Separate names with a comma.
Discussion in 'AMD Processors' started by DuronBurgerMan, Dec 4, 2018.
Take this with as much salt as you like. It does seem semi-plausible though.
The best part is at the end . Since "we" expected good news about Ryzen 3000 series already.
not giving it a click. theres not going to be a 16 core over 5ghz chip.
Video says 16 core 4.3 base, 5.1 GHz max single core turbo on halo ($500 135 watt TDP) part. Also, will be three chips. One I/O die, and two 8 core chiplets.
We all want 16 cores 32 threads with a (even a single core) speed of 5ghz. But it'll never happen. I'd upgrade at least 3 of my machines if this was true.
Not really I'm not looking forward to having to water cool that cpu at 5ghz.
I'm currently cooling a single core 5ghz processor... I don't have any problems.
Yeah the TDP part is rather unsettling. All the time the people were beating the crap out of AMD for not being power efficient . The AMD forums around the world were hell with those smart Intel people scolding that it uses to much power.
So don't buy it? The rest of us will just buy adequate cooling...
The question is if AMD can make an 8 core chiplet that can hit 5.1 GHz single core boost. If that's plausible, then 2x chiplets + i/o die may very well work out as the leak says.
If they can't, obviously the leak is bogus.
It's a leak, so it's just discussion. I neither believe or disbelieve, but it's PLAUSIBLE that 7nm could bring a ~5GHz single core turbo. And I highly suspect the chiplet + smaller i/o die than Rome is actually true. It would make sense for AMD to be manufacturing the same chiplets for all CPUs they can. So 16 cores is not impossible. It's more a question on if those two features can be combined. I dunno. Maybe?
Even bigger question: what is the utility of sixteen cores without significantly increased memory bandwidth?
At some point, which is perhaps testable, the workloads that will put sixteen cores to good use will be limited by bandwidth.
The same thing was said about Prescott back in the day.
How many tests are there out there where the quad channel RAM and the added bandwidth makes very little difference in many applications? Plus, there's the HEDT platforms for that.
Realistically, 16C is overkill for mainstream use (right now anyway), but if AMD has a 8C/16T 5Ghz part for $229 and Intels is $490+ and they are within 10% of each other. There's really no reason to spend the extra for Intel. And it's a DROP IN REPLACMENT with a bios update.
The other interesting part of that video is the Navi info and Vega 56 performance in the RX580 price point and Vega 64+ performance at the $250 price point with less power.
Well, supposedly zen 2 has double the L3 cache, so that will help alleviate the bandwidth issues (higher hit rate, and possibly higher bandwidth).
But they will still need to bump the officially supported DDR4 clock to not starve the chips. 'I'm pretty sure they could do 3200 without gouging people for the memory (2700x supports 2933 in stock mode).
It was interesting that Adored suggests the 6c, 8c, and 12c Zen 2 will drop in to X370 and X470. But 16c will only work on X570, due to increased Threadripper-like power requirements.
Which is strange because if you load up an all core OC, you're probably way beyond the 135W that the 16C is at already and current boards handle it. It might come down to a case by case basis with each board where higher end ones get support, and lower end ones don't.
The number of consumer applications that really benefit from sixteen cores over two is pretty limited. If you're not buying sixteen cores just because you can, you're buying because the extra resources actually make a difference in those few applications. This is even more true as core counts rise.
Which gives reason for concern if they're cranking core counts but not feeding memory bandwidth; many of the applications that would make positive use of the cores, like say VMs, would be better suited to an HEDT or workstation platform with more PCIe lanes and ECC.
...and your off-topic distraction reveals your intent.
The 12 core part looked the most interesting to me.
So comparing CPU's between brands is off topic? That doesn't really make much sense when discussing CPU's.
That'll help a little bit, but hopefully they have faster memory support planned for those that would actually want to use those sixteen cores.
"Hey look at this off-topic thing, just to prove my bias and that I'm more right!"
Sorry, yeah, we're not comparing brands, we're talking about high core-count CPUs on a consumer socket, and here you going on about low core-count CPUs as a distraction.
Welcome to the lose/lose conundrum of forward compatibility
You're just being ridiculous. The video had CPU's all over the spectrum and the topic is "New Zen 2 Leak" which includes everything from the 3300 to the 3850X. The 8C/16T part is in there. As you said, not everyone needs a 16C part, so it's clearly relevant. I don't understand why a chip that performs similar to the $490 Intel part at half price is a bad thing for consumers. I don't give a shit about Intel or AMD. I like performance per dollar .
Well, I mean I like their version of it compared to Intel's (admittedly as a bitter Z270 chipset owner ).
I didn't reference the video, purposefully- we know the lower core-count products will be effective, with near certainty. At least I have faith that AMD will execute. I'm referencing the higher core-count parts specifically because I'm wondering where memory bandwidth limitations will kick in.
They had two choices- forward potential compatibility, or a new product line with guaranteed compatibility. AMD has gone the route of potential forward compatibility, and I understand why they did whereas Intel didn't. I see Intel focused more on OEMs, and AMD, for the moment, working to garner as much enthusiast sympathy as they can. And.. it's working
Seems plausible on the leak and Adored TV tends to be pretty accurate.
I'm just saying, DDR4 is still largely untapped. You could bump the supported speeds up to 3200 and 3466, and handle the new memory hunger completely.
IT's been a year, and DDR4 3200 is now cheap. Stock speed, it would give you a 20% bandwidth bump over the 2700x. That, plus improved cache should be enough to cover most processing loads.
It's the 'should' that I'm looking at. I get that Zen2 should be an overall upgrade, I'm just wondering where it would fall short of a 2950X due to memory bandwidth, which despite Zen2 likely moving up in terms of memory speed, will still fall significantly behind.
Ryzen 3000 Series is literally a miniature version of Rome.
It contains two chiplets and 1/4 I/O die.
Alternatively, one of the chiplet can be replaced with an iGPU
If this leak is true (IF), it's what I've been saying since the Rome reveal. It would make a lot of sense. Make one chiplet die. Bin and segment. Make two i/o dies, a big one for Epyc & TR, and a small one for AM4. Make AM4 use two chiplets OR one chiplet and one GPU. Cheap way to salvage semi-shitty GPU dies, too.
It all makes a lot of sense from AMD's perspective. But who the hell knows. I'm always skeptical of leaks.
So AMD is launching 7 CPU's in 4 weeks and the only leak we have is this one? Nothing on SiS or any other CPU bench software? No mobo/chipset leaks?
And despite the other X570 leak showing Computex?
"Take it with a grain of salt" = I'm talking out of my ass for easy clickbait.
I will say if this is accurate my Intel parts are going on Ebay day one. Won't even bother waiting 6+ months for Ice Lake.
I dunno Adorned usually is pretty good with its leaks and sources, I Would believe them over Semiaccurate or Wcctech....But who knows, we will find out soon. Lets just say we KNOW AMD has 7nm ready, just how ready is the real question.
Hopefully the GPU 'die' has some local DRAM.
If Zen 2 hits 5.0ghz on 8, 12, or 16 cores at those price points then intel is in trouble. I'd immediately switch from my 7820x. For 9th gen HEDT Intel just soldered the IHS, slightly more cache, and more Lanes on the same architecture no IPC change. Intel is so lame.
Which is still more than AMD is doing...? And if you go above 16 cores on Threadripper you have to contend with NUMA. 10nm vs 7nm might be different, but that's what we have today.
Hopefully Intel has their next process in gear. If not, we might all be running AMD...
Where did it say its got all 16 cores at 5ghz? I dont see any. Totally believable. Keep in mind this isn't single die with 16 cores. Its chiplet design. Heck if with OC they can all reach 4.5ghz. That's money right there still.
I honestly suspect them to get up to 4.7 ish.
3600X or 3700X looking very appealing at those prices.
My MSI X470 AC7 WIFI actually has 2 8pin CPU connectors already. So I think for the 16 core chips they will likely have a fine print about that. I see no reason why it won't work with my board as it has 2 8pin connectors which I have connected already. I think they will likely say check your motherboard vendor for compatibility on the 16 core chips. AND my VRMs run cool as ice on the board. LIke in the 50-60c range under AIDA64.
Man if those NAVI performance numbers and pricing are true That shit is nuts. It really looks like they are over delivering on the 7nm and if they are getting that much performance from navi 20 at 150 watts. I think Lisa might not have been joking when they said they will be there competing in the high end. Imagine a full navi chip in second half when the process is even more mature and they can release a Navi 20 with full die and 250w. That could literally be competing with 2080+.
Rumors about Navi is looking better than they expected seems to be coming to fruition. I will keep expectations low though but that naming scheme looks like something AMD would do haha, like they did with intel. Also lisa usually doesn't hype things but even though she was careful with her words, it seems she was fairly confident in what they have in the labs.
Don't most chips generally do all-core OC that matches their stock single-core boost? At minimum. So the boost clocks here would represent a theoretical max all-core OC.
Hopefully the extra cache mitigates the extra latency of ALL memory being across a inf fabric connection on the separate IO die. uniform access to all memory on single socket boxen at least, would solve window's shit scheduler with NUMA.