Quote from: NikoB on September 08, 2023, 16:50:46This is a shame for Intel - the Apple M2 Max has a soldered memory with a bandwidth of 400 GByte/s, and not 120Gbyte/s like Intel, i.e. more than 3 times faster.

Again it is clear that Intel is counting on suckers, not knowledgeable buyers - they should have soldered 16-32GB HBM2/3, but not the shameful LPDDR5 7500.

I've been writing for a long time that given the availability of 16-24GB VRAM with a bandwidth of 500GB/s on x86 in discrete cards, it's high time to make the processor have direct access to this memory and make it possible to execute code or use it as an L4 cache for running the OS /software. Since very often people even buy gaming laptops not for games at all and VRAM stupidly and stupidly sits idle in most cases, although even on 6GB VRAM you can easily run a full kernel with a bunch of W10/W11 drivers with a 5-10 times better response than on brake DDR5 5200-5600 in slots with a shameful 80-90GB/s, instead of 500GB/s+ for VRAM video cards.


Intel, AMD, when will you at least have a 512-bit memory controller or at least a 256-bit one in the top consumer series?

To much vram usage will burn up an ssd very fast lessening its life span 

If you achieve desoldering of 32GB HBM3 with 1TB/s on a SoC substrate with a processor and provide a direct GPU link to them, then in general the video cards will not require their own VRAM. 8Gb for the OS and 24 as VRAM will be enough, and slow DDR5 will be used as a medium for programs that are less bandwidth-demanding. That's how I see it.

There is no point in having a separate GPU - it should be next to the processor on a large substrate and with shared super-fast HBM3 memory.

Sound cards have gradually disappeared from PCs, and now it's time to combine the CPU and GPU along with super-fast buffer memory for both.

Although from the point of view of neural networks, which require terabytes of RAM, for something really serious - it is obvious that we are at a dead end - for the next breakthrough at the household level with neural networks, terabytes of memory are needed in a regular, average PC and laptop. Those. prices for 512-1024 bit memory controllers and HBM3 must fall by 2 orders of magnitude or more for complex trainable neural networks to really blossom in everyday life.

Just like now, SSDs are slowly approaching HDDs. Although they still cannot compete with them as a more reliable and capacious means of long-term data storage for less money.

This is a shame for Intel - the Apple M2 Max has a soldered memory with a bandwidth of 400 GByte/s, and not 120Gbyte/s like Intel, i.e. more than 3 times faster.

Again it is clear that Intel is counting on suckers, not knowledgeable buyers - they should have soldered 16-32GB HBM2/3, but not the shameful LPDDR5 7500.

I've been writing for a long time that given the availability of 16-24GB VRAM with a bandwidth of 500GB/s on x86 in discrete cards, it's high time to make the processor have direct access to this memory and make it possible to execute code or use it as an L4 cache for running the OS /software. Since very often people even buy gaming laptops not for games at all and VRAM stupidly and stupidly sits idle in most cases, although even on 6GB VRAM you can easily run a full kernel with a bunch of W10/W11 drivers with a 5-10 times better response than on brake DDR5 5200-5600 in slots with a shameful 80-90GB/s, instead of 500GB/s+ for VRAM video cards.


Intel, AMD, when will you at least have a 512-bit memory controller or at least a 256-bit one in the top consumer series?

Nobody mentions what other news says, and the temperature of the processor along with the RAM.
What happens if one of those modules blows, we will have to change everything, spending $600 to change everything when the processor is still in good condition, are we going crazy or what????
We will see how this affects the prices that will be imposed by Intel, processor price + RAM = surcharge.

Quote from: davidm on September 07, 2023, 19:57:18die based high bandwidth RAM could be huge. It's why the M2 chips are so capable, particularly at AI tasks. It would truly be about time for Intel to catch up on this.

Intel is the wrong comparison. For AI, compare with Nvidia!

M1 versus M2:

www.cpu-monkey.com/en/compare_cpu-apple_m2-vs-apple_m1

"Performance for Artificial Intelligence (AI) and Machine Learning (ML) [...] The performance is given in the number (trillions) of arithmetic operations per second (TOPS)."

Apple M1: 11.0 TOPS = Factor 1
Apple M2: 15.8 TOPS = Factor 1.436

Nvidia versus Apple M:

RTX 4070 is the factor 32.5 faster than Apple M1 in machine learning (KataGo visits/s)

Hence RTX 4070 is the factor 32.5 / 1.436 = 22.63 faster than Apple M2.

This is how NOT so capable Apple M2 is for AI tasks.

People who want to upgrade their RAM "later" are just wasting time. Buy enough RAM at the start. In two years, you're better off getting a new computer than adding more RAM (and computers should be easier to recycle). The only real reason for complaint is companies are stingy with RAM and charge way too much for it, and it's cheaper to buy aftermarket if it's an option.

However, die based high bandwidth RAM could be huge. It's why the M2 chips are so capable, particularly at AI tasks. It would truly be about time for Intel to catch up on this. Though it looks like they are doing baby steps rather than finding ways to include lots of RAM out of the gate.

But will it get any performance gains from the ram being closer to the CPU?

-Will