At the Apple conference half a month ago, I guess many friends went up to make a scene.
New phones, new computers, new chips...
One of the most impressive is definitely the new top chip:
M1 Ultra。
Why do I say that? Because of the performance of this chip, it is simply a cow x heaven!
Wait, this parameter is so familiar.
You wait a while, I will go to Apple's official website to give you a screenshot of last year...
Can't say it's exactly the same, it can only be said... Exactly doubled.
It's as if there's a feeling in the dark:
Apple should not have glued the two M1 Maxes together!!!
Don't you say, Apple really does this!!!
And the M1 Ultra's multi-core running score is just twice that of the M1 Max, and Apple's original take-off chip has become doubled.
>/ Real Glue? Fake glue?
But ha... A small partner who has a little understanding of computer knowledge may say it.
How is this a little wrong!
Because, like the M1 Ultra, such a "glue chip", not no one has done before, and the result is that everyone has overturned.
More than a decade ago, Intel and AMD both engaged in it, and both of them were still arguing for a "true and false quad-core".
The result is finally a multi-core processor created by splicing chips, because the performance coordination across chips is not good, and the actual performance is one than the other.
Not only the CPU, Lao Huang also did gtX 690 and other "dual-core graphics cards", known as doubling performance.
As a result, most games can only call one of the cores for rendering.
Why do I know so clearly? Because I spent 8,000 yuan to buy a young leek for this card!
Hey, the sad things of that year, don't mention it.
In other words, in most cases in the past.
"Glue chips" are all - pay the money for double chips, and do a chip work.
that... Why is this Apple's M1 Ultra different from the old predecessors who walked on this road before, through the splicing of two chips, it can achieve 100% performance double?
Because ah, Apple has done enough in the communication between the two chips.
Don't look at this thing so thin, it can carry 2.5 TB/s of data traffic.
Note, it's 2.5 TB! Every second!
What is this concept? Let's take a few examples that may not be appropriate.
5G is fast enough, and we don't feel that we need this speed in our daily lives.
Ultra Fusion is 1000 times the theoretical limiting speed of 5G 2.5GB/s, which is nearly 80 times the theoretical speed of PCIE 4.0 x16 slot on computer graphics cards.
According to Apple's press conference, the performance of the Ultra Fusion is more than 4 times higher than other people's flagship multi-chip interconnect technology.
Even if lao Huang just announced the latest glue architecture NVIDIA Grace Hopper on gtC, the on-chip interconnection speed is only 900GB/s
What kind of process Apple uses this time, we can't say clearly.
Some people say that it is a CoWoS-S package with TSMC, and some people speculate that it is an INFO-LSI package.
But we don't have to bother to remember these names, as long as we know that they are very, very, very NB.
In the past, those "glue chips" that overturned.
Although the multi-core scheduling effect of the chip itself is very good, the result is that they cannot coordinate with each other to do what tasks to do.
Apple's approach this time is very simple.
You're communicating too slowly, aren't you?
You can't communicate well, right?
I'll give you a 2.5 TB/s communication bandwidth to play slowly.
Rough, direct, but useful.
In Apple's official publicity, most of the units between the two chips can directly exchange information and data through Ultra Fusion.
Thus achieving full utilization between different CPUs and GPU compute units, greatly reducing the latency of data.
So, ah, unlike the previous glue chips, after solving the problem of on-chip communication.
Apple's glue chip doesn't need to be treated as a "dual-core chip" at all.
For developers, there is no need for additional processing of multi-chip processes as in the past.
Because Apple has already helped you optimize the two chips into a whole.
>/ Chiplet? Big chips?
However, there may be some small partners who will be curious:
Why does Apple have to study the trouble of putting two chips together?
Directly like the previous M1 M1 Pro M1 Max, wouldn't it be possible to design a complete chip that is one larger?
Isn't this more efficient and hassle-free?
Actually, Apple wants to.
It's just that the cost of producing a single-chip oversized processor, even if a big local tycoon like Apple pinches his finger, he feels that he can't draw it and can't carry it.
Speaking of this, we have to talk about the cost control of chip research and development:
We all know that the chip is cut on the wafer.
Wafers, wafers, wafers, as the name suggests, are round components, while common chips are mostly square structures.
When cutting, it will be embarrassing to find that the larger the square chip is... The easier it is to waste and cause waste on the side.
This waste of scraps is all white flower ticket wow.
And ah, we make chips, but also have to take into account a yield rate problem.
After a wafer is lithographed, more or less some circuit damage will occur.
In general, vendors address this effect through redundancy (backup circuits) of the circuit design.
But if the main and backup circuits of a function are all carved, then the chip must be gg.
If the design is a small chip is fine, itself is small in area, not easy to fail, each production of shipments is also more, broken is not painful.
And the super-large size of the chip... On the other hand, the production itself is small, and it is easy to break...
So Apple, which uses a glue chip design, actually leaves enough way back for itself:
When it is ready to be produced, it is directly made according to the specifications of the M1 Max.
After the production is complete, find out the adjacent and intact circuits, let them be paired in pairs, install the Ultra Fusion, and cut them off, that is, a brand new M1 Ultra.
The remaining chips that are not matched in pairs, but have good performance in themselves, can continue to sell them as normal M1 Max.
Speaking of which, it's not over, what if this M1 Max is also broken?
If this M1 Max is just broken at the bottom, then you can also cut off the lower half of the GPU and memory controller, and continue to sell it as an M1 Pro.
That said, don't look at the M1 family with four brothers.
But in theory, as long as two production lines are opened, they can produce them all...
This knife method... Lao Huang was ashamed of himself.
In fact, as early as before the conference, many netizens were already speculating that Apple would give everyone such a "glue chip".
However, at that time, everyone's speculation was wilder, and everyone thought that Apple would also bring a four-chip interconnect M1 Super Special Final Ultra Pro Max.
Long so.
Or maybe it's that way.
That power... But I really can't imagine it.
>/ Single core? Glue multi-core?
Although this time for the M1 Ultra, we ended up seeing only glue stickers for two chips.
But at the end of the conference, Apple also emphasized that the replacement of the entire desktop-level processing is not over.
Maybe by the fall, the apples that appear in front of us can go further.
Glue the four chips together and install them on your Mac Pro.
Then again, the chip at Apple's conference, although the performance is ruined...
However, for our ordinary consumers, it may also be a pleasure.
Who can afford to eat tea eggs (dog head) ▼
But for some chip manufacturers, Apple's path is also giving them a sample.
In the past few years, many people have been seen singing about Moore's Law every year.
With the continuous advancement of semiconductor process technology, the cost of producing chips is also rising.
But the cost has come up, and the products made cannot satisfy consumers.
Especially on the mobile phone SoC, the past two years have been frequently overturned.
In this environment, the vision of chip manufacturers has once again fallen to multi-core integration.
Single core is not enough, multi-core to make up; multi-core is not enough, glue is top.
Although the "glue chip" can only solve the urgent need for performance, it cannot fundamentally increase the density of the transistor.
But in fact, the manufacturer has not put it down in recent years.
Like Intel and NVIDIA, although the quality of the glue made in those years was not good, it has also continued to introduce new "synthetic large chips" in recent years.
AMD is not to mention, in the past few years, it was able to kill the consumer market, relying on this hand glue chip Ryzen Ryzen and EYPC Xiaolong.
Don't look at today, Apple with 2.5TB/s "glue" for a time to dominate.
However, the development of technology is also changing with each passing day, and the new technologies and new processes of various manufacturers are also emerging in an endless stream.
Who can laugh last in this semiconductor war without smoke and smoke, and there is no letter.
I want to really make a "One More Thing" that makes everyone "shine".
Glue alone is not enough...