cerebras.net
Waferlets can't compete.
Nvidia BTFO
Other urls found in this thread:
fortune.com
twitter.com
How do you even cool this?
Liquid nitrogen, probably.
Why contain it?
>AI optimized
>rand()
>12cm CPUs
Intel has finally outdone themselves.
What would that even be used for?
RTX OFF
RTX ON
Giant copper bathtub on top of the chip filled with liquid nitrogen.
Lol
deep learning
I don't know, seems too thin.
>46,225mm^2
am i stupid or does that translate to almost 7 meters across?
i'm ready for my deep learning lessons
You're a big chip
yeah 7.19 meters
Yeah, you're pretty stupid.
100 mm^2 = 1 cm^2
10,000 cm^2 = m^2
This thing, 462.225 cm^ or 0.0462225m^2
ah, i see
For you
needs moar cores.
If I pull off that coolant pump, will you overheat?
If I pull that mask off, will you die?
Do you?
It would be very expensive
did everyone get it haha
A large area in itself isn't actually inherently hard to cool. You simply upscale the watercooling system with a larger block with more channels and greater water throughput, and a bigger pump and more radiator area. The problem with cooling modern processors is that the ever shrinking process results in the same heat in an ever smaller area, or ever more heat in the same area. Depending on the process this mega processor may actually be pretty simple to cool.
I'm pretty sure both Nvidia and AMD could make something like that if they really wanted, since it's basically an uncut wafer of GPUs, with the same way of dealing with defects (by fusing off faulty cores). But how do you connect memory to this thing?
You are a fried guy
>400,000 cores
How? wouldn't latency and Vdroop across the core be absolutely atrocious?
and frequency scaling would have to be in the hundreds of Mhz best case.
>check out the site
>not one single specification
>all buzzwords
ahh thats how.
Cool?
fucking kek
That's not how it works.
Probably costs over 40k to buy. 56x the area of the next largest chip, means ~(x)^56 probability that a flaw occurs. I doubt the yield will be anything other than disgusting.
>latency
Why assume that would be a problem? Really.
>Vdroop
Yeah it's obviously impossible to have a distributed power network...
>frequency scaling
Could be asynchronous or having a slow clock = no problem.
what if they just glue 56 normal sized wafers together?
The yield on this must be so awful, just 1 defect anywhere means the entire wafer will be thrown out
everyone is asking the wrong questions
the real question is
how many chrome tabs can I have open?
>The most expensive chip ever build
>The lowest yields ever built
This is your tech on monolithic dies
zero because it's not a general purpose processor nor is it ram
Only IBM and AMD have the glue
Run Crysis at 240 FPS
Price? Power? Why can't I just install a bunch of GPUs?
Based and photolithography-pilled
40k? More like 4 million.
Does it run Skyrim with mods?
So this is the power of hole technology.. woah
So we spent 30 years shrinking transistors and now we're back to make them bigger.
Sad thing is the next asteroid could take thousands of years to wipe us out.
Will it run Crysis okay?
The transistors on this are still smol. It just has 1.2 trillion transistors vs the 18.6 billion on the Titan RTX
kek
>56x larger than the biggest GPU ever made
>Uncut 84-GPU wafer
No, drivers are shit
Who fabbed this?
>15 kilowatts of power
underrated
ray tracing with acceptable framerate
>15 kilowatts
Jesus Christ
So we're going from chips per wafer to wafers per chip. Well isn't that great! I'm sure everyone was waiting for a single chip with even worse price/perf then Intels lineup! How can this ever fail in 2019, the year of chiplets and distributed conputing!
200*200=40000
So each side is around 200mm
have sex amdincel
Apparently they use redundant circuits to route around possible defects, so it's just fuck you expensive, not literally impossible because you'd need a whole wafer to come out without defects.
G L U E
L
U
E
Well I can see how you could make such a fuck huge design work. However, it's still retarded. We've (well AMD) have finally found out a very efficient way to communicate between chiplets which makes baking chips more affordable and effecient, there are more and better ways to do distributed computing then ever and for some reason the biggest monolith monstrosity shows up and it's supposed to be useful?
Be honest, even if there is a way to make this behemoth of a monolith work, there is no real market or use case for this thing.
Naisu
Just believe in Mother Earth and her super volcanoes desu.
Nope, the chip can accept some defects and still mostly work.
user, AMD has literally, exactly, completely, absolutely zero marketshare in the area this chip is targeting. It's all Nvidia and some legacy Intel servers. So long as it beats Nvidia by a large margin then it's an autowin.
Too big for an ai chip. Just one tumble will cause it to snap in two, producing a dead android.
Let it spill into the schools and churches
You're also stupid
>Distributed power network
Only to each core maybe
>It's all Nvidia and some legacy Intel servers
nvidia and jewgle's custom shit
this is the coolest shit I've seen all year
Used as a plate to put my burgers on.
The chip is designed with redundancies so the whole wafer can handle dozens of defects. It has 18 gigabytes of on-chip cache. 100 petabits of bandwidth (aggregate, obviously.) It has the IO capacity for 10 petabits of RAM. It's probably built on 28/20nm for the sake of cost and yields, because throwing away a 14/16nm die costs more than 10 grand.
It's advertised as 8.5x8.5 inches, or 216mmx216mm, or 46656mm2
That's whole system power. With a design like that I'm willing to bet that half the power goes to IO, RAM, and supporting hardware.
Even if it were 15KW for the chip itself that is .32watts/mm2 - well below the 1W and higher we have had on CPUs for over ten years.
This system is meant to replace 8x V100s with dual 8180s while providing >10x the performance.
The designer has said "We've seen training times decrease from months to minutes" - take that with a grain of salt, but still.
An 8X Quadro with Xeon setup is going to use 5KW per unit and perform multiples if not orders of magnitude worse.
I for one want to see "chips" like this developed and built to simulate a complete human brain in the next 10 years.
Only 65watt tdp, cooler sold separately
I read somewhere it's manufactured on TSMC 16nm. That power density is almost exactly the same as the big Nvidia GPUs. I'm interested to see how much of an improvement the SLA cores are over the tensor cores everyone else is making and how good the efficiency gain from the SRAM is. Might be interesting if they chop up defective wafers and sell the tiles individually too.
They have a system that automatically routes around broken sectors.
Custom solution involving a custom circuit board with small vertical channels continually pumping water.
Total power usage of the "appliance" they sell is about 15 kW according to reports.
precisely one of the things it couldn't be used for
>power density is exactly the same as
But it isn't.
The 15KW figure is for a fully assembled system, not the chip itself. And like I said I am willing to bet that the supporting hardware uses half of the total power. That's a lot of RAM and I/O to support a billion billion bits/s with 400,000 cores worth of data.
we can't make them much smaller anymore
God I wish that was me
Its got memory on die integral to the compute cores. DL is massively memory limited, so its the only way to not choke this monster to death.
Its going to be really interesting to see how its going to stack up against the other players in datacenter AI.
At this point you can just CNC a waterblock out of a huge hunk of steel.
You like charcoal huh?
>nvidia btfo
as if nvidia is leading on anything A.I
>15 Kilowatts
>That's 20 horsepower
They even made a chip more powerful than my motorcycle... Fuck your A1 license EU
Holly shit
this thing is bigger than my Dick
really big fan
>VERY COLD IS BETTER, FUCK HEAT TRANSFER AND SHIEETT
You should watch less hollywood movies, faggot.
In case you haven't noticed it, it consists of nanowelded smaller cores, the final chip was not made entirely on the same wafer.
Same thing they do with medium format camera sensors.
Well that's neat, and makes sense, a shame it isn't cost effective to fab entire wafers like this.
A fully integrated wafer would be somewhat more efficient and stable than a thousand blocks stitched up.