• Hey Guest. Check out your NeoGAF Wrapped 2025 results here!

Intel Knights Landing: How does 16GB eDRAM as a cache and 500GB/s bandwidth sound?

Status
Not open for further replies.
T

tipoo

Banned
http://techreport.com/news/25848/real-world-tech-analyzes-knights-landing-rumors#0

Intel has revealed relatively few details about Knights Landing, the next-gen processor set to join the Xeon Phi family. This massively parallel chip is designed for high-performance computing applications, and it's set to be manufactured on a 14-nm process. Knights Landing will feature on-package memory and support 512-bit AVX instructions. It's also slated to be available as both a standalone processor and on a PCI Express card.

That's about all we know for sure. However, there are plenty of rumors and leaked slides circulating around the web. Real World Tech's David Kanter has sifted through them all and come up with some informed speculation on what Knights Landing entails. In short, it looks like a beast.

According to Kanter, Knights Landing is likely to have up to 72 individual cores based on a modified version of the Silvermont architecture. The current generation of Bay Trail Atom chips is also based on Silvermont, but the implementation destined for Knights Landing will need to be tweaked to support AVX-512 instructions. These modified Silvermont cores are expected to be arranged in dual-core "tiles" that feature a shared L2 cache that Kanter suspects will weigh in at 512KB.

Knights Landing's embedded DRAM will total up to 16GB, according to slides leaked by VR-Zone. The eDRAM will reportedly be split between eight individual controllers and boast 500GB/s of sustained bandwidth. On top of that, Knights Landing is said to feature a six-channel interface that supports up to 384GB of separate DDR4 memory. Kanter says those memory specifications are accurate, and he suspects Intel's QPI interface will serve as the system interconnect.

Even though it will be fabbed on a 14-nm process, Knights Landing is expected to be a huge chip. Kanter suggests the silicon will measure 700 mm²—nearly three times the size of Ivy Bridge-E, whose silicon has a 257 mm² footprint. Ivy-E has only six cores, but each one is much larger than Silvermont. Also, Ivy-E is fabricated on larger 22-nm process.

Knights Landing isn't due until 2015, so it could be a while before we have definitive details on the chip's specifications.
Click to expand...
 
Read about this a bit ago. Sounds super powerful and incredibly expensive but damn if I'm not impressed by the potential.
 
NotMyRealName said:
Who is this intended for?
Click to expand...

Supercomputers. I think the fastest supercomputer right now uses Xeon Phi (which had the codename of Knights something, so it's a precursor to this.)

I think it's interesteing that Knights Landing can either be stand-alone or add-on via PCIE...
 
After reading that, I really have no idea what to do with this boner. Thanks for nothing OP :-|
 
Would this be of any use in a gaming PC tho?
I imagine games wouldn't take advantage of the extra shit, but how would it compare to a top of the line i7?
 
The whole industry needs to go to massively parallel computing as vast as possible.

AMD have the right idea by focusing on more cores, rather than die shrinks. Even though I'm really looking forward to Intel's 14 nm chip this years.
Many gamers don't understand the benefits, as their games run better on fewer Intel core, than AMD's focus on more cores. But a 4 core market focus isn't sustainable. And massively parallel process, is the only way to keep up Moore's Law. And developers are going to have to become used to developing for many cores.

The mind is more powerful than computers of today, because it is a massively parallel system. Computers should go down the same path. It will make them much more powerful from learning from the mind. This is one reason I'm happy that new consoles are focused on AMD multi-core processors. As this means more software will be focused on multi-core development.
 
The eDRAM will reportedly be split between eight individual controllers and boast 500GB/s of sustained bandwidth.
Click to expand...

That is an insane amount of memory throughput.... bearing in mind the fastest standard DDR3 is around 17GB/s. Will there be any reason to even put ram in a normal desktop?
 
HolyBaikal said:
The whole industry needs to go to massively parallel computing as vast as possible.

AMD have the right idea by focusing on more cores, rather than die shrinks. Even though I'm really looking forward to Intel's 14 nm chip this years.
Many gamers don't understand the benefits, as their games run better on fewer Intel core, than AMD's focus on more cores. But a 4 core market focus isn't sustainable. And massively parallel process, is the only way to keep up Moore's Law. And developers are going to have to become used to developing for many cores.

The mind is more powerful than computers of today, because it is a massively parallel system. Computers should go down the same path. It will make them much more powerful from learning from the mind. This is one reason I'm happy that new consoles are focused on AMD multi-core processors. As this means more software will be focused on multi-core development.
Click to expand...

Die shrinks and more cores are the same side of the coin. The smaller you can get a singular processor the more you can fit on the same die. And then you end up with less power draw and heat production which means you can scale higher.

Fewer efficient and faster processors can be better than many less efficient ones, GPU's notwithstanding since it's a slightly different situation. That and there's only so many processes you can throw at a problem until you end up hitting a wall where you won't get any more performance no matter how many processors you throw at a problem.

Side note: Intel can absolutely push just as many cores as AMD and they do, but there is absolutely no reason for them to do so for the general market since they have a near monopoly at the moment.
 
Sir British said:
Woah there, developers just got the hang of multithreading. Baby steps.
Click to expand...

Doesn't matter for extremely parallel workloads. A developer doesn't have to think about exactly how many cores are in a GPU to use it well for OpenCL or CUDA. Same concept. Supercomputing is already highly parallel by nature and necessity. THey won't be sitting around thinking about how to split up 72 threads, part of it is automagical at this level.


HolyBaikal said:
AMD have the right idea by focusing on more cores, rather than die shrinks. Even though I'm really looking forward to Intel's 14 nm chip this years.
Click to expand...

That doesn't make sense, die shrinks enable more cores to be put on the same chip size.
 
Don't even care who it's for. Bring it out and push tech advancements like this. Those specs sound mighty fine to me.
 
SRG01 said:
700 mm^2 on a 14nm process? Yields must be terrible.
Click to expand...

For supercomputing, they can charge a mighty markup to cancel out those low yields. And what else is there like it? Sure, GPGPU, but nothing with x86 compatibility with other general processor features.
 
terrisus said:
Bitcoin miners
Click to expand...

Scrypt for Litecoin is extremely bandwidth dependant, more so than Bitcoin, I'm salivating wondering what a chip with access to 16GB 500GB/s sustained bandwidth just for an eDRAM cache plus more DDR could hit in that.
 
thecharrr said:
Looks like I know what I'm putting in my next rig when I win the lottery.
Click to expand...

Actually for games it won't be of much help since most barely use more than 4 threads and the individual cores here are relatively weak Atom-derived ones - there's just a LOT of them and they have tons of cache bandwidth.
 
This kind of makes me lament where we may have been if by now Larrabee had been released and was three or four generations revised as a graphics oriented discreet part from Intel. I wonder what that could have been like.
 
tokkun said:
The only workloads I can think of that would benefit from 16GB of cache would be database-related.
Click to expand...

Supercomputing of all sorts. Weather simulation, fluid simulation, radioactivity simulation, quantum physics stuff, yada yada.

This is more GPU-ey than when you think of a typical CPU and its bandwidth requirements. High end compute cards are already 300GB/s-ish.
 
tipoo said:
Supercomputing of all sorts. Weather simulation, fluid simulation, radioactivity simulation, quantum physics stuff, yada yada.

This is more GPU-ey than when you think of a typical CPU and its bandwidth requirements. High end compute cards are already 300GB/s-ish.
Click to expand...

Scientific computing usually has more regular and predictable memory access patterns, which is why they are able to be accelerated on the GPU. You don't need such a large cache if you just need high memory bandwidth. The large cache size suggests a huge working set or very irregular access patterns.

goomba said:
640K ought to be enough for everybody
Click to expand...

This is coming out next year. We're talking about what it will be used for now, not what it might be used for a decade from now.
 
Status
Not open for further replies.

Similar threads

How does the real 2026 compare to 1927’s 2026?
20
943
sono replied
Does anyone need a Landscaper? I'm getting fired. Lol
15
705
EverydayBeast replied
And just like that, we got a mech as a prototype by a respected robotics company. Unitree presents their GD01.
42
284
Idleyes replied
OpenAI is creating a phone - are you interested? Does it have any chance of competing?
28
270
ResurrectedContrarian replied
Does anyone here browse reddit occasionally
2
98
1K
Avid Reading Horse replied
Top Bottom