Friday, April 20th 2018

Intel Stratix 10: Capable of 10 Trillion Calculations per Second

(Editor's Note: Intel says the Stratix 10 contains some 30 billion transistors - and they say that's more than triple the amount in CPUs that run in the fastest desktop processors today. They're really the ones to know it, since Intel has decided to cut on disclosing transistor count on its CPUs for some time now. The amount of data these FPGAs can process in a single second is nothing short of mind-blowing, though: Intel says they can process the data equivalent to 420 Blu-ray Discs... in just one second. If that doesn't spell an unimaginable future in terms of processing power, I don't know what does.)

Because of the Intel Stratix 10's unique design, it can whip through calculations at blinding speeds - often 10 to 100 times faster than the chips in consumer devices. Intel Stratix 10 FPGAs - the latest version came out in February - are capable of 10 TFLOPS, or 10 trillion floating point operations per second. The Stratix 10 is the fastest chip of its kind in the world.

FPGAs, or field programmable gate arrays, are a special class of computer chip that is surging in importance with the rise of applications like speech-recognition, artificial intelligence, next-generation wireless networks, advanced search engines and high-performance computing. Unlike traditional central processing units (CPUs) that power today's laptops and desktops, FPGAs can be customized - or reprogrammed remotely and on the fly - to perform highly specialized computing tas
Source: Intel
Add your own comment

23 Comments on Intel Stratix 10: Capable of 10 Trillion Calculations per Second

#2
BadFrog
phanbueyNice, but aren't GPUs currently breaking 100Tflops in HPC applications?
Yes they are. But, I don't think they care about that but more about trying to spoil AMD's launch day by advertising the "fastest" chip in the world :)
Posted on Reply
#3
phanbuey
BadFrogYes they are. But, I don't think they care about that but more about trying to spoil AMD's launch day by advertising the "fastest" chip in the world :)
They bought Altera in Austin some time ago, this has Altera written all over it.



Later today they will sneak onto the AMD campus and fill their R&D labs with thousands of balloons.

EDIT:

yep... www.altera.com/products/fpga/stratix-series.html
Intel only in name.
Posted on Reply
#4
Yukikaze
phanbueyIntel only in name.
It is produced on Intel's 14nm process. The previous generation was a TSMC 28nm part. So while the design was likely was in the works for a while (likely from before the acquisition), changing a design to a different production process is not a trivial task, which leaves us with two options:
1) It was designed, at least in part, by Intel.
2) It was a completed design that was re-designed for the Intel manufacturing process.

In either case, "Intel only in name" is a false statement.
Posted on Reply
#5
Vya Domus
phanbueyNice, but aren't GPUs currently breaking 100Tflops in HPC applications?
It's much more nuanced that that , you can make a DSP that's a fraction of the size of GPU ,uses a fraction of the power , but can still do a trillion operations a second as well (www.androidauthority.com/pixel-visual-core-808182/).

But that DSP fits a very specific use case and it's extremely limited in terms of programmability and robustness. GPUs are , in nature , much more programmable and flexible but that comes at the cost of computational capability. So how are those GPU breaking 100Tflops then ? Well they simply integrate the same fixed function DSP type hardware blocks that I've just mentioned , so there is nothing special about this. To answer your question , GPUs aren't really braking 100Tflops , it's the additional hardware blocks inside of them that do that but those are essentially more or less the same type of things that power products like the Stratix as well. (By the way , you might have mistaken Tflops for "Trillion operations" it seems , GPUs still do calculations within TeraFlops)
Posted on Reply
#6
eidairaman1
The Exiled Airman
BadFrogYes they are. But, I don't think they care about that but more about trying to spoil AMD's launch day by advertising the "fastest" chip in the world :)
Won't affect the current market as thet won't be released to us, probably more for Server loads... being remotely programmed sounds like a tinfoil hat situation though, and with intels poor security batting record, I'd be wary of a Skynet deal.
Posted on Reply
#7
phanbuey
YukikazeIt is produced on Intel's 14nm process. The previous generation was a TSMC 28nm part. So while the design was likely was in the works for a while (likely from before the acquisition), changing a design to a different production process is not a trivial task, which leaves us with two options:
1) It was designed, at least in part, by Intel.
2) It was a completed design that was re-designed for the Intel manufacturing process.

In either case, "Intel only in name" is a false statement.
Intel played the same part TSMC or any other Fab would have played.

This chip would have been built with or without Intel was the sentiment behind that statement.
Posted on Reply
#8
Nuke Dukem
Intel says they can process the data equivalent to 420 Blu-ray Discs...
420
Posted on 4/20 (the US way of writing it)

Come on, Intel, pass the joint already :D
Posted on Reply
#9
eidairaman1
The Exiled Airman
Nuke DukemPosted on 4/20 (the US way of writing it)

Come on, Intel, pass the joint already :D
Er, many ways of writing the date, 20180420
Posted on Reply
#10
R-T-B
eidairaman1Er, many ways of writing the date, 20180420
But only one that fits the joke...
Posted on Reply
#11
eidairaman1
The Exiled Airman
R-T-BBut only one that fits the joke...
Erm i said 0420 lol
Posted on Reply
#12
ArdWar
YukikazeIt is produced on Intel's 14nm process. The previous generation was a TSMC 28nm part. So while the design was likely was in the works for a while (likely from before the acquisition), changing a design to a different production process is not a trivial task, which leaves us with two options:
1) It was designed, at least in part, by Intel.
2) It was a completed design that was re-designed for the Intel manufacturing process.

In either case, "Intel only in name" is a false statement.
Stratix 10 itself is already a quite mature family. The GX (generic) and SX (SoC) variant already shipped from 2013. The MX (HBM) variant just shipped a while ago. All Stratix 10 are already fabbed in Intel's 14nm process long before it's acquisition by Intel. The only thing new from this, uh..., press release is the TX version of Stratix 10. The long overdue high-speed transceiver version. There's no production process to change between them, besides the obvious incremental process optimization change.

I feel that after their acquisition by Intel, there's just more buzz around their product announcement. Either because of Intel's marketing trying to market it to AI-this, accelerator, coprocessor-that, or just because it's a release from Intel (come on, do we get to see this much coverage when they're still Altera, or if it's from Xilinx, or Lattice?)
Posted on Reply
#13
cdawall
where the hell are my stars
Jokes on everyone here this is a bitcoin mining chip :roll:
Posted on Reply
#14
the54thvoid
Super Intoxicated Moderator
I was out with my brother last night (electronic engineer, PCB designer) who works on FPGA chips and whose company is doing rather well. We were talking about this chip and he has first-hand experience of it. Says it does not do as Intel say (speed is far less). Also, a buyer for the chip was so unimpressed they walked away and went with Nvidia instead as their option was far better. Lols.
Posted on Reply
#15
Vya Domus
the54thvoidI was out with my brother last night (electronic engineer, PCB designer) who works on FPGA chips and whose company is doing rather well. We were talking about this chip and he has first-hand experience of it.
FPGAs are useful because they are configurable after manufacturing , they are a very cost effective solution for companies that want to distribute a small amount of hardware that they constantly need to reiterate upon or customize with minimal costs and do it quickly. But these things just aren't speed demons in nature due to the way they are built/programmed.

Achieving 10 Tflops with an FPGA is certainly a big feat , but depending on the application it might be better to go ahead and make your own custom silicon or get something that already exists on the market. Intel certainly isn't looking to take away market share from Nvidia/AMD with this , these things are simply meant for an entirely different market.
the54thvoidAlso, a buyer for the chip was so unimpressed they walked away and went with Nvidia instead as their option was far better.
To be honest it looks to me like that buyer didn't really knew what he was supposed to look for in the first place if that was the case.
Posted on Reply
#16
jabbadap
Tflops in what actual precision? 10 TFlops of fp64 is massive, of fp32 very good still, of fp16 uhm yeah ok I guess, of fp8 why even bother.
Vya Domus...
To be honest it looks to me like that buyer didn't really knew what he was supposed to look for in the first place if that was the case.
Maybe they fell to that marketing propaganda of AI et al.
Posted on Reply
#17
the54thvoid
Super Intoxicated Moderator
Vya DomusTo be honest it looks to me like that buyer didn't really knew what he was supposed to look for in the first place if that was the case.
It didn't actually perform as well as Intel said it would. Apparently, Intel is quite bad for this - promise the earth and deliver far from it.
Posted on Reply
#18
Vya Domus
the54thvoidIt didn't actually perform as well as Intel said it would. Apparently, Intel is quite bad for this - promise the earth and deliver far from it.
That's a different story if they literally lied about their product , though I would be very surprised if they did. This isn't like a consumer CPU than they can sell to thousands of people and get away with lying about it just because the people that bought it typically don't have a clue about these sorts of things.
Posted on Reply
#19
the54thvoid
Super Intoxicated Moderator
Vya DomusThat's a different story if they literally lied about their product , though I would be very surprised if they did. This isn't like a consumer CPU than they can sell to thousands of people and get away with lying about it just because the people that bought it typically don't have a clue about these sorts of things.
Lying about a product? A big company?

Well, that wouldn't be a first, would it?
Posted on Reply
#20
Vya Domus
I don't doubt their ability to lie a second, especially if we are talking about Intel. But this one would be odd , they just can't get get away with it , these things aren't used by your average consumer like I mentioned. The success of these products rely on the fact that they function as advertised.

If that's how they hope to run even these sorts of businesses , they are thoroughly screwed.
Posted on Reply
#21
the54thvoid
Super Intoxicated Moderator
I guess we can equate it to the auto industry. A car will do 'x' performance. But... It needs specific tyres, specific surface, specific environment. In other words, the stated performance is highly dependent on very controlled variables, unlikely to be replicated in real world use.
Posted on Reply
#22
Unter_Dog
Much rather buy parts from xilinx. I personally get much better support from them. Ymmv
Posted on Reply
#23
Xaled
10 times faster from Intel (or nVidia) would mean 10 times more expensive.. so why should we care?..
Posted on Reply
Add your own comment
Dec 24th, 2024 12:08 EST change timezone

New Forum Posts

Popular Reviews

Controversial News Posts