Chipworks has provided our first look inside the A9X chip that powers the iPad Pro, thanks to a die shot obtained by Motley Fool.
One of the fascinating things about the incredibly small processes used for today’s chips is that even a die shot doesn’t provide a definitive answer to what is found inside one. There’s still a certain amount of interpretation involved, though both Chipworks and Motley Fool‘s chip specialist Ashraf Eassa agree on what we are looking at in the full image (below) …
Chipworks’ Dick James tells me that he sees a 12-cluster GPU, two CPU cores, and an absence of the level-three cache memory found inside the A9 chip […]
I agree with his assessment. The two CPU cores can be seen in the green box, and I believe that inside of each blue box are two GPU clusters, for a total of 12 clusters.
Eassa says that the memory interface is twice as wide as the one on the A9, explaining the absence of the cache – the chip can read & write data to RAM twice as fast, so doesn’t need it. The far larger size of the iPad Pro, enabling it to much more easily dissipate heat, is the likely reason Apple was able to take this approach.
The GPU design appears to be a custom one specifically for this chip. Apple’s GPU supplier, Imagination Technologies, doesn’t offer the 12-cluster design seen here. Eassa says the entire chip is an extremely impressive piece of engineering.
I believe that the A9X is easily the most advanced mobile system-on-a-chip available today. It has best-in-class CPU/graphics performance and it is an extremely large chip built on a bleeding-edge foundry manufacturing process.
It is little wonder that the chip performs as well as it does in the various performance tests available online.
The green area indicates the two CPU cores, while the blue areas represent the 12-cluster GPU (two in each box).
FTC: We use income earning auto affiliate links. More.
Cant wait to see the S2 chip bring the power of ipad pro to the apple watch 2. LOL
The article: > The A9X features a memory interface that’s twice as wide as the one on the A9, allowing the chip to pass data to and from memory at twice the rate that the A9 can.
You: > the chip can read & write data to RAM twice as fast, so doesn’t need it.
It’s not as simple as that. If you expand a highway from 2 lane to 4 lane, cars are not going twice as fast there, the highway is simply capable of taking in twice as many cars but they’re all at the same speed. A9X supports DDR4 that happens to be faster as well, the performance can match even at half of the memory interface. If you pass data that doesn’t take up the entire highway, it doesn’t mean the CPU process the data twice as fast, it will still have to wait for the next pass to grab the next wave of data.
The level 3 cache is much faster than the memory chip with very low latency, that’s the purpose of this low level cache. He assumes the cache just takes up too much of the room to justify the extra cost over the extra wide memory interface. Level 3 cache will come back in the future when there’s room for it, possibly in the 10nm shrink.
Useful insight, thanks.
Very Interesting.
I completely agree. The only reason you drop L3 cache is because you don’t have room for it.
Bringing 12 CPU cores inside the SoC means you do not need L4 cache, but L3 is still at least 10x lower latency than external RAM.
More than 15 years ago (a long time ago in a galaxy far, far away…) we fought to keep our programs in RAM so they would be fast, because a “RAM miss” that went to spinning rust was so expensive. Today we fight to keep everything we need in cache and L2 cache is not nearly big enough.
But, if you need to move 1,000 cars from point A to point B, and, if it takes you one hour using two lanes, you should be able to complete the process in one half of the time, or twice as quickly, if you were able to use four lanes.
This is the chip that should have been in the new Apple TV. Then it would have been able to handle accurate ports of last gen titles.
You are comparing a device that costs $799 with a device that costs $149. How does that make any sense? Apple could have made a console to compete with the Xbox One and PS4, but it wouldn’t have cost $149.
A PS4 costs £300, and for that you get a desktop class GPU, an 8 core laptop class CPU, 8gb of RAM, a blu-ray drive, and a Bluetooth controller. So I’m quite sure Apple could have managed to put an Xbox 360 level GPU into the Apple TV without increasing the cost.
Clearly you don’t understand how the game industry works for console gaming. A PS4 is sold below or close to the actual build price. Sony is probably losing money on each and every PS4 being sold. It’s still an interesting business because they get compensation from the game developers selling very expensive games.
Apple has never been interested in selling hardware without a healthy margin. And iOS games seldom cost more then $5. Apple concentrates on the thing they’re good at: making hardware and providing a great platform for developers.
That’s why they’re the electronics company everyone wants to be and why Sony seems to be on their way to oblivion. They forgot what it takes to sell products: great hardware and a healthy ecosystem of apps and services that all work together.
the A8X in the iPad Air 2 has three cpu cores, this only two. Why did they went back to two cores ?
Because having great single core performance is much better for most tasks. Even for properly programmed parallel tasks, the 2 cores of the A9X are faster than the 3 cores of the A8X. More cores need more power, which is another downside.
better singlethreaded compute power its still relevant and easier to have a beefy dual core