Announcement

Collapse
No announcement yet.

Linus Torvalds Switches To AMD Ryzen Threadripper After 15 Years Of Intel Systems

Collapse
X
 
  • Filter
  • Time
  • Show
Clear All
new posts

  • Originally posted by coder View Post
    Well, there was an outrage flareup over the GPUs in AMD's new APUs, and then another one about the morality of AMD allowing notebook makers to sell them in single-channel memory configurations.

    So, I'd say brain chemistry is an improvement. And if you don't agree, have some Acetyl-L-Tyrosine and maybe you'll reconsider.
    As long as you don't want to give me Haloperidol....btw I like Tryptophan rich food.

    Comment


    • According to ZDNet, his last build was an i9-9900k.

      Comment


      • Originally posted by starshipeleven View Post
        AM4 can use up to 128 GB (4x banks of 32GB UDIMM), which is already pretty damn high as is.
        For a consumer desktop. Not for a workstation or server. Primary use case would be dozens of VMs or even hundreds of containers. I've seen databases in production that have eaten more than 1T of physical RAM.

        Comment


        • Originally posted by GI_Jack View Post
          For a consumer desktop. Not for a workstation or server.
          Yeah I was implying that for a consumer motherboard with a consumer chipset. In the sense of "affordable" and "not sounding like a jet engine".

          To be fair it's pretty damn high for most workstations too. Only some professions that are running research applications crunching large datasets need that.

          Obviously even midrange servers can do better, but then again it's kind of expected for hardware that costs an order of magnitude more.

          Comment


          • Originally posted by starshipeleven View Post
            Yeah I was implying that for a consumer motherboard with a consumer chipset. In the sense of "affordable" and "not sounding like a jet engine".

            To be fair it's pretty damn high for most workstations too. Only some professions that are running research applications crunching large datasets need that.

            Obviously even midrange servers can do better, but then again it's kind of expected for hardware that costs an order of magnitude more.
            It isn't so much a question of cost for motherboard. It's a question of getting the last fraction of a nanosecond faster RAM speed. With registered RAM, it would have been easier to support huge amounts of RAM. With unbuffered RAM, few chip manufacturers sells large enough RAM chips so the total chip count will not end up loading the signals too much. Amd optimized for gaming, making workstation users suffer.

            When it comes to motherboards, the Intel side wins because of this. But their Xeon line of processors ends up ridiculously expensive instead.

            Comment


            • Originally posted by zyxxel View Post
              It isn't so much a question of cost for motherboard. It's a question of getting the last fraction of a nanosecond faster RAM speed.
              What? Registered and Load Reduced (the server RAM type with higher density than Registered) are slower and/or have more latency than unbuffered. Because there is a buffer, or a register between the RAM chips and the controller.

              Servers work around this by having 4 channels or more (while in consumer boards like AM4 you are still dualchannel, on enthusiast ones you get quadchannel usually), that's not as good as using faster low-latency RAM in the first place, but it sure beats having to mount 50 DIMMS per socket or something like that, as it's also going to screw with timings and speed.

              With unbuffered RAM, few chip manufacturers sells large enough RAM chips so the total chip count will not end up loading the signals too much.
              That's kind of fuzzy way to talk about it.
              In unbuffered RAM the RAM chips are connected physically to the pins on the bottom, so there is a finite amount of chips you can cram into a DIMM bank.
              There is no "signal overload", it's a pure "number of pins available" limit.

              For registered and load-reduced yes the limit is the register or the buffer, if you add too much RAM chips it's overloaded.

              Amd optimized for gaming, making workstation users suffer.
              Are you aware that Threadripper boards can mount up to 8 DIMM banks, and that 32GB DDR4 dimm banks are available (also ECC)?
              That's 256GB of RAM. I can concede that SOME workstations might need more than that, but it's very uncommon.

              When it comes to motherboards, the Intel side wins because of this.
              Maybe? If you are going to buy like 2-2.5k RAM (that's the price for filling all the slots with reg or lrdimms) and a 1-2k processor (or better), and probably add some Quadros at silly high prices too (let's be honest, who uses AMD GPUs at this level), the fact that AMD's offering costs 400 euro more (because Epyc motherboards cost around double what enthusiast Intel boards cost) isn't going to be that significant. At this price point you are mostly looking at performance, and the Epyc board can usually mount better processors than the Intel enthusiast board so it really might be a better choice.
              Whoever is working on this PC will be a very expensive employee, and not wasting his work hours is more important than saving a few hundred euro when you buy the system.

              Not saying it's 100% perfect choice by AMD to go like this, just saying that it's not as bad as you make it look like.
              Last edited by starshipeleven; 13 July 2020, 03:09 PM.

              Comment


              • Why are you formulating your post as if you don't agree and then agree with my statements? Do you work with electronics? As in designing? Or just in a fight mode?

                Originally posted by starshipeleven View Post
                What? Registered and Load Reduced (the server RAM type with higher density than Registered) are slower and/or have more latency than unbuffered. Because there is a buffer, or a register between the RAM chips and the controller.

                Servers work around this by having 4 channels or more (while in consumer boards like AM4 you are still dualchannel, on enthusiast ones you get quadchannel usually), that's not as good as using faster low-latency RAM in the first place, but it sure beats having to mount 50 DIMMS per socket or something like that, as it's also going to screw with timings and speed.
                The buffers adds transfer delay, meaning more latency. More memory channels steps up the total bandwidth - but doesn't change the latency added by buffering.

                That's kind of fuzzy way to talk about it.
                In unbuffered RAM the RAM chips are connected physically to the pins on the bottom, so there is a finite amount of chips you can cram into a DIMM bank.
                There is no "signal overload", it's a pure "number of pins available" limit.
                Not a "number of pins available" limit. A total capacitance allowed per signal trace. You can have a huge number of chip pins connected to the same signal if stepping down the frequency.

                The buffers are there to handle signal load - each signal pin adds capacitance. In older TTL, the fan-out/fan-in was a question of static current load - i.e. resistance. So there was a hard limit on number of inputs an output could source or sink.

                In modern chips it's all about the time needed to charge/discharge the capacitance load from each signal pin (let's leave the properties of the signal trace itself aside for this debate) - not the static currents to keep a signal at a fixed level. With buffers, there can be many RAM chips on the stick, that all listens to the address data and depending on address range drives the same data lines. Without buffers, the RAM chips needs to be fewer to reduce the capacitive load. So each chip needs to have larger storage capacity. This capacitance definitely represents a load - it looks like the signals are basically short-circuited until the capacitance has been charged/discharged to the new voltage. That's also a reason why the world got low-voltage signalling - the size of the capacitance and the size of the voltage step defines the amount of energy needed to toggle a signal.

                In the end, you have a bit simplified a RC circuit where the resistance limits how large current that is available to charge the inputs. Add more inputs and the additional capacitance adds to the time constant. And with a larger time constant, the bus speed needs to be reduced.

                For registered and load-reduced yes the limit is the register or the buffer, if you add too much RAM chips it's overloaded.
                It's the same rules with or without buffers. The driving side can only manage to force a limited amount of current over the signal trace. Buffers just adds a "heavy lifter" in the middle.

                With the buffers, there is one input capacitance for the motherboard to drive irrespective of number of chips on the memory module. Which is required to have the voltage stabilize within allowed time.


                Are you aware that Threadripper boards can mount up to 8 DIMM banks, and that 32GB DDR4 dimm banks are available (also ECC)?
                That's 256GB of RAM. I can concede that SOME workstations might need more than that, but it's very uncommon.
                Am I aware? Yes, I'm very much aware - I own such a machine with a Threadripper 3970x.

                When I bought it, I only found one brand of modules to get 32 GB/module unbuffered and with ECC. And 32 GB/module is what I need to reach 256 GB with 8 memory modules fitted. 256 GB really isn't that much RAM, depending on what you are doing.

                With my Xeon-based predecessor machine, I can find a large number of suppliers of 32 GB memory modules because each memory module can be fitted with twice as many RAM chips that each are half the capacity. Or I can even fit 64 GB memory modules to reach 512 GB.


                Maybe? If you are going to buy like 2-2.5k RAM (that's the price for filling all the slots with reg or lrdimms) and a 1-2k processor (or better), and probably add some Quadros at silly high prices too (let's be honest, who uses AMD GPUs at this level), the fact that AMD's offering costs 400 euro more (because Epyc motherboards cost around double what enthusiast Intel boards cost) isn't going to be that significant. At this price point you are mostly looking at performance, and the Epyc board can usually mount better processors than the Intel enthusiast board so it really might be a better choice.
                Whoever is working on this PC will be a very expensive employee, and not wasting his work hours is more important than saving a few hundred euro when you buy the system.

                Not saying it's 100% perfect choice by AMD to go like this, just saying that it's not as bad as you make it look like.
                Note that a workstation doesn't mean the need for a Quadro or other high-end graphics cards. A workstation doesn't mean a focus on graphics - even if all the motherboards for the 3970x chip all optimized for 4 PCIx slots for four graphics cards.

                But the AMD chipset is a definite limitation. I'm looking into getting one more machine just because the 3970x is a wonderful processor but the total system ends up limiting. I can't duplicate what I can do with my previous Xeon-based machine. With the AMD chipset, I need to run two machines. And only because AMD did release a gaming-class chipset, leaving a hole in the workstation segment. Then Epyc steps in for the server segment.

                Comment


                • Originally posted by zyxxel View Post
                  Why are you formulating your post as if you don't agree and then agree with my statements?
                  I'm clarifying some stuff because it's unclear, it's not necessarily disagreeing.

                  The buffers adds transfer delay, meaning more latency. More memory channels steps up the total bandwidth - but doesn't change the latency added by buffering.
                  Yeah but you said "faster" which is unclear, speed is usually bandwith, while latency is latency.

                  Not a "number of pins available" limit. A total capacitance allowed per signal trace. You can have a huge number of chip pins connected to the same signal if stepping down the frequency.
                  intersting

                  When I bought it, I only found one brand of modules to get 32 GB/module unbuffered and with ECC.
                  yeah, the Samsung DIMMS, I also use them. https://www.samsung.com/semiconducto...8A4G43MB1-CTD/

                  256 GB really isn't that much RAM, depending on what you are doing.
                  I said needing more than that is uncommon outside of some fields like research.
                  What is your field? What are you filling your RAM with?

                  Note that a workstation doesn't mean the need for a Quadro or other high-end graphics cards. A workstation doesn't mean a focus on graphics - even if all the motherboards for the 3970x chip all optimized for 4 PCIx slots for four graphics cards.
                  Most workstations greatly benefit from having a GPU with CUDA inside even if they aren't doing 3D work, since most work applications do have some way of offloading calculations or filters or simulations or whatever to a GPU.

                  With the AMD chipset, I need to run two machines.
                  How about getting a Epyc workstation board, as I mentioned above? It's a bit more expensive, but not by that much for the price of a complete system, and it's definitely better than the price of two full systems.
                  https://www.asrockrack.com/general/p...Specifications
                  https://www.asrockrack.com/general/p...Specifications
                  https://www.gigabyte.com/Server-Motherboard?fid=2301

                  Comment


                  • Originally posted by starshipeleven View Post
                    I'm clarifying some stuff because it's unclear, it's not necessarily disagreeing.

                    Yeah but you said "faster" which is unclear, speed is usually bandwith, while latency is latency.

                    intersting

                    yeah, the Samsung DIMMS, I also use them. https://www.samsung.com/semiconducto...8A4G43MB1-CTD/

                    I said needing more than that is uncommon outside of some fields like research.
                    What is your field? What are you filling your RAM with?

                    Most workstations greatly benefit from having a GPU with CUDA inside even if they aren't doing 3D work, since most work applications do have some way of offloading calculations or filters or simulations or whatever to a GPU.

                    How about getting a Epyc workstation board, as I mentioned above? It's a bit more expensive, but not by that much for the price of a complete system, and it's definitely better than the price of two full systems.
                    https://www.asrockrack.com/general/p...Specifications
                    https://www.asrockrack.com/general/p...Specifications
                    https://www.gigabyte.com/Server-Motherboard?fid=2301
                    I haven't double-checked the part number, but it was Samsung memory I bought. I started with 128 GB since I couldn't find any reference to someone who had used them with the 39xx series motherboard and they weren't in the motherboard list of compatible memory. 256 GB costs too much to buy just on a whim.

                    On one hand, I run lots of virtual machines. Some could be run on a server, but they often have some form of pass-through of hardware so they own graphics card, USB devices etc. So I end up with a single physical machine used as multiple workstations.

                    But lots of other virtualizations too, for different environments etc. Some could be automatically packaged and sent over to server machines, but it's often practical to run it directly on the machine in front of me. Especially since that gives extra bandwidth to pick up debug/trace information when doing simulations.

                    In some situations, I work with larger in-memory databases. In some situations, I set up simulated systems with millions of networked clients hammering on a server while supervising both sides to look at latencies etc. In memory access gives way better bandwidth than having to add a separate network interface to stream out data - and the 3970 has enough cores that it can directly process the debug data while leaving enough cores for simulated clients and simulated server.

                    So for me, the Threadripper processors runs rings around my Xeon chips. While my Xeon-based motherboards can do tricks that the more gaming-optimized Threadripper motherboards can't. Life would have been wonderful if I could have had 7 or 8 PCIe slots and registered RAM.

                    I haven't kept up to speed - if any newer motherboards have arrived that would motivate me to rebuild.

                    Comment


                    • And the toy/gaming chipset for the Threadripper family seems to be explained by this:

                      https://www.phoronix.com/scan.php?pa...pper-pro&num=1

                      Comment

                      Working...
                      X