Why doesnt the CPU make the area larger, using area and quantity to trade for performance?
The original question of the topic has been modified to clarify the direction of the inquiry It is divided into three specific question directions Why not stack more cores on larger wafers to trade area for performance? For example, why divide it into chiplets? Why not make it into a single chip to save power consumption in chiplet communication? For example, why do GPUs generally have larger single chips? 2 Why not stack more and wider execution units in a single core to trade area for performance? For example, why not extend SIMD width to 1024 bits or even higher? For example, why is the number of SIMD FPUs in desktop big core CPUs surprisingly consistent? Cant more be added? (Currently, the SIMD FPU throughput of Intels various Lakes, AMD Ryzen, Apple Firestorm, or Cortex X1/2/3 is all 512 bits/clk) 3 Why not stack more chiplets? Trade quantity for performance? For example, fill the entire motherboard with CPUs, one topping a hundred
Comparison of Intel and AMD server platforms
In fact, this is how Intel and AMD server platforms work.
As we all know, Intel’s desktop processors can be divided into MSDT[1] and HEDT[2], which both have corresponding workstation/server versions. These Xeon processors and C-series chipsets use the same wafers and packaging as their corresponding Core processors/consumer chipsets, but they are not compatible with each other in principle. These processors have no difference in size compared to consumer-grade processors.
However, above the HEDT, Intel also has higher-end server platforms (LGA3647, followed by LGA4189), and these Xeon processors are called Xeon Scalable processors. Xeon Scalable processors have more cores, support multiple CPU interconnections (dual/quad/octa), and support 6-channel memory. The package size of these processors is much larger than that of MSDT and HEDT processors.
The appearance and size comparison of LGA3647, LGA2066, LGA1151, LGA4094, and AM4 processors are shown in the figure[3].
Platform
MSDT
HEDT/Server
Server
Microarchitecture
Skylake
Skylake
Skylake
Process
14nm
14nm
14nm
CPU Package
LGA1151
LGA2066
LGA3647
Chipset
C23X, C24X
X299, C422
C62X
Number of CPU Cores
Up to 4/6/8
Up to 18
Up to 28
Multi-CPU Interconnect
1
1
4/8
RDIMM Support
No
No/Yes
Yes
Number of Memory Channels
2
4
6
Number of PCIe Channels
16
28/44/48
48
These three ATX motherboards (E3C246D4M-4L[4], C422 WS/IPMI[5], and EPC621D8A[6]) respectively support LGA1151, LGA2066, and LGA3647 processors, and it is not difficult to see that the scalability of the latter two is much stronger than the former.
AMD Ryzen™ Threadripper™ PRO 5995WX is a powerful CPU.
You can take a look at AMD’s Threadripper, which can be used on a regular PC and is nearly twice the size of a regular CPU, about the size of a palm.
For example, compare the newly released Threadripper, the AMD Ryzen Threadripper PRO 5995WX, with a default TDP of 280 watts. The new 7000 series, on the other hand, can reach up to 350 watts.
However, it’s recommended to check the prices:
Intel also has large-sized CPUs, like the one below, also rated at 350 watts:
Although the price of the Intel CPU is much higher than that of the aforementioned AMD one, Intel’s strength lies in the fact that there are people willing to purchase it.
Sometimes it’s not that these miraculous devices don’t exist, it’s just that our own abilities don’t allow us to catch a glimpse of them too early.
Apple: Suspicion of clicking on me!
Apple: I suspect you’re tapping on me!
M-Series Chip Cases: M1 expands to M1 Pro, further expands to M1 Max, then doubles to M1 Ultra~
Cerebras - Large-scale Chip Solution
Product - Chip - Cerebras
A whole 12-inch wafer directly produced. Pay attention, this is a chip, not uncut, it’s just this big. How about this? Can’t get any bigger, right?
Some answerers talk about global clocks, what a load of nonsense.
In reality, the reason why chips are not usually made this large is one - poverty. Or in a more emotionally intelligent way of saying it, the cost is too high.
If you have money, then go ahead.
Limitations of Cost and Performance
This question is actually quite good.
Just like people say, why not make compact car models bigger, just like SUVs?
Why are airplane economy class cabins so cramped? Can’t they make them more spacious like business class? At least provide comfortable footrests.
Or, why can’t computer chips all come with i9 processors? Or make CPUs as large as the screens, wouldn’t that increase computing power? Beat AMD’s Threadripper with fists, and step on AMD’s R9 consumer processors with feet.
So, everything revolves around cost. If we talk about the quality of a product without considering cost, it’s like an unruly competition.
The cost of a wafer is limited, and now the number of semiconductor diodes that can be made on a single wafer is increasing, effectively lowering the cost.
Expanding the area, wouldn’t that be a problem for oneself?
The increase in performance is not unlimited either. Based on the current lithography process, achieving simultaneous progress in single-core and multi-core performance is already very good~
PS: For fun, I recommend two computers.
Expensive and large chips for different user levels.
You can’t afford the big ones. Enterprise-level users are willing to pay, so of course the size can be increased. The Sapphire Rapids 8480 with 1600mm2 and the EPYC 9654 with 1200mm2 are both priced at $10,000 per piece. And these are still in MCM packaging. If you look at monolithic ones, the colossal Nvidia H100 GPU with 800mm2 is priced at $50,000 per piece, with delivery scheduled for next year. It’s really charitable to sell consumer-grade chips that are only 200-300mm2 for $500.
Thread Ripper, Aurora Dragon, XEON
Yes, there is. This thing has a large area, multiple cores, and strong performance, but it is expensive, costing tens of thousands for one, even the size of a palm.
Thread Ripper, Night Dragon, XEON
CPU Size Depends on Usage.
I saw some answerers mention that the reason for not making CPUs bigger is because of poverty or high cost, but that’s not the case. The main factor is where you are using the CPU. If it’s for a laptop, then smaller is definitely better because it allows for a more compact and portable size. Even for desktop computers, a smaller size is still preferred.
Therefore, whether a CPU is big or small depends on its usage.
Difficulties in heat dissipation, cost, and data interaction between chips.
Complexities:
-
AMD’s Ryzen architecture can combine multiple standard CCDs, such as 7950 and 7900.
-
Apple’s M-series processors, like the M2 Ultra, use multiple M2 Max chips combined.
Parallelism with multiple independent processors:
Commonly seen in servers with numerous CPUs. Some, yes, it’s just the data interaction between heat dissipation, cost, and chips that’s a bit tricky.
For scalability:
-
AMD’s Ryzen architecture can have multiple standard CCD combinations, such as 7950 and 7900, which are two CCDs.
-
Apple’s M-series processors, such as M2 Ultra, are created by combining two M2 Max chips.
Parallel processing with multiple independent processors:
The most common example is servers, which have numerous CPUs.
Bigger, more powerful devices for you.
Area matters, it’s all about money.
If you are willing to pay, there are large cups and extra-large cups available. Pro max ultra, 4-core, 8-core, 12-core, 16-core…
In addition, as the power increases, as long as we match it with heat dissipation and power supply (battery), it shouldn’t be a problem for mobile devices to carry a few hundred grams extra.