Don't be so quick to bag on Apple. They are seeing reduced shipments of phones and tablets; on the other hand, their CPU division continues to produce stunning improvements that are almost ignored by their core market. AMD will face A-series chips sometime in the future, possibly 3-4 years from now. But I do not think it will be a direct challenge from Apple. Apple is first-and-foremost a company that sells digital lifestyles. They do not want to make internal changes that will disrupt their corporate culture or force them to recreate divisions that have atrophied in recent years (notably their server division, which is now non-existent). So I suspect they will license A-series tech to companies that want to challenge AMD and Intel for marketshare. It will open up new revenue sources for Apple without forcing them to restructure their organization. The A12x is already a scary, scary chip. AMD is not ready to face it today, and they have to keep the pedal to the metal to stay ahead.
That chip is a dual core/quad core x86 at best. That won't be disrupting anything other than the laptop and low-end desktop market (if that and if they somehow get smart enough to license it and make it available at a competitive price). ARM has its place in the market and that will grow, but they are nowhere near being a threat to the traditional PC market and nonexistent in the server market. Apple is going to get bagged on because they made a clear cut decision to milk the snot out of their customers w/ insane margins and that has consequences
Clearly as we head into a global economic down turn. The core market for those chips is disintegrating at the stupid levels they've priced their phones :
https://mobile.slashdot.org/story/1...-9-billion-in-iphone-sales-due-to-weak-demand
9 Billion? What is that, like 9 iPhones? (Score:5, Insightful)
Guess they should have realized they were pricing themselves out of the market earlier.
^Everyone can see this and its become a common joke...
Intel did the same thing and is paying for it and so is Nvidia whose stock price has been cut in half towards 2/3rds from October of 2018. Never take a dump on your core market. Never do so while simultaneously jacking prices. Never do both of these while also chasing new non-existent markets based on hype. Apple deserves to get bagged on. They have managed to milk the snot out of people for some time w/ no consequences. Like everyone else, gravity will eventually catch up w/ them. The longer they avoid it, the harder it will come down on them. The hype has died down and they have to start accepting lower margins or they're going to get crushed. It's a global market not a milk americans market. Huwaei has surpassed them and they have yet to prove what you suggest with the A-series chip.
That slide deck is a bit weird. No update to A320 or B450. Threadripper should see an update in July/August so don't worry about that so much. I/O is not really AMD's weak point right now, since they are mostly ahead of all current and future competition in that regard on AM4 (and definitely ahead on I/O on Threadripper). I would not be too quick to ask for more DIMM slots until AMD can produce an IMC that can elegantly handle more than 2 8GB DIMMs on their consumer socket.
At higher core counts, I feel they very much have to do something about I/O. A 16 core on existing I/O capacity on AM4 would be quite weird. Threadripper, while ahead of competition is a balanced platform. North of 16 cores, clearly things got weird, silly, and I/O starved. So, again, they need to also do something w/ I/O if they ramp the core count on threadripper. The only thing that makes sense on Threadripper is 16 core (max). The other combos north of that core count have significant reductions in efficiency and sensibility. As far as how they do that, it's the same socket as EPYC.. time to stop gimping it and as I detailed in length, it's time to get past these stupid and antiquated form factors for motherboards. ATX and even E-ATX were never meant for the current core counts, I/O capacity, and fully featured dual slot full length cards that need better spacing for air flow.
We can want a knock-down, drag-out fight all we want, but right now Intel is suffering from a lull. I don't see them re-emerging until 2021 when their 7nm comes online. That may be their last shot before they have to start rethinking their business strategy. I can see them giving themselves an out by shifting towards their products like Loihi as well . . .
Agreed. They're irrelevant until they do something majorly different towards 2021 and
fix their stupid prices.
A few points here:
The original HSA paradigm as sold during the Kaveri era was a developer-facing set of confusing APIs that could be used to achieve SVM (Shared Virtual Memory). The idea was simple: if you give your GPU local access to system memory at the same basic latency as the CPU, then you can avoid all data copy between CPU and GPU. Data duplication between memory blocks made GPGPU something only useful in massively-parallel operations that had no branching behavior at all, period. You loaded a slew of data, GB at a time, onto the GPU over the PCIe bus, you performed a set of operations on the data using a kernel you'd have to build for the GPU, and then you'd sit back and let it all run. SVM meant being able to build kernels for one or two waveforms at a time, send them to the GPU, and return a result to the program with barely a hiccup. With Carrizo/Bristol Ridge you could even send multiple kernels and let the GPU return results in a non-sequential fashion (GPGPU task switching). That was really handy for sending kernels to the GPU from multiple threads.
Sadly it never went anywhere.
It better down the road because 7nm and Rome concludes the innovation based on the antiquated computer architecture we've been riding on. There's nothing left to drive the break neck pace of sales. RDMA and other great scaling technology has been expertly kept away from the desktop and its time they cut out the shenanigans. As I said, software has driven the market and now its time for hardware to do the same. Maybe there will come a point in the very near future then where software will have to set a brand new path for hardware because HSA is a must.
That is not coming back. OpenCL2.0 duplicates most (if not all) of what AMD was trying to do back then, assuming there is base driver support for it. AMD appears to have used their kernel fusion driver (amdkfd) model as a part of their OpenCL driver stack for SVM operations. So HSA isn't dead, it's just in the background . . . I guess. Sadly there still isn't a working kfd for Windows yet of which I am aware. If you want to do anything SVM-related on an AMD APU, you have to do it under Linux.
AMD has always had issues in this area. I think its due to resource starvation. Something that will be resolved as their CPU sales become much more established. OpenCL is a joke and their driver/dev stack has a long way to go to even reach parity with Nvidia's CUDA. As far as windows support, no one does any major development for such compute on windows. Everything major happens on Linux. So, that's not surprising. The future is Linux oriented. Even Microsoft gets that and is aligning accordingly.
The hope that there would be off-the-shelf "big" APUs from AMD based on Vega and Zen is basically dead, too. To date we have had Raven Ridge and that's it. In 2019 it's Picasso. Snowy Owl or whatever "big" server APU we read about in past rumours either never materialized or only exists for select customers willing to pay well for AMD to produce such products as semi-custom projects.
My commentary centers moreso around AMD opening up infinity fabric like they said they would to outside third party devices. There's a limit to my understanding of this area technically. I recall a user corrected and detailed this to me earlier in this thread regarding CCIX/Gen Z which allows AMD to do so over PCIE-4.0. But i'm referring to :
Also, as just detailed :
https://www.servethehome.com/xilinx-alveo-u280-launched-possibly-with-amd-epyc-ccix-support/
The U280 acceleration card includes CCIX support to leverage existing server interconnect infrastructure for high bandwidth, low latency cache coherent shared memory access with CCIX enabled processors including Arm and AMD. (Source: Xilinx Alveo U280 whitepaper WP50 (v1.0) accessed 16 November 2018)
So, we are very much enroute to a much different architecture and paradigm in the future.
There's no reason for the CPU to be loaded down so much with cross traffic communication between PCIE devices and memory OPs to and from I/O devices. The pipe needs to become bigger, openly standardized, lower latency, and accessible to all. This needed to be resolved yesterday. The CPU will become one of many citizens hanging off of a big pipe. It will no longer be the center of attention. It's already trending that way and established.