- Dec 15, 2021
- 2,356
- 1,273
- 106
Memory bandwidth is one of the greatest constraints in the advancement of the computing power. With CPUs becoming more powerful, they add more caches and support for faster RAM.
Yet, historically Android SoC makers (Qualcomm, Mediatek, Samsung) have been stingy in including big caches in their SoCs.
Apple on the other hand hasn't been so, instead lavishing their SoCs with humongous amounts of cache.
Now there's the classic explanation why they don't do it because adding cache takes up a substantial die area, which in turn substantially increases the cost of the chip.
But I just had a thought. If you have a massive cache (ideally a huge system level cache like Apple does), then you get a twofold benefit. (1) Less accesses to DRAM improves power efficiency, as crucial data is kept on the chip. (2) Having a massive SLC offsets the need for faster DRAM because the massive cache comes with massive bandwidth in itself.
So you can have two setups:
A: Small caches, cutting edge DRAM (eg: LPDDR5X)
B: Big caches, non-cutting edge DRAM (eg:
LPDDR5)
Now I don't know the exact cost values of a LPDDR5X and LPDDR5 RAM chip, but it's certainly a substantial amount, considering LPDDR5X is the latest tech.
Adding a big cache will take a few mm² of area in the die, but the cost increment for the chip is only going to be a few $ right? Should be similar to the price difference between LPDDR5 and LPDDR5X.
Therefore I think the cost of both A and B has to be pretty similar. In such a situation isn't it smarter to choose Option B? Because the having big caches leads to better power efficiency and performance as well.
So Option A and B have similar cost, but B is better due to benefits like superior power efficiency.
As a matter of fact, it seems this is EXACTLY what Apple has been doing. In recent memory they never used the latest DRAM tech.
When Android competitions were using LPDDR5, they stuck to LPDDR4X. When Android manufacturers switched to LPDDR5X, only then did Apple switch to LPDDR5. And in the mean time their chips contain big caches!
Now following B might not be viable for Qualcomm and Mediatek, who have to design SoCs and then sell it to phone OEMs.
On the hand though, I think Option B may well suit Samsung and Google, who make their own Exynos and Tensor smartphonr chips, to be used in their own devices, thus eliminating the middle man.
What are your thoughts?
Yet, historically Android SoC makers (Qualcomm, Mediatek, Samsung) have been stingy in including big caches in their SoCs.
Apple on the other hand hasn't been so, instead lavishing their SoCs with humongous amounts of cache.
Now there's the classic explanation why they don't do it because adding cache takes up a substantial die area, which in turn substantially increases the cost of the chip.
But I just had a thought. If you have a massive cache (ideally a huge system level cache like Apple does), then you get a twofold benefit. (1) Less accesses to DRAM improves power efficiency, as crucial data is kept on the chip. (2) Having a massive SLC offsets the need for faster DRAM because the massive cache comes with massive bandwidth in itself.
So you can have two setups:
A: Small caches, cutting edge DRAM (eg: LPDDR5X)
B: Big caches, non-cutting edge DRAM (eg:
LPDDR5)
Now I don't know the exact cost values of a LPDDR5X and LPDDR5 RAM chip, but it's certainly a substantial amount, considering LPDDR5X is the latest tech.
Adding a big cache will take a few mm² of area in the die, but the cost increment for the chip is only going to be a few $ right? Should be similar to the price difference between LPDDR5 and LPDDR5X.
Therefore I think the cost of both A and B has to be pretty similar. In such a situation isn't it smarter to choose Option B? Because the having big caches leads to better power efficiency and performance as well.
So Option A and B have similar cost, but B is better due to benefits like superior power efficiency.
As a matter of fact, it seems this is EXACTLY what Apple has been doing. In recent memory they never used the latest DRAM tech.
When Android competitions were using LPDDR5, they stuck to LPDDR4X. When Android manufacturers switched to LPDDR5X, only then did Apple switch to LPDDR5. And in the mean time their chips contain big caches!
Now following B might not be viable for Qualcomm and Mediatek, who have to design SoCs and then sell it to phone OEMs.
On the hand though, I think Option B may well suit Samsung and Google, who make their own Exynos and Tensor smartphonr chips, to be used in their own devices, thus eliminating the middle man.
What are your thoughts?
Facebook
Twitter