4 ms·
Apple is doing something interesting but it's not at all unclear how they're doing it: > outpace on a per watt basis with AMD/Intel (and be at least competitiv
by AnthonyMouse 3y ago
Apple is doing something interesting but it's not at all unclear how they're doing it:
> outpace on a per watt basis with AMD/Intel (and be at least competitive with on an absolute basis)
They're the first on TSMC's new process nodes, so whenever a new node comes out, people will compare the new Apple chip on a new node to the previous generation AMD chip on the previous node. If you compare them this way then the newer node outperforms the older one as expected. Then AMD releases one on the new node and the advantage basically disappears, or is just making different trade offs, e.g. selling chips with more cores which consequently have a higher TDP in exchange for higher multi-thread performance.
Intel hasn't been competitive with either of them on power consumption for some time because Intel's fabrication is less power efficient than TSMC's.
But if Apple's advantage is just outbidding everyone at TSMC, that only lasts until TSMC builds more fabs (the lead times for which get shorter as the COVID issues subside), or someone else makes a better process.
> For example, they have a unified 800GB/s memory bandwidth with zero copy sharing between GPU and CPU. That’s not something anyone else has managed yet.
It's also not something which is at all difficult to do. It's a straightforward combination of two known technologies. Integrated GPUs have unified memory and discrete GPUs have high memory bandwidth.
There is no secret to attaching high bandwidth memory to a CPU with an integrated GPU, it's just a trade off that traditionally isn't worth it. CPUs typically have more memory than GPUs but if it's unified and you want to use the fast memory you're either going to get less of it or pay more, and CPU applications that actually benefit from that amount of memory bandwidth are uncommon.
One of the rare applications that do benefit from it are LLMs, and it's possible that's enough to create market demand, but there's no real question of if they can figure out how to make that -- everybody knows how. It's only a question of if customers want to pay for it.
And what we may see is something better -- high bandwidth memory as a unified L4 cache. So then your CPU+iGPU gets the amount of HBM traditionally found on a discrete GPU and the amount of DRAM traditionally attached to a CPU and you get the best of both worlds with no increase in cost over a CPU + discrete GPU. But currently nobody offers that -- the closest is the Xeon Max which has HBM but no integrated GPU.
And none of these are why Qualcomm isn't competitive with Apple -- they're not competitive with AMD or Intel either. It's not the fabs or the trade offs. Their designs just aren't as good. But all that means is they should hire more/better engineers.
- lern_too_spel 3y agoAre you looking at the same benchmarks the rest of us are looking at? https://news.ycombinator.com/item?id=38076018 https://news.ycombinator.com/item?id=38076018