28 - Core Intel Demo Questioned on Cooling

haste. said:
Can't really argue any of that. I do think that we should be expecting 2 to 3 years instead of the 4 to 5 you stated, but that's for the basic arch and then they have to get the silicone to market. So assuming he got to work right away, arch done by 2020ish and to market no later than YE2023 (and that would be very late). I'm sure Intel was working on something and brought Keller in to optimize seeing what he did with less resources and crunched for time at AMD.
Click to expand...
By 2023, x86 might be declining in market share to the point Intel may never recover in this market.
Amd on gpu and cpu fronts are going massively threaded/multi die and SOC. Bit like Phi. Amd can see the writing on the wall, adapt or die.
In a decade or two we may be running SOC entire systems the size of an m2..
 
DuronBurgerMan said:
Easily is a stretch. TR2 will have memory latency and bandwidth issues because two dies will have disabled memory controllers.
Click to expand...
Adding to what the others already said, AMD does give us a hint of how much faster the 24 core TR2 is compared to the Intel 7980XE, and thus allow indirect comparison to other CPUs.

In the AMD Computex presentation, they pitched a water-cooled Intel 7980XE with 3200 MHz memory against an air-cooled 24-core TR2 with 2666 MHz memory.
The rendering is first shown around 1:15:09 mark, with the Intel CPU maybe getting 1 second or so head start.
AMD finishes at 1:15:37 at 29 seconds according to AMD commentary
Intel finishes at 1:15:46 at 38 seconds (or 39 if we assume 1 second headstart)

The AMD system is around 30-35% faster in this situation.

From earlier benchmark data we know that the 7980XE beats the 1950X by around 16% in Blender (for example here: https://www.golem.de/news/core-i9-7...8-kerne-um-amd-zu-schlagen-1709-130083-3.html with 2666 MHz memory on Intel side and 2400 MHz memory on AMD side).

This means overall, in this test setup, the 24 core TR2 performs roughly 50-55% faster than a 1950X which would be very much in line with the increase in core count, if we assume identical clocks.
 
KazeoHin said:
The IF has a really good bandwidth throughput, its the latency which it effects, so in things like offline rendering I wouldn't imagine a huge loss in core scaling.

If the data being processed can be done entirely in cache, you're looking at perfect scaling.
Click to expand...

It's more a case of the bandwidth having to be shared between more cores. Same memory bandwidth as the 16 core version. Twice as many cores fighting for that bandwidth. The latency for the two new dies will be an issue for sure.

But I am sure AMD did their best to mitigate this, and on tasks that are not latency sensitive, or on data that can be processed in cache (like you say), things will be better. Nonetheless, when the benchmarks come in, I'm still expecting two die TRs to scale better than four die TRs. How much of an effect this will really have... hard to say until we see benchmarks.
 
DuronBurgerMan said:
But I am sure AMD did their best to mitigate this, and on tasks that are not latency sensitive, or on data that can be processed in cache (like you say), things will be better.
Click to expand...
But we were talking about achieving 6000 points in Cinebench R15. The characteristics of the Cinebench R15 benchmark heavily favor multiprocessing scenarios and are usually not very bandwidth/latency sensitive.

Based on what I wrote above for Blender, I believe that beating a Cinebench R15 score of 6000 is quite doable for 32-core TR2, and I would be surprised if it falls seriously short of that.

Edit: Found this, someone on LTT forums tested Ryzen 1600 in single-channel DDR4-2133 vs. dual-channel DDR4-2666: https://linustechtips.com/main/topic/774946-small-ryzen-ram-test-8gb-2133-vs-16gb-2666/

Results:
Cinebench R15: ( i only ran this once, and with priority set to realtime.)
SC 8gb 2133 Score: 1136
DC 16 gb 2666 Score 1152
Click to expand...
The results barely changed at all. So latency and bandwidth apparently do not affect the Cinebench R15 benchmark.
 
Last edited:
chithanh said:
But we were talking about achieving 6000 points in Cinebench R15. The characteristics of the Cinebench R15 benchmark heavily favor multiprocessing scenarios and are usually not very bandwidth/latency sensitive.

Based on what I wrote above for Blender, I believe that beating a Cinebench R15 score of 6000 is quite doable for 32-core TR2, and I would be surprised if it falls seriously short of that.
Click to expand...

Probably. Let's not forget that Cinebench is a benchmark, though. Real world rendering and encoding tasks are what these things are really going to be used for - and in some of those tasks, the scaling is likely to be worse. I don't want to seem like I'm shitting on AMD for this, though. They found a way to get 32 cores working on TR, and that's definitely not a bad thing. I just want to be careful about how far the hype train goes before we see the real benchmark figures.

It's my hope that AMD figures out a way to get 8 channel working on future motherboards (obviously the traces are all wrong for the existing ones), and then this won't be a problem.
 
DuronBurgerMan said:
Probably. Let's not forget that Cinebench is a benchmark, though. Real world rendering and encoding tasks are what these things are really going to be used for - and in some of those tasks, the scaling is likely to be worse. I don't want to seem like I'm shitting on AMD for this, though.
Click to expand...
Oh, I think we can all agree that there will be scenarios where memory latency and bandwidh matters. AMD themselves admitted this, and recommended Epyc for these situations (see Anandtech article).

I just wanted to get this discussion straight:
chithanh said:
So Intel feared that a 32-core TR 2000 beats 6000 points easily, and this had to be prevented at all cost.
Click to expand...
FLCLimax said:
This score is in line with the ASUS setup on regular water cooling. 6K give or take, which TR2 will easily beat.
Click to expand...
DuronBurgerMan said:
Easily is a stretch. TR2 will have memory latency and bandwidth issues because two dies will have disabled memory controllers.
Click to expand...
Because there is nothing which makes me expect TR2 to miss the 6000 mark, on the contrary, anything seriously below that would be quite surprising.
 
chithanh said:
Oh, I think we can all agree that there will be scenarios where memory latency and bandwidh matters. AMD themselves admitted this, and recommended Epyc for these situations (see Anandtech article).

I just wanted to get this discussion straight:


Because there is nothing which makes me expect TR2 to miss the 6000 mark, on the contrary, anything seriously below that would be quite surprising.
Click to expand...

Maybe. I don't know. If it doesn't happen, we'll at least know why.

Intel's shit was lol, though. They clearly were afraid of Threadripper 2 stealing the show.
 
FLCLimax said:
So...the chiller unit that Intel used for their dog and pony show is illegal in the US and EU. LMAO! It uses the banned R22 gas.

https://www.out-law.com/en/sectors/...ir-conditioning-costs-and-the-banning-of-r22/

Hilarious. Would they even have been able to stage such a stunt at a Tech Show in the US? Too funny.
Click to expand...
Well, still not illegal per se in USA till 2020...After that date R22 systems cannot be imported into the US or manufactured here.


As per the US Environmental Protection Agency, R22 will become illegal in the United States on January 1, 2020. After that R22 refrigerant phase out date, R22 can no longer be manufactured or imported into the US.
 
Hagrid said:
But it is expensive as all hell. We would just replace the units with 410a systems.

They must of grabbed an old system real quick. Just as fast up as it was put down! :) I really did not think they would stoop to such a low level. It really shows that AMD somehow put them on guard again?
Click to expand...
Not really the point, however, that unit is still sold with R22.
 
alamox said:
i don't get it, intel knows AMD got the better architecture, scallability, cost, efficiency, they are just better equiped for core wars, and the outcome is easy to predict from ryzen and threadripper lineup, so why does intel double down and initiate the core war knowing fully well it will make them look dumb, why not play on their strenght ?
reasonable amount of cores, but push the clock and optimisation , so even if they lose in benchs they would still be able to spin it, that their 18-20 core cpu with less cores can still beat AMD in gaming, and somewhat get close in workloads, increasing the value of their cores and cheapning AMD's.
but what Intel si doing rightnow is just stupid and shortsighted, their pride cannot be that oversized, it's better to lose with dignity and spin the loss, than to(yet) lose and make yourself look stupid in the process.
seeing the last couple years i honestly think i could make a much better job if i ran intel PR departement.
Click to expand...
I don't think you or many Intel fans realise, AMD since ryzen has them beat on efficiency in most cases. They used nearly 1/3rd of the power of Intel's 7999999x ten core whatever last cpu they tried to fight with, and this was just 1800x not even TR, it was about 70-80%+ of the performance oc to oc and under half the cost witness cores, the latency was no better on intel either.
So now the problem is they can't beat AMD on per core with this many cores, hence they pulled a Damn chiller out!
Not to mention yield rates for Intel will be laughable. That's why they stormed out like a kid who's been found out when AMD dropped an air cooled 32 core that has similar performance to their chiller cooled mega oc 28 core. Make sense now?
Tldr: Intel is beaten in most cases where power matters beyond 6+ cores already. Badly.
 
Splorch said:


Paul's Hardware gives a nice close up of the components and the mobo. Correction* 32 (4x8) pin power for the cpu. Apparently 32 phase power delivery. Absolutely insane.
Click to expand...


Running on Windows 8?
 

The 28C chip Intel used during demo was a Skylake Xeon with modified microcode, not the real HEDT chip (Cascade-Lake) that will be launched. The Xeon used on the demo is a 14nm+ chip with ordinary TIM. The future HEDT chip is 14nm++ with STIM (Solder TIM).

The exotic cooling was used because the demo didn't even use a sample of the real product, but simulated the performance of the future product with a current Xeon. Intel pushed up that Skylake Xeon to 5GHz to demonstrate performance of the future 28C HEDT chip, not because you will need that kind of setup for the future Cascade-Lake chip. There are rumors that stock turbo will be 5GHz for the 28C part, just as in current i7-8086k. And Intel confirmed the real product will overclock to 5GHz


Apparently there is also a Cascade Lake AP series coming (AP = Advanced Processor) but it seems to be destined to the HPC market.
 
Nightfire said:
Well whoopdy do. I am sure AMD has alot of neat products in development as well. The product AMD showcased, however, will soon be launched.
Click to expand...

That AMD product targets Q3 launch. The Intel product targets Q4 launch.

TR2_Slide_Car_678x452.jpg


https://www.anandtech.com/show/12893/intels-28core-5-ghz-cpu-coming-in-q4
 
You must log in or register to reply here.
Back
Top