Howdy, Stranger!
It looks like you're new here. If you want to get involved, click one of these buttons!
Quick Links
New AMD 3000 series
Been reading a lot of speculation about their 7nm CPU's In this thread someone mentioned that they have the new entry level 6 core 3300 running at 5ghz on air cooling.
https://www.hardocp.com/news/2019/01/02/ryzen_3000_processors_listed_on_russian_retail_site/
If any of this is true, Intel is going to be in a world of hurt with its server business.
Looks like it is time to build a new system when these become available. If the price is reasonable, I can see a 3300 system in my future.
https://www.hardocp.com/news/2019/01/02/ryzen_3000_processors_listed_on_russian_retail_site/
If any of this is true, Intel is going to be in a world of hurt with its server business.
Looks like it is time to build a new system when these become available. If the price is reasonable, I can see a 3300 system in my future.
Comments
Next week will be announcement, not launch.
That being said - the G APUs rumored in Q3'19. No note of what the graphics cores would be, but good guess on at least Vega, if not Navi. Wonder how far those would be from what is anticipated to be in the next gen consoles.... XB1X right now has 40 CU (possibly Polaris) graphics and 8 Jaguar CPU cores @ 2.3Ghz. PS4Pro has the same CPU cores @2.1Ghz and 36CU graphics.
Jumping from Jaguar to Zen3, and from Polaris to ??? (probably Navi) .... probably looking at something a lot like what is listed here, with a lower CPU clock and more GPU cores to be able to fit the TDP budget (probably between 60W and 90W, give or take and based on console).
Intel could have remained competitive in the desktop market if Zen 2 simply couldn't clock very high. That was unknown, and to a considerable degree, still is, as these are just rumors. But if Zen 2 cores can match Coffee Lake Refresh in single-threaded performance, in addition to having much better well-threaded performance, then Intel's desktop market will also be in a world of hurt, not just their server market.
And if Zen 2 can run at extremely low power at idle, then Intel's laptop market will also be in a world of hurt. But there aren't even rumors about that yet. If not, then Intel could remain the preferred vendor for a considerable chunk of the laptop market by virtue of having longer battery life when idle.
But one thing at a time, and if AMD had their choice between dominating the desktop, laptop, or server market, they'd choose the server market. And did, more or less. It's arguably better to think of Zen 1 desktop parts as being 1/4 of a server part rather than an EPYC CPU being 4 desktop parts.
https://videocardz.com/newz/amd-ryzen-9-3800x-matisse-listed-with-16-cores-and-125w-tdp
Or at least the ones listed as being announced at CES do.
Remember that x86 CPUs have an enormous markup, as they're not large dies at all. The die in a Ryzen 7 2700X costs about as much as the GPU die in a Radeon RX 580. Yet AMD charges a lot more for the former than the latter, even though the video card costs a lot more to built, as it also needs a big PCB, some VRMs, a bunch of memory, and other things to make a completed video card.
Certainly, if AMD has an Intel-killing part, they're not going to sell their top end for $200 as they did with the FX-8350. But a top end Socket AM4 part costing $500 is plausible, as that probably only costs something like $50 each to build. AMD doesn't want to have an awesome part and then price themselves out of consideration from most of the market.
Remember that the higher end game consoles aren't relying on DDR4 to feed their GPUs. They're relying on GDDR5. That's able to get plenty of bandwidth to the GPU. The problem is that it caps you at 8 GB of total system memory. That's very limiting for a desktop. It also means high idle power consumption, which is a complete non-starter for a laptop.
If AMD does build a big integrated GPU with a stack of HBM2 on it, I don't see them selling that for $200. AMD might try to sell it for $400, claiming correctly that it matches the combined performance of a $300 CPU and a $300 discrete GPU, so it's a great value. But that would be a much higher end part than the $200 and under G-series parts rumored above. It might also take a different socket.
At least until the next one, if 5 nm ends up being junk, too.
The Russian site put in the same info minus the 3850x. It could be a placeholder using the leaked material above. You’ll see the videocardz reference to adored TV (that’ll be the YouTube channel where I posted a video before that you didn’t like because it was 30 minutes long lol).
I think that the leak from Adored has some merit, a lot say it’s too good to be true. Should be exciting next week with the keynote from Lisa Su!
One thing that might concern AMD in the future is Samsung. 7nm production started last October, 5/4nm planned for this year, 3nm for next year.
As for Samsung, it's not clear that they offer process nodes appropriate to high performance hardware. AMD used Global Foundries' 14 nm, which was basically just a licensed version of Samsung's 14 nm, because they were contractually obligated to do so, not because they wanted to. It seems likely that for high performance hardware, it was inferior to TSMC 16 nm.
For starters, Nvidia could have used GloFo/Samsung 14 nm if they wanted to, but didn't. Furthermore, at the end of 28 nm, AMD and Nvidia had about as good of architectures, with Fiji (Fury X) and GM200 (GTX 980 Ti) close in the various efficiency metrics. You could argue that Nvidia was slightly ahead, but it was close.
Then they did a die shrink of essentially the same architectures, with AMD moving to Global Foundries' 14 nm and Nvidia to TSMC 16 nm, and then it wasn't close, with Nvidia way ahead. Either AMD botched the shrink with Polaris/Vega, Nvidia botched Maxwell on 28 nm (which seems unlikely, considering how large of an advance it was over Kepler), or TSMC's 16 nm process node was a lot better for high performance GPUs than Global Foundries' 14 nm.
That's not to say that Samsung's process nodes are junk. They're simply targeted at lower power parts, such as cell phone SoCs. Samsung's 7 nm EUV process node is in production today, and both AMD and Nvidia are free to use it if they want. They just don't want to.
-------------------------------------------------
Samsung's new fabrication - to early to know what it is capable of. Or how it stacks up with TSMC's 7nm. Probably have to have multiple products from both plants. Or how this might change with any die shrinkage - to 5/4nm say.
Then its a case of what IBM, NVidia (is this how they plan to respond to AMD on 7nm?), Qualcomm etc. And - of course - Samsumg themselves.
Until we see products though I don't think we can make statements about what is / is not possible.
And then - as with TSMC's fab - there is the question of cost. The suggestion that Samsung may, in part, produce chips for low cost phones and tablets however may indicate that anticipated run costs may be "acceptable". (Set up costs are huge of course).
I also think the pricing is placeholder as a test to see reactions from leaks.
Regarding silicon and a summary of AdoredTV's analysis (but would recommend to watch his videos) -
- Ryzen gets all the stuff in between (defective cores or low clocks), which with 2 dies per socket module means defects on up to 1/2/4 cores die can still be used and can be some of the super high clocking ones that would never make it into a console or server. Thus combos of 4/6/8/12/16 cores on the Ryzen platform is possible.
I think next week we'll see the x570 boards guessing it will raise the board design guidelines and revises the chipset itself with PCI-E 4.0 support. Mainly notable from the aspect of being able to bump up the supplementary lanes from the chipset from 2.0 to 3.0. Board availability wouldn't have to go hand in hand with actual Zen 2 products so that would be cool to see. PCIe 4.0 support is pretty big for anyone using more than 1 M.2 SSD with some situations saturating the bandwidth of PCIe 3.0 (copying files from one SSD to another).Stream/Casting certainly doesn't take as much horsepower as an AMD APU would likely have -- not that AMD couldn't make one, just that any number of already-existing ARM variants already have enough umph, and the Shield already can do so in 4k.
As an Android mini-console.... well, ok there you could use more horsepower, but your still limited by Android. Why not just do something like a NUC and make a full-fledged HTPC out of the thing?
AMD's chiplet strategy for the next generation is going to be different from this one. This time, AMD had ambitions of making server chips with 32 x86 cores, 128 PCI Express lanes, and 8 DDR4 channels. What they did is to make a multi-chip module where four dies each had 8 x86 cores, 32 PCI Express lanes, 2 DDR4 channels, and three infinity fabric links to connect the dies.
Once you have a die with 8 x86 cores, 32 PCI Express lanes, and 2 DDR4 channels, that has what you need for a desktop part. So you can use it in desktops, and AMD did. Threadripper uses two such dies, and came unusually late in the development process, rather than being something planned several years ahead of time as most major products are.
For their next generation Rome server chips, AMD has ambitions of making server chips with 64 x86 cores, 128 PCI Express lanes, and 8 DDR4 channels. This time, they're not making all of the chiplets identical. Rather, they're making 8 chiplets that each have 8 x86 cores, an infinity fabric link, and no PCI Express or DDR4. They're also making a giant I/O chip on a 14 nm process node with 128 PCI Express lanes, 8 DDR4 channels, a bunch of infinity fabric links, and no x86 cores at all.
I'm going to go not very far out on a limb and predict that the giant I/O die will not be used for mainstream desktop parts. But the small chiplets with the x86 cores might not, either. Eight x86 cores and an infinity fabric link isn't what you need for mainstream consumer parts. At minimum, you need DDR4 controllers on the die to keep the memory latency down. Having PCI Express on the die makes sense, too. Having to take multiple hops to get to memory kills your latency.
Yes, it causes latency problems for servers, too. But if you try to make a giant die with 64 x86 cores, 8 DDR4 channels, and 128 PCI Express lanes all on an enormous, monolithic die, yields are going to be terrible at best. It might even be impossible to manufacture such a die at all, as process nodes have caps on how large dies can be.
AMD went with a chiplet strategy for the sake of yields in ambitious server chips. Once you're going to use multiple dies anyway, it's just a question of how you want to do it.
But for mainstream consumer desktop parts with 8 x86 cores, 32 PCI Express lanes, and 2 DDR4 channels, you can fit all of that in a single die without causing yield problems. That took under 200 mm^2 on 14 nm. On 7 nm, it would probably be closer to 100 mm^2--not a problem. If a single die with everything results in the die being pretty small, then that's what you do. You only jump to a multi-chip module if the total die size that you want is awkward to fit in a single die.
TL:DW
Highly unlikely these specs are 100% true or even come close.
We will just have to wait and see, but I am thinking these leaked specs are not that far off.
4 cores: Ryzen 5 2500X: 4 GHz
6 cores: Ryzen 5 2600X: 4.2 GHz
8 cores: Ryzen 7 2700X: 4.3 GHz
12 cores: Ryzen Threadripper 2920X: 4.3 GHz
16 cores: Ryzen Threadripper 2950X: 4.4 GHz
They could easily make a quad core with a max turbo of 4.4 GHz if they wanted to. They just don't want to, as they want you to have to spend more for their best single-threaded performance.