE. Even Intel's low-end mobile parts have more power that that. If AMD want a high performance mobile part, they would just shut off the built-in GPU, raise the TDP to 30W+ and then bundle a dGPU in the 550 or so range. No switchable graphics needed as Polaris can get very low idle power when undervolted and underclocked. Or they could go back to the Bulldozer era and pair an equally performing dGPU to match the iGPU and have you use crossfire. The GPU on ryzen mobile is deliberately undersized for the CPU because of memory bandwidth. You're grasping at straws now mate.

• neblogai - Thursday, April 11, 2019 - link

  Yea, I also feel like I'm wasting time..

• SaberKOG91 - Thursday, April 11, 2019 - link

  Well, you were never going to get me to agree with you :p You keep arguing about ambitious and lofty goals that AMD doesn't have the kind of spare cash to justify chasing. If this were Intel we were talking about, none of this would be unreasonable. Instead it's a company that makes an order of magnitude less revenue each quarter. Don't get me wrong, I like AMD, but this just isn't how they do business.

• neblogai - Thursday, April 11, 2019 - link

  Well, they could be making more cash by reusing their tech more widely. They actually seem to try to do that with scaling Zen and Zen2 dies for several markets at once. But yes- this would be one more project, and maybe AMD do not have the capacity to pursue everything. Simply, this is my take on what would be very reasonable for them to do, in contrast to AdoreTV rumors of Zen2 desktop APU with 14-20CUs- which do not make sense.

• TheinsanegamerN - Thursday, April 11, 2019 - link

  So is AMD going to support these ones properly, or are they going to let them wither on the vine and require OEMs to update driver for AMD?
  
  That is what killed first gen mobile ryzen, along with poor battery life and power usage in general. A tale that is as old as time for AMD.

• Irata - Friday, April 12, 2019 - link

  You can now run the regular Adrenalin drivers, battery life is great if the laptop is properly configured and built and power usage is also fine. Get 9+ hours on my 2500U based laptop.
  
  The reported issue was idle only, so if you get a laptop to let it idle for hours on battery then maybe mobile Ryzen is not for you

• TheinsanegamerN - Friday, April 12, 2019 - link

  Sure, its fine NOW, over a year after release.
  
  Is AMD going to do this again? Because if they do, then mobile 2nd gen will be just as much of a sales failure as the 1st gen. Some of us dont want to wait a year before our products work like they are supposed to.

• Soulkeeper - Tuesday, April 16, 2019 - link

  I wish the ecc memory I purchased worked with my 2700U
  It is a busines "pro" scam afterall ...