|
The GPU Megathread is here, this thread is about CPUs, APUs, and AMD platforms in general. After six goddamn years of being completely uncompetitive and terrible in almost anything but the cheapest parts of the CPU market, AMD has finally managed to produce a completely new design that isn't trying to salvage anything from the construction core dumpster fire. The Zen microarchitecture is a modern design created in part by Jim Keller, the heroic lead designer of the Athlon (K7) and Athlon 64 (K8) architectures, and it manages to actually be quite competitive with Intel's best offerings. Models: Currently, the R5 and R7 lines of desktop processors have been released. code:Benchmark Summary  In terms of price for performance, the superiority of Ryzen over Intel's Broadwell-E CPUs is indisputable, with the only options from Intel that's strictly superior to the Ryzen chips being the i7 6900K and i7 6950X, which cost $1,049 and $1,649 respectively. The R5 1600X is essentially equivalent in terms of performance to the lower end Broadwell-E hexacore, the i7 6800K, although the gap in single-threaded performance does allow the i7 6850K to be slightly ahead overall. More controversial is the comparison with Kaby Lake. In productivity tasks and multi-thread optimized games, the hexacores (much less the octocore R7s) are much faster than even the flagship i7-7700K because they have so many more threads of execution available to be utilized, even if each is not quite as performant as the higher-clocked Intel quad core. However, when one is exclusively playing purely/mostly single threaded games while not using their computer in other ways, the i7-7700K is faster in CPU bottlenecked games. Some focus on this quite heavily, but the fundamental truth of the matter is that the performance difference will be small in any reasonable setup. As long as you're not pairing a Titan Xp with a $249 CPU or more broadly aren't spending huge amounts on a gaming PC to merely run games at 1080p or lower resolutions, the bottleneck will almost always be the GPU rather than the CPU. In addition, most benchmarks are done with NVIDIA cards (understandable considering the fact that AMD hasn't put out an actual high end GPU in ages), which seem to have performance problems on Ryzen, compared to AMD's own graphics cards. In addition, with DX12 and Vulkan, games will be getting more and more multithreaded, thanks in part to the use of old, crappy low-power AMD cores in the consoles this generation (including the PS4 Pro and Scorpio). Overall, looking forward, Ryzen is a much better value proposition than Intel's offerings, at least at the R5 1600 level and above. The Ryzen quad cores, however, are only really competing against the bottom of the barrel i5s and the i3-7350K, which they are indeed superior to. Motherboards: Ryzen uses the new AM4 socket, and motherboards are available with the following chipsets:  However, at the moment, the motherboard situation isn't ideal. It's been a long time since AMD has launched a new CPU, and there have been some significant teething issues with BIOSes and especially the memory support. Ryzen is quite picky about the RAM timings that are used with it, and the long story short is to actually pay attention to the QVL list for whatever motherboard you are getting. In addition, there are some fairly important motherboard features beyond the standard motherboard issues (like VRMs), with special relevance to using high speed memory, which is far more important for Ryzen's performance than it is for Intel CPUs, because of how it affects the clock rate that the internal communication bus between the modular complexes that make up a Ryzen CPU. In addition to the quality of the actual electrical design of the board, some boards offer external clock generators that may allow higher memory speeds and (depending on the settings) either better or worse stability than the default 100 MHz. The following motherboards have BCLK generators:
However, there are some reports that BCLK overclocking may cause problems with peripherial clocks (PCI-E, USB 3), but if you are planning on overclocking it is worth getting a board that has it, since it is hardly more expensive than any of the other top quality X370 boards. Sinestro fucked around with this message at 21:37 on Apr 14, 2017 |
#
¿
Apr 14, 2017 21:16
|
|
|
|
| # ¿ May 5, 2024 03:23 |
|
|
repiv posted:Those should be 16 threads not 12 SwissArmyDruid posted:Your thread counts are wrong for the R7 parts, Sinestro. Fixed.
|
|
#
¿
Apr 14, 2017 21:38
|
|
|
priznat posted:I'm curious if the current version of InfinityFabric phy can support PCIe Gen4 speeds (8GHz, 16GT/s), because it would be pretty impressive if they could jump right into the Gen4 ecosystem. There's a few new features they'd have to implement to get to the Gen4 spec of course. It's 32 bytes wide, so it's exactly that fast. 16GT/s * 32 bytes/T = 512 GB/s.
|
|
#
¿
Jul 10, 2017 06:19
|
|
|
The issue with IF as implemented in Zen 1 is that the memory controller also handles generating the clock for the bus, and that clock is at the same rate as the one that goes out to the DDR4 PHY. So in the most pessimistic case of DDR4-2133, it's limited to just 68GB/s, and in the perfect case of DDR4-3200 you're still only getting 102GB/s. Hopefully future revisions will fix that... frankly inexplicable design choice.
|
|
#
¿
Jul 10, 2017 20:03
|
|
|
Yeah, I greatly overstated it, because there is an obvious explanation that, indeed, it does work well enough and doesn't involve anything with asynchronous clock domains or any such dark wizardry that would be hell to deal with. I'm not an expert on actual hardware implementation by any means but I know enough that it'd be a nightmare. I don't think that having it run at the full 8 GT/s would be feasible at all and definitely wouldn't be worth the effort when they're trying to get Zen 1 out the door, but it wouldn't have been that hard with a multi-master bus to run it at double the RAM speed, kept in phase, which would double the inter-core bandwidth. It wouldn't be an unimaginable surprise if something like that happened even with the same silicon on the MCM parts, because trying to have 32 cores squabbling over what in the case of a server is almost certainly 68GB/s bandwidth doesn't seem like it's a recipe for success.
|
|
#
¿
Jul 11, 2017 01:59
|
|
|
I know I'm an awful failure baby but I kinda love that stuff a lot of the time. At least in the sense that I can make a custom color scheme than the full on hypnotic cuttlefish.
|
|
#
¿
Jul 26, 2017 03:10
|
|
|
Nah, it's just changed rules for SIGGRAPH.
|
|
#
¿
Jul 29, 2017 19:43
|
|
|
That Geil RAM looks so ugly and weird. Why did they give it such tall heat spreaders it almost looks like SODIMMs?
|
|
#
¿
Aug 4, 2017 03:52
|
|
|
Oh, I know why there's heatspreaders that do nothing in the first place, I'm just confused by why they look so terrible and not even in an  way.
|
|
#
¿
Aug 4, 2017 17:13
|
|
|
It's stunningly specific. They should really take the microcode that checks if it's running a C compiler and creates random errors out, I don't think that's going to be their silver bullet versus Intel. Wait, gently caress. The conspiracy theorist in me almost imagines some sort of insane "Reflections on Trusting Trust" scheme that's not quite working right.
|
|
#
¿
Aug 5, 2017 18:21
|
|
|
You can use Xen and the performance is great, I've helped my best friend set it up on his Ryzen desktop. You have to patch the driver to fix the Code 43 "bug", and that's it.
|
|
#
¿
Aug 9, 2017 09:07
|
|
|
The appeal of the octocore TR part is the idea of being able to upgrade in the next generations, along with the issues that Ryzen has with 16GB DIMM clock speeds and wanting 64GB for VMs.
|
|
#
¿
Aug 11, 2017 03:38
|
|
|
Malcolm XML posted:hmm perfect for my bunker oil burning generator powered wanking station I can only masturbate if I know it's harming polar bears.
|
|
#
¿
Oct 28, 2017 20:54
|
|
|
https://www.youtube.com/watch?v=ucMQermB9wQ I'm really not sure how I feel about this video. It's pretty crazy, but nothing that's like... 100% not believable.
|
|
#
¿
Feb 22, 2018 07:12
|
|
|
SwissCM posted:Nation state hackers, cryptocurrency, manipulation of the masses through social media technology and CPU hardware exploits being used to manipulate the stock market, we truly live in the lamest cyberpunk future year 2018. Where's the cool poo poo like jacking off your brain into microchips or whatever. Is the brainstem your brain's dick? -- The above sentence contributes about as much to the world of security as this AMDFlaws.com poo poo that looks and feels straight out of a clickbait "related content" ad next to "Local Semiconductor Process is 10, looks 7!". Sinestro fucked around with this message at 10:24 on Mar 14, 2018 |
|
#
¿
Mar 14, 2018 10:21
|
|
|
Paul MaudDib posted:See also: armchair architects who decided that since Ryzen used a MCM packaging, that AMD would obviously be introducing a MCM strategy with Navi as well. Since, y'know, it's not like people have been trying to make multi-GPU work for say, the last 20 years or anything, they didn't have the magic of It's more to do with the performance of the interconnect. AMD was prototyping a MCM design all the way back with Cypress, but HyperTransport maxed at 25.6 GB/s, and it was pushing at the limits of SerDes design and the transistors themselves to do so. The per-channel memory bandwidth of Cypress was 38.4GB/s, so even using one link per memory channel, that'd be a huge drop in performance. Looking at the RX 580, since that sort of midrange design is exactly what would be used for the component dies of such a design, it also has four channels of memory but now with 64 GB/s of bandwidth per channel. The intra-socket component of Infinity Fabric, running at 1333 MHz as in the case of Epyc/TR with DDR4-2666, has 42.6GB/s of bandwidth, but it is designed to be flexible with clock speed and I don't doubt that it would be possible for a GPU to clock it at 2000 MHz, which would mean that each link would be equivalent performance to a channel of memory, which is how they are allocated in the CPUs.
|
|
#
¿
Jul 19, 2018 20:13
|
|
|
Malcolm XML posted:I suspect any MCM style GPU will have to hide that it's a MCM and simply figure out a scaling method that is transparent which is gonna be really hard. Not really, at least in theory. IF allows them to connect with the same protocol as the network-on-chip (and the same bandwidth, if they're using enough links), and then the schedulers can distribute work to compute units on other dies (essentially) transparently, with their memory requests going out to controllers on other dies equally transparently. "At least in theory" is the key part there.
|
|
#
¿
Jul 20, 2018 04:27
|
|
|
I can't find an option on my Crosshair VII Hero to change which GPU is used to display the bootsplash/EFI. I got a secondary GPU to use as a host adaptor for virtualization stuff, and I want the boot stuff on that, ideally without moving my real GPU out of the slot that it's in currently, both for bandwidth reasons and because I don't think I could get the GPU's AIO tubing to work out like that. Does anyone know if there's some option that controls this, or am I poo poo out of luck?
|
|
#
¿
Dec 17, 2018 20:54
|
|
|
|
| # ¿ May 5, 2024 03:23 |
|
|
I’m pro elf tiddy but only if it’s aimed at cornering the lesbian market.
|
|
#
¿
Jul 2, 2019 21:51
|
|