07-27-2015, 10:50 PM
Hmm, that is interesting about the base clock - it seems that we've been stuck with 100Mhz BCLK forever. This must be why the new edram "L4 Crystal Well cache" is helping Intel's 14nm CPUs so much (like massive L2 cache helping Maxwell overcome most of its relative bandwidth limitation).
Next year, perhaps we'll be seeing cheap DDR4 memory clocked at least 3200MHz "effective".
Then as Intel and AMD stubbornly stick to such low base clock, either one of these companies might go ahead and integrate HBM memory with the CPU, while still allowing for add-on DDR4 expansion.
This is what AMD should've done with their Fury cards, that were limited to only 4GB of HBM1 memory. Just add old-fashioned GDDR5 chips to make it say, 8-12GB or even 16GB, to take a bit of the thunder away from Titan X. It would have been more complicated, but with enough bright engineers, it should have been as "do-able" as the L3 and "L4" caches on Intel's CPUs. The remaining GDDR5 memory would be slower for certain, but Nvidia did that before with GTX 970 and even GTX 550 Ti where a certain portion of memory had lower bandwidth than the rest.
Ideally, every high-end CPU should come with 4-16GB of HBM2 memory while still allowing for QuadDataRate memory capacity of up to 128GB or so. It's like having an integrated GPU with the CPU, making it easier for users to troubleshoot their PC's if their dGPU's go bad - same for memory sticks (if it's suspected to be bad, the PC would still work without the sticks).
Imagine a i5-4690K with 4GB HBM2 , i7-4790K with perhaps 8GB HBM2, and the "E"-variants with up to 16GB. Intel would be more justified with getting gamers to cough up $400+ for a CPU that has more than just the new generation integrated GPU that most of them still won't ever really be using.
Next year, perhaps we'll be seeing cheap DDR4 memory clocked at least 3200MHz "effective".
Then as Intel and AMD stubbornly stick to such low base clock, either one of these companies might go ahead and integrate HBM memory with the CPU, while still allowing for add-on DDR4 expansion.
This is what AMD should've done with their Fury cards, that were limited to only 4GB of HBM1 memory. Just add old-fashioned GDDR5 chips to make it say, 8-12GB or even 16GB, to take a bit of the thunder away from Titan X. It would have been more complicated, but with enough bright engineers, it should have been as "do-able" as the L3 and "L4" caches on Intel's CPUs. The remaining GDDR5 memory would be slower for certain, but Nvidia did that before with GTX 970 and even GTX 550 Ti where a certain portion of memory had lower bandwidth than the rest.
Ideally, every high-end CPU should come with 4-16GB of HBM2 memory while still allowing for QuadDataRate memory capacity of up to 128GB or so. It's like having an integrated GPU with the CPU, making it easier for users to troubleshoot their PC's if their dGPU's go bad - same for memory sticks (if it's suspected to be bad, the PC would still work without the sticks).
Imagine a i5-4690K with 4GB HBM2 , i7-4790K with perhaps 8GB HBM2, and the "E"-variants with up to 16GB. Intel would be more justified with getting gamers to cough up $400+ for a CPU that has more than just the new generation integrated GPU that most of them still won't ever really be using.
Ok with science that the big bang theory requires that fundamental scientific laws do not exist for the first few minutes, but not ok for the creator to defy these laws... 