Intel Will launch a 28 core Desktop Processor at End Of Year

I'm not sure either. Guess we'll see.

I wasn't willing to go that far - the 1950X was already stretching it. If I had more money I would like to tinker with a server setup.

 

To bad the trick me down effect rarely works.

 

It's R15.

 

I don't this it impressive, Just imagine that 800W+ concentrated in such small die- it's crazy to think about the cooler they used!.
if we look on Ryzen 7 2700 with it's 65W TDP, then AMD can easily make 32 core TR2 with ~250W that could be cooled by normal AIO water cooler and we can get over 6000 point's in CB R15.
http://www.guru3d.com/articles_pages/amd_ryzen_7_2700_review,9.html
I hope tomorrow AMD will shaw why is called ThreadRipper!.

P.S we aready know about Ryzen 7 2700E with it's 45W TDP and I think it going to be on Ryzen 7 1700 performance levels, so if we take this into consideration TR2 32 cores with TDP under 200W can score over 5000 CB R15.
P.S2: This about how expensive those binned 28 core dies will cost, on AMD side they are just 4 binned Ryzen 7 .
RIP Intel overpriced XEONS.

 

7980x @ 4.6ghz, 16 cores, full load: CinenBench: 550w
AIDA64 stress: 510w

Threadripper 1950X @ 3.95GHZ. 16 cores, full load: CinenBench: 515w
AIDA64 Stress: 520w




There go your "predictions", not to mention how power consumption skyrockets when 2600x, 2700x are OC-ed.

Take this into consideration, so Threadripper @4.6ghz would use even more power and @5ghz would melt.

2600x uses more power than 8700k and it is slower, lower clocked.

But nice try.

And when they are not OC:

7980x stock CinenBench score: 3334, power consuption: 275w

1950x stock CinenBench score: 2985, power consuption: 330w


And If I am right, 7980x has turbo set @ 3.4ghz when all the cores are utilized.

You have to overclock 1950x to 3.95ghz to beat it in CinenBench, then its score is 3415.

7980x @4.6ghz is 4452 while the difference in power consumption is 35w in favour of AMD.


So threadripper isn't ripping that much, is it?

You get now why is it cheaper?

 

Yup let's just pull numbers out of nowhere - what a great way to prove a point. Actually... what exactly is your point? I don't see how Threadripper is relevant to this discussion.
Regardless, I personally haven't been able to find any OC results comparing the 7980X to the 1950X, but if you look at the chart from PC World's article of the 1950X vs 7960X at stock speeds shows the 7960X has a pretty substantially higher wattage, despite having very similar [boosted] clock speeds and the same core count. With all threads enabled, it came out to a nearly 50W difference. It is generally faster, but not in every test and not by a wide margin.
Everyone knows Ryzen's efficiency really starts to suffer once you exceed 3.8GHz. That being said, I'm sure the 7960X (or even the 7980XE for that matter) would start to push ahead when overclocked. But the fact of the matter is these chips already run super hot, and many reviewers have commented how the power cords were starting to get warm. A 28-core is going to suffer even more with heat dissipation and power delivery - that's ultimately the point here.

Again, cite your sources. I checked 5 sources (including Guru3D) and in 3 of them, the 2600X used less power under load at stock speeds. Power delivery dramatically changes depending on your workload. If I were to cherry-pick results like you do, I could show wattage results of AVX2 calculations and be like "hurr durr look how horribly inefficient Intel is!" even though it's actually very efficient considering how quickly it crunches data. It seems you (along with many people) forget that high power consumption is not directly correlated to efficiency. The most important metric to consider is performance-per-watt. But again, that's totally besides the point - the PPW is irrelevant when you've got to figure out how to sustain proper cooling for 28 cores.

So yeah, nice try indeed.

 

I think its the ring bus effect. More cores in the ring = diminishing returns.

 

I stand corrected my score was 1300 not 3100. oops my bad.

 

Power consumption pulled out of (__!__)
Fictional 4.6 and 5.0GHz on TR.
Comparing intel's chip which is 50% more expensive to AMD's in metric which is irrelevant comparison of those 2 chips.
Then topping it by comparing intel's chip which costs 120% more than AMD's chip in comparison in even more absurd way.
And your belief that OCing 7980x enough for it to get 33.5% on score will result in 32% increased power consumption...

I get why it is cheaper. In business, where you have use for multiple threads, you can pay $2000 for intel's chip and it delivers certain performance per watt.
But you can as well pay 2x $900 for AMD's chips and keep them in power efficient clock range, having 50% higher performance at same power consumption.

You do not understand many core business. Moment you have use for them, you really care little if it is 10 or 40 cores. All that matters is price of purchase and cost of delivered computational result.

 

der8auer managed to partially post an epyc cpu on a tr4 board all 4 dies powered on, 1detect pin must be covered, they are definitely not electrically incompatible for the most part, the socket is less different than skylake and coffeelake., also even though it would be lacking memory channels, you can still boot epyc and thread ripper in single channel mode, it just means memory access has to happen over the infinity fabric for any dies not directly connected. 32core threadripper is definitely possible.

 

Maybe electrically incompatible isn't the totally accurate way of putting it, but I'm pretty sure they're electrically different (otherwise, why call them different sockets, rather than just specify which chipsets work with which CPUs?). I'm sure the situation is a lot like how modders got some Coffee Lake CPUs to work on older chipsets - the sockets between both generations are actually different, but not enough to totally prevent some CPUs from working. It is, however, different enough to prevent full backward compatibility.

Anyway, AMD is planning on a 64-core Epyc, so assuming these new Threadrippers are derived from that, they'll still have 2x dies.

 

the only true electrical incompatibility/important difference on tr4 is the detect pin, which tells it if its an epyc or threadripper cpu, thats it.(they are different sockets because tr4 doesnt support 8 channel memory and 128pcie lanes , the extra pins for those simply arent connected to anything)
coffee lake is the same story, the 6 core chips have a detect pin which needs to be shorted inorder to allow it to post, the rest is firmware related, with coffee lake there are also other socket changes , mainly the extra power pins , which are not critical ( it is possible to boot any coffeelake cpu on even z170 boards with the pin short +firmware edit), on tr4 there are no such extra differences, all amd has to do is slightly modify the pcb of the cpu so that it outputs a different signal to that pin and put out a firmware update, and presto you have 4 dies on tr4 , changing the pcb in this way is a really minor thing.

also if i recall der8auer xrayed a threadripper cpu and the pcb has all the traces to connect all 4 dies, which wouldnt make sense to do if the pcb is significantly electrically different and is never intended to run more than 2 dies.

 

I would be surprised if they were to sell it for anything less than $5K per unit.

 

The numbers are authentic, I got them from the kitguru.net review of 7980x, check them, but that's not important. You also missed that it is an 18 core, not 16. Regardless, this was a bait for you if you pick on me once again without checking the numbers as usually. One thing I am curious about and wanted to ask, aren't you concerned about a 32 core Threadripper potentionally having only the quad channel?

 

Seems like what Zen2 would look like.

Is it 7nm? I can't imagine how it suppose to be cooled without serious 7nm magic.

 

https://www.tomshardware.com/news/intel-28-core-processor-5ghz-motherboard,37213.html

THG did some analysis on it. Looks like they were using a water chiller that was hidden under the table.

 

"We confirmed that Intel was using a water chiller in the 5 GHz demo, a Hailea HC-1000B, which is a 1 HP water chiller good for 1500-4000 liters per hour and uses the R124 refrigerant to reduce the temperature of the water to 4 degrees Celsius. Technically this unit has a cooling power of 1770W, which correlates to the fact that a Corsair AX1600i power unit was being used for the system."
By anandtech.


SaLaDiN666 - where are you hiding? we got news for you: TR2 32core @250W !

So yesterday I was spot on .

 

My god, you actually have to actively chill water so it keep up with heat. Not air cooling, not water cooling, but freaking chilling the water in pipes.

I guess it still does make sense, since people who are willing to pay 4 digits $ for this CPU, will probably get active water cooler.
Total system cost might end up being 5 digits $.

 

Sorry for missing one nonsense...
And who are you again? If you feel need to create bait for me specifically to find out if I am chasing you, then you have problem. Not with me. I can't care less, I just call BS when it is just too big to let it be.
If you wanted to ask, you could have done it in 1st place. Not really, Since I expect AMD to improve on memory compatibility and therefore it will be quite OK.
TR1950x (3.4GHz base) has 2666MHz Quad Channel IMC and is limited to 180W at stock. 32Core (3GHz base) limitation will be 250W. With that in mind, you can safely presume that bandwidth requirement per core will be lower than 16Core chip. And may be sufficiently fed by 2933MHz memories. But in pure mathematical look at it TR1950x has 60% higher memory bandwidth available per core. (That's worst case scenario.)
From what you can see in comparison of 1700x/2700x and TR1900x, ... TR1900x does not really benefit from Quad Channel IMC, that means 8C/16T chip is not bandwidth starved with 2Channel memory. It benefits from IF OC which happens due presence of faster memory.

That means 32C TR will have up to 37.5% less bandwidth per core once all cores are fully utilized in memory bandwidth dependent task. If TR1950x was hitting wall, then this wall would be felt more on 32C TR, but you do not get to see articles all over internet about TR being bandwidth starved. You see almost exclusively praise. Get one, run geekbench at constant CPU clock and decrease memory bandwidth in steps. I'll gladly calculate for you if any of those workloads there are bandwidth limited on TR.

On intel's side GB did reveal that certain workloads will tank badly on intel's 8C/16T if paired with dual channel w/o crazy fast memory. But that's with highly clocked chips. 4.5GHz all core OC on intel's chip is 50% more than 3GHz on all core clock in that 32C TR.

 

Sure they will have... 5-15% of performance increase by not using tooth paste but soldering their CPUs again, making 6 cents less profit per 1000$ CPU. It's just not anything really new.