Same thing happened to Nvidia, and I'm very perplexed why this is not widely reported.
I wonder how the 3070 and 3060 are going to compare, in terms of performance per watt, when compared to the 2080/2070 series. Based on the current numbers, they'll possibly show very little improvement.
What “current numbers” do you refer to? Of course it may depend on workload, but the 3080 has, at least in one benchmark[1], better performance per watt than all compared cards (7% better than 2080 Ti, 21% than 2070, 32% than 2080, 67% than 1080 Ti). Total power consumption is up quite a bit (25% over 2080 Ti), but still get more performance than more power.
NVIDIA pushed the 3080's stock performance a little too high up the perf/watt curve. If you limit it to the TDP of the 2080 Ti, you lose 4% performance but you get much better efficiency: https://www.computerbase.de/2020-09/geforce-rtx-3080-test/6/
It’s not clear yet, since we’ve no idea if there will be a 3080 Ti. The 3090 throws out the naming convention from the past few generations, leaving it a bit of a mystery. Nvidia may do as they did with the 1080 Ti: not release till nearly a year after the 1080 (whereas 2080 & 2080 Ti were launched just a week apart).
Given the 3090 is not too much faster than 3080, it seems there may not be room there. But then again, the 1080 Ti was as fast as the Titan X. So…yea.
This is actually what I find extremely dishonest (which is, in other words, extremely good marketing) - a lot of websites are comparing the 3080 with the 2080, and reporting massive improvements... which doesn't make much sense.
I think 3080 vs 2080 Ti is the only possible comparison, but websites should make it very clear that it's an unfair comparison.
Probably the only comparison that actually makes sense is the (future) 3070 vs 2080 Ti. I'm not suprised that Nvidia pushed the release (it was supposed to be released earlier).
The end of Moore’s law has been broadly reported for the last decade and is widely understood. Nobody is likely to get huge efficiency improvements in general processors ever again, not like what we saw in the past, and it isn’t news anymore.
Algorithmic improvements, custom domain specific ASICs, and maybe quantum computing or other physical processes in the future are where large efficiency deltas might come, but for now small improvements are here to stay for all chip makers.
Moore's law may be considered dead at Intel, but TSMC does not agree.
>Wong, who is vice president of corporate research at Taiwan Semiconductor Manufacturing Corp, gave a presentation at the recent Hot Chips conference where he claimed that not only is Moore’s Law alive and well, but with the right bag of technology tricks it will remain viable for the next three decades.
“It’s not dead,” he told the Hot Chips attendees. It’s not slowing down. It’s not even sick.”
I'm not sure chip fabs are ever going to say Moore's law is dead -- I interned at Intel last summer, and Moore's law was pretty much all they could talk about. (In fact, they made very similar claims at the exact same conference [0]).
In fact, multiple gates can be created in the same transistor, in an effect SFN calls “multi-tunnel.” Multiple NOR and OR gates can thus be created from a single Bizen transistor, allowing creation of logic circuits with many fewer devices. This can result in a three-fold increase in gate density with a corresponding reduction in die size for integrated circuits based on the transistors. Summerland said that SFN is also creating a reduced device count processor architecture to enable analogue computing with Bizen transistors.
There's a lot more to the perf improvement tapering than changes in Moore's law (which is about circuit complexity increasing at given cost), namely problems translating the increasing transistor budget to ipc improvement or failing that, solving parallel programming. And clock speed improvements.
FWIW, I don’t think perf improvements are slowing down, I just think efficiency improvements in ICs are. Flops per watt of general compute isn’t moving quickly, and can’t anymore. But we can still make bigger parallel machines, design better algorithms, solve new problems, etc.
Outside HPC/ML I think our programs are now trading off useful ops per watt to take some advantage of the elusive beast called thread level parallelism. A web browser is happy to get a speedup of N by throwing 2N or 4N spinning threads at the problem if correctness and stability can be retained.
What makes sense to accelerate, how to integrate it and balance accelerators vs. general cpus, and how to expose it all to the programmer all seem like fun and interesting problems.
It is a cool time! Yeah I totally agree, and I think it’s awesome that you’re looking at it as an opportunity to learn and have fun doing it. Some people worry, and others embrace the change and make good things happen. I think I can attest to your vision since I work for a chip maker and I’m involved in the hardware & software design of some domain specific computing - it has been a blast, and we are learning all kinds of fun things.
The press follows, rather than identifies, the trend. It doesn't help that in most 'news' organizations, the news is considered entertainment and isn't terribly rigorous.
The 2xxx series wasn't terribly impressive to many so they've just started to peak out in the way that Intel did at least 5 years ago. We're heading into AMD's 4th iteration of processors that basically mop the floor with Intel from a price/performance and performance/watt standpoint and it's just starting to become accepted in the mainstream that Intel is in trouble. It will take another generation or two of products before the press catches on to the fact that what nVidia is telling everyone isn't true. Of course it will be helpful if there's some competition to point to that helps make the case.
Coincidentally there was a Twitter fight about that topic last night. From a consumer perspective there's nothing that can really be done about it. Chip designers have known for years that "Moore's Law" is slowing so it's not news to them.
Was there an issue with any of the math in the video? It was just straight forward performance per watt / performance per watt calculations as far as I could tell.
I wonder how the 3070 and 3060 are going to compare, in terms of performance per watt, when compared to the 2080/2070 series. Based on the current numbers, they'll possibly show very little improvement.