New Ryzen 3000 kinda looks like the Sandy Bridge of AMD: Single thread performance boost, major clock gain, and power reduction all at the same time. Gonna be a great.
There were plenty of people who felt trapped by Intel's HEDT lineup at the time that saved a bunch of money going AM4 in 2017.
They EoL quickly and go bargain basement. Just like what happened with AMD's AM3+ and FM2+ products in 2017. AMD can't afford to be picky about what happens to yesterday's products when they have a lean and hungry monster chasing them from the back. They need to be positioned against their next competitor. Not just Intel. Hector Ruiz is the kind of a guy who would say, "ehh let's hold back on consumer Zen2 so we can clear out the channel a bit and protect Threadripper". Look at what he did with the x2. Lisa Su can not afford to make the same mistake, albeit for different reasons. The challenges that face AMD in 2021/2022 may be more difficult to overcome than what they saw in 2006. Now is the time to prepare, not only technologically but also in terms of market share and mindshare.
All current CPU prices will fall, much like Ryzen 1/TR 1 CPU prices fell once Ryzen 2/TR2 CPUs launched. This is progression, old CPU values are falling and new CPUs take their place.
TR2 16C 32T 2950X currently selling for $900, when Ryzen 3 16C 32T will launch it will be on the market for at least 6 months, its price will start to fall to the $600 price level. Also remember that TR platform offers more features than AM4 (Quad Memory Channel, more Cores, more PCI-e lanes etc), so even if you have a 16C 32T AM4 CPU at $499, a TR2 16C 32T CPU can easily be priced at $599. Also users of TR platform will be able to upgrade up to64C 128T32C 64T TR3 some time in the future and that is another selling point for the TR platform.
I think you're missing something big here in your analysis. AMD will beat the 9900K with its own 8c 3600X, and this will be the CPU that is priced at half the 9900K price. The 3600X will not be the best CPU that AMD will be offering on AM4, so it cannot then command the premium for being the best.
IMO, AMD will price the 12c 3700X below the 9900K, and the 16c 3800X at 9900K pricing.
We're not saying that the best CPU will be at half the price of a 9900K, only that AMD's most comparable CPU will be. That's what matters. That's what kills Intel completely outside of OEMs.
We don't know how much the increases will be in these areas. Nor do we know the price, regardless of what some are insisting on.New Ryzen 3000 kinda looks like the Sandy Bridge of AMD: Single thread performance boost, major clock gain, and power reduction all at the same time. Gonna be a great.
We don't know how much the increases will be in these areas. Nor do we know the price, regardless of what some are insisting on.
The expectations are through the roof in this thread. If the numbers don't end up being what AdoredTV has listed, the beatdown by the pro-Intel camp will be brutal.
I don't see how else they could price it if a) it is better than 9900K, and b) there are 12c and 16c variants.
As others have said, they can't make the 16c more expensive that a TR, so that immediately places that at current 9900K pricing.
I do think Ryzen 2 will be an excellent product. But you're right. The speculation is running rampant. Temper your expectations folks. Leave some room to be pleasantly surprised. Allow yourself to be happy. ;-)Exactly. The "leaks" are rumors and nothing has been confirmed. I hope that the leaks are true! I'd love to not have to make a huge compromise by going with AMD and there's a sizable amount of ground to catch up with in gaming.
I don't disagree. I'll add however that with a 70+% share price jump AMD needs to be at least somewhat cautious about how it prices its products.
That chip is a dual core/quad core x86 at best.
HSA is a must.
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.
My commentary centers moreso around AMD opening up infinity fabric like they said they would to outside third party devices.
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)
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.
I do think Ryzen 2 will be an excellent product. But you're right. The speculation is running rampant. Temper your expectations folks. Leave some room to be pleasantly surprised. Allow yourself to be happy. ;-)
Cavium already has product :Don't be blind. Cavium could license the core from Apple and build a 48c or 64c version out of it and smash records for performance.
My understanding is that companies other than apple already have competent ARM complexes they can leverage. I'm not sure why someone would go out of their way to license a mobile first platform with tons of fixed accelerator functions that they'd have to strip out. Regarding Qualcomm, this just happened today :They could even get core size/cost down by stripping out the fixed-function hardware and GPU (unless they wanted it in there for compute purposes, which would be weird/interesting). They could also stick to the Vortex cores. And that's 2018's chip. If Cavium licensed their designs today, we wouldn't see product for another 2-3 years. We would be facing the next A-series iteration. Like I said, Apple isn't the one that's going to give anyone trouble. Someone is going to either chase Qualcomm into the notebook/desktop space or go after the server room. Apple has the best of the ARM world right now, and Apple would be insane not to license it to someone operating outside of their core market.
Tomato tomatoe, essentially I/O devices need to be freed up from the CPU and have direct access to ranges of memory/storage and be able to directly work with each other on an open standard. Everything is going towards this. This paradigm is where computing goes next. Gate shrink beyond 7nm will be costly. Were up to 64 cores now. Storage and memory need to see some innovation and peripherals need to be freed from prison. That's where the slow downs are. My point was that the core war is almost over. AMD won. Intel is mia until 2020/2021. By then, the next stage of focus is on interconnects/new io/mem paradigms.All I'm saying is, HSA as we know it is dead. If they work towards an SVM model, it will be in a different format. And it will probably be something less proprietary/better supported across the industry.
The problem is that OpenCL is a nightmare to develop on and AMD has problems when it comes to driver/dev support likely due to minimal staff. This kills off any adaptation. CUDA is a factor because its what you can far more easily develop on now. Given this, there are a lot of things left on the table which Nvidia keeps segmented in pro hardware that needs to be freed up.CUDA isn't even a factor here. CUDA does have support for UVM (Unified Virtual Memory), but nVidia doesn't have any product on the market that would really benefit from existing SVM models (APUs) outside of their APUs that are . . . well let's just say that they're operating in different ecosystems than AMD and Intel and leave it at that.
It's a mess and "restricted" is code word for that. No mention of when you get deep into develop and things just aren't there. I considered it for a while and abandoned it. Maybe at some point it will mature and I will be more than happy to jump on board. Apple dropped it as well.As far as OpenCL being a joke goes . . . I don't even know why you would say that? The OpenCL2.0 ecosystem is functional. AMD's driver stack is just restricted, that's all. Maybe you just like CUDA better (I think a lot of developers do, actually). But anything you would want to do with SVM on Raven Ridge is doable under Linux right now, assuming your end-user has amdkfd loaded. Not sure if that's standard to distros yet or if you have to use a hacked kernel.
https://www.zdnet.com/article/microsoft-releases-its-first-linux-product/If we consumers are ever to see SVM or UVM compute becoming "a thing" in our software, the first thing that will have to happen is that Windows will need to support it. At least theoretically, nVidia could already support UVM with their consumer dGPUs assuming someone would improve PCIe performance to the point that we would ever want to use it. But until there is a kfd under Windows, we won't see SVM or UVM support in Windows for any of AMD's products.
I linked to a talk. It centered on CCIX, GenZ, and OpenCAPI.Now you are maybe talking about UVM. I think UVM is about treating the memory pool of all devices attached to a system as one large memory pool, and allowing reads and writes to every device in the system. This becomes tricky since you may be crossing clock domains and dealing with heavy latency. Even without heavy latency, you don't want CPU-critical data - stuff you might need to load into cache at some point - winding up in the memory pool of one of your PCIe devices. At the same time, you want to try to keep all your data in one exact memory block to reduce writes and writebacks. SVM solves this problem by forcing the GPU and CPU to share one memory pool. UVM makes things a lot more complicated. And don't expect future consumer dGPUs to ship without memory. That is not going to happen.
There are proprietary high-end interconnects out there right now that essentially let CPUs do that already, which is what they use in supercomputers. Note that even using those interconnects - or stuff like nVidia's NVLink - you still can't put dGPUs into the system as "equal partners" since their memory controllers and cache structures are different. There's no cache coherence. There will have to be a major change in the way they build dGPUs in the future if they want them to have cache coherency with CPUs.
Sure there is. GPUs and CPUs are still not built using the same memory model. Why do you think stuff like HSA was so hard to begin with? And that was using an iGPU sharing the same physical memory block as the CPU. Also, do you think anyone wants to build a NIC or SATA controller with a CPU-like cache structure that would be necessary for it to operate in a cache-coherent device ecosystem like CCIX? That would increase costs by quite a bit on otherwise-simple devices.
Cavium already has product :
https://www.servethehome.com/cavium-thunderx2-review-benchmarks-real-arm-server-option/
The barrier to adoption is that its non-x86 is it not?
I'm not sure what Apple's custom ARM IP has to do w/ a server chip as Apple's ARM complex is tuned mainly for specific application flows (mobile)
Then comes the realization of the pricing Apple would apply to it.
I stand by my statement that you eventually pay for milking customers.
I'm not sure why someone would go out of their way to license a mobile first platform with tons of fixed accelerator functions that they'd have to strip out.
Apple would be better off proving the versatility of their ARM IP within their own ecosystem
by extending that chip into a range of MBPs. Again though, just like Intel, people are getting vocally tired of their milking. Popping an even cheaper arm processor in a mbp and charging $1,500... People are going to get tired of that.
Tomato tomatoe
essentially I/O devices need to be freed up from the CPU and have direct access to ranges of memory/storage and be able to directly work with each other on an open standard.
It's a mess and "restricted" is code word for that.
CCIX is likely coming with Rome w/ peripheral support.
https://semiaccurate.com/2017/10/26/arm-talks-ccix-details/
AMD could graft on CCIX coherency ports on the updated Vega to give Radeon Instinct cards something that looks and smells like NVLink ports on the high-end Tesla cards; we expect that the updated Vega GPU will plug into PCI-Express 4.0 x16 slots, which means a lot more bandwidth between the CPU and the GPU.
No CPU involved... Just smart nics attached to nvme
The problem I see here is ..
That performance discrepancy happens due to higher memory latency, cross-CCX latency penalty due to threads being shuffled across the CCXs and L3$ being a victim cache. AMD has not provided any detailed architectural overview of Zen 2 so we don't know how these issues have been addressed, if they've been addressed at all.Is AMD going to fix whatever made Ryzen slower in gaming despite all else being basically equal? Same clocks, same RAM, same number of cores and threads, yet Ryzen was often significantly slower than Broadwell-E despite all these things being equal. Sometimes it was a nasty difference and that hasn't changed enough even to this day, so I'm wondering if these new CPUs will have some changes that mitigate that odd disparity in gaming performance. I'm just being real here from a performance gaming perspective.
I see current benchmarks from the likes of gamersnexus doing yet another revisit of the 2600K, and guess what? A Ryzen 7 2700X still can't convincingly beat an overclocked 2600K. They are in the same general performance bracket down on the bottom half of the charts, consistently, while every Intel CPU starting at Haswell and newer occupies the top half of the performance chart. Can AMD fix whatever is causing THAT to happen? Because if they add 15% IPC but don't fix whatever the hell is causing that performance issue, then they still won't be able to reliably catch something like an overclocked Haswell or newer.
Is AMD going to fix whatever made Ryzen slower in gaming despite all else being basically equal? Same clocks, same RAM, same number of cores and threads, yet Ryzen was often significantly slower than Broadwell-E despite all these things being equal. Sometimes it was a nasty difference and that hasn't changed enough even to this day, so I'm wondering if these new CPUs will have some changes that mitigate that odd disparity in gaming performance. I'm just being real here from a performance gaming perspective.
I see current benchmarks from the likes of gamersnexus doing yet another revisit of the 2600K, and guess what? A Ryzen 7 2700X still can't convincingly beat an overclocked 2600K. They are in the same general performance bracket down on the bottom half of the charts, consistently, while every Intel CPU starting at Haswell and newer occupies the top half of the performance chart. Can AMD fix whatever is causing THAT to happen? Because if they add 15% IPC but don't fix whatever the hell is causing that performance issue, then they still won't be able to reliably catch something like an overclocked Haswell or newer.
Thank you for sticking with my commentary and providing detailed replies. It was very informative and provides me some pointers to look into. I do agree that ARM is going to play a much more prominent role in the future beyond their current ecosystem. On HSA, I had a rough idea and picture but I'm definitely technically out of my depth in terms of navigating a discussion. Thank you for clarifying and correcting things for me. I also seemed to have ignored the important thing you pointed out which is how much the board and board components would run to achieve this.
To loop back around, where do you think AMD is possible going to go w/ this Rome launch (PCIE 4.0/CCIX)?
What would you compare this to w.r.t existing tech? Is it like Nvlink over PCIE 4.0?
When code running on a CPU or GPU accesses data allocated this way (often called CUDA managed data), the CUDA system software and/or the hardware takes care of migrating memory pages to the memory of the accessing processor.
Do you think they will expose this on AM4?
Threadripper? Where could this go down the line?
Have any thoughts on NvME 1.4 and beyond and where that's headed and how that might integrate with AMD's roadmap?
Any new sys memory paradigms on the horizon? When can we expect something beyond DDR4 sys mem?
Nobody knows but my guess is nope. AMD has ALWAYS had problems with latency compared to Intel and that hurts gaming performance significantly. It should be better than Zen+, but I think halo gamers will be disappointed with the gaming performance from Zen2. The 9900k will retain the gaming crown but Zen2 should narrow the gap some.
As a 8700k owner I think the relevant question for gamers at the end of the day is; at a certain price point, who gives the best cpu for gaming?Is AMD going to fix whatever made Ryzen slower in gaming despite all else being basically equal? Same clocks, same RAM, same number of cores and threads, yet Ryzen was often significantly slower than Broadwell-E despite all these things being equal. Sometimes it was a nasty difference and that hasn't changed enough even to this day, so I'm wondering if these new CPUs will have some changes that mitigate that odd disparity in gaming performance. I'm just being real here from a performance gaming perspective.
I see current benchmarks from the likes of gamersnexus doing yet another revisit of the 2600K, and guess what? A Ryzen 7 2700X still can't convincingly beat an overclocked 2600K. They are in the same general performance bracket down on the bottom half of the charts, consistently, while every Intel CPU starting at Haswell and newer occupies the top half of the performance chart. Can AMD fix whatever is causing THAT to happen? Because if they add 15% IPC but don't fix whatever the hell is causing that performance issue, then they still won't be able to reliably catch something like an overclocked Haswell or newer.
Gen-Z Memory Module (ZMM); http://www.smartm.com/products/dram/Gen-Z_ZMM.aspAny new sys memory paradigms on the horizon? When can we expect something beyond DDR4 sys mem?
Yes, 5D0108BBM8SH2_37 . This one would be even stranger for obvious reasons if those three numbers were about the clockspeed, we have only a 010 there! As for the 37 at the end, it could mean either a 3.7GHz base or boost (fixed frequency sample?) going purely off of the old way of decoding these, IIRC. So yeah, I don't think those three numbers have anything to do with the clock, they refer to something else. The clocks of this 12 core sample are therefore a mystery, since it doesn't even have the two numbers at the end like the 8 core did.edit: if you do a Google search on the SKU you'll find evidence that there's an 8c ES floating around as well.