Discussion AMD Ryzen MegaThread! FAQ and Resources

[-Fran-]

AI won't replace engineers, but will simplify the tasks to the point where instead of a full team of them, you just need 1 or 2 to tweak and supervise the AI doing the work.

There's always going to be a human element in everything, because laws make it so that no AI in the foreseeable future can cope with the stupidness of regulations and their changes.

And I agree with gamerk on the specialized cores argument. It makes little sense to add dedicated cores to stuff when they come with an implicit expiration date. Moreover, for security things, you'll be tied to upgrade cycles and, more importantly, security holes. More pragmatically, for dedicated hardware you also need specialized software that can use that effectively. When you write software for a general purpose CPU, it's easy to get it to market fast, but when you need to support specific stuff, it becomes a big PITA.

Cheers!

 

[Eximo]

AI can also be truly stupid when you give it bad or biased data. Sometimes it will pick out the criteria that is least important and blow it out of proportion. I recall one study that was quite funny. They tasked AI with filtering a pile of resumes down to candidates worth interviewing. End result was that it learned to determine the difference between male and female applicants by their use of language and began rejecting all female applicants. It was completely ignoring qualifications and experience because the data of successful retention and hiring heavily favored men. A good follow up study would have been to run it through a female dominated profession and see how it did there.

I remember reading about AI designed electronics many years ago, iterative evolution design or something like that. Don't know if that ever became a thing. It was essentially waveform in, waveform out, and it would try adding components and see if if got farther or closer to the desired outcome, and do this millions of times. It made truly impressive designs, but often used hundreds more components then a standard design.

 

[goldstone77]

I'm a mechanical engineer- the 'AI can design it better' is a huge sales pitch in the CAD industry at the moment. The problem though, is that as you say you have to give the AI the rules to follow- which is usually based on an FEA study of a part. In my experience all the AI does is remove every scrap of material it can whilst still meeting the load case given the the FEA study- however that results in a part that has no capacity to withstand unexpected forces (e.g. side loading, shock loading etc). The parts are also totally impossible to manufacture save for 3D printing (another buzz term that is being held up as the answer to everything yet in reality is only suited to very limited production volume and parts produced in this way have a lot of limitations).

There are positives and negatives, but it's only going to improve over time. Everything is going to become automated at some point, because that is the direction everything is heading. On a side note, what is going to happen when unemployment starts to skyrocket because of automation?

 

[goldstone77]

Also, for the record I would like to say I welcome our A.I. god, and robot overlords!

 

[gamerk316]

I'm a mechanical engineer- the 'AI can design it better' is a huge sales pitch in the CAD industry at the moment. The problem though, is that as you say you have to give the AI the rules to follow- which is usually based on an FEA study of a part. In my experience all the AI does is remove every scrap of material it can whilst still meeting the load case given the the FEA study- however that results in a part that has no capacity to withstand unexpected forces (e.g. side loading, shock loading etc). The parts are also totally impossible to manufacture save for 3D printing (another buzz term that is being held up as the answer to everything yet in reality is only suited to very limited production volume and parts produced in this way have a lot of limitations).

In a case like this, you could argue the problem is the rules given to the AI aren't stringent enough.

 

[gamerk316]

There are positives and negatives, but it's only going to improve over time. Everything is going to become automated at some point, because that is the direction everything is heading. On a side note, what is going to happen when unemployment starts to skyrocket because of automation?

I honestly believe that's where we are heading, and when that happens is going to be defined by how quickly the AI field advances. I do feel jobs are going to be at a premium in 20 years or so; we can only destroy so many fields before growth in other industries can't pick up the slack.

At the end of the day, only us Software Engineers are truly safe.

 

[cdrkf]

In a case like this, you could argue the problem is the rules given to the AI aren't stringent enough.

My point is that what is being hailed as "AI" is just a dumb optimisation tool. It isn't intelligent- it's function is fixed.

I don't think there's much risk to jobs in the design and engineering fields any time soon. The tools people are talking about are just that - tools that can allow the task to be done better (when used in the correct context) and arguably reduce time by automating the donkey work. They aren't actually taking over the "thinking" part though.

Show me an AI system that can plan out a real design (involving more than a single component), based on nothing more than a brief and then I'll start to worry... I've yet to see anything that can actually do that.

 

[-Fran-]

My point is that what is being hailed as "AI" is just a dumb optimisation tool. It isn't intelligent- it's function is fixed.

I don't think there's much risk to jobs in the design and engineering fields any time soon. The tools people are talking about are just that - tools that can allow the task to be done better (when used in the correct context) and arguably reduce time by automating the donkey work. They aren't actually taking over the "thinking" part though.

Show me an AI system that can plan out a real design (involving more than a single component), based on nothing more than a brief and then I'll start to worry... I've yet to see anything that can actually do that.

Oof... This is actually a big bag to discuss, as it depends on everyone's understanding of what constitutes "thinking". I'd argue, within the context of AI's, the term is used a bit loosely compared to, say, psychology. Machines won't have creative thinking (which is one of the base elements of "thinking"), but they will have pretty much everything else: learning, decision making and memory. If you simplify the "creative" aspect to just a more complex method of "try-adapt-learn-try again", you could argue machine AI, even at this stage, has a degree of intelligence and can replace a lot of common-ground tasks (simple ones).

I've seen lab reports of AIs getting close to parrot's level of intelligence. So we're not that far away from getting them to perform menial tasks and fully replace job positions that don't require complex decisioning or complex creative process. And for those, there's a HUGE bunch.

Cheers!

 

[gamerk316]

Oof... This is actually a big bag to discuss, as it depends on everyone's understanding of what constitutes "thinking". I'd argue, within the context of AI's, the term is used a bit loosely compared to, say, psychology. Machines won't have creative thinking (which is one of the base elements of "thinking"), but they will have pretty much everything else: learning, decision making and memory. If you simplify the "creative" aspect to just a more complex method of "try-adapt-learn-try again", you could argue machine AI, even at this stage, has a degree of intelligence and can replace a lot of common-ground tasks (simple ones).

I've seen lab reports of AIs getting close to parrot's level of intelligence. So we're not that far away from getting them to perform menial tasks and fully replace job positions that don't require complex decisioning or complex creative process. And for those, there's a HUGE bunch.

Cheers!

I think the real problem here is one of definition; "intelligence" itself doesn't have a clear definition.

I would take the side that what we have today are just highly-advanced cases of specialized learning. There's no way for these AIs to learn new behaviors spontaneously, and there's no way to apply said behaviors to a slightly different use case. Making a good SC2 AI doesn't mean it will be anything but rubbish in SC1. Making an AI that can beat World 1-1 of SMB won't help it in World 1-2. They learn for very limited use cases, but have no ability to apply that knowledge to different, but similar cases. To me, that implies no real intelligence.

 

[goldstone77]

https://www.bbc.com/news/business-48760799
Up to 20 million manufacturing jobs around the world could be replaced by robots by 2030, according to analysis firm Oxford Economics.

 

[Rogue Leader]

https://www.bbc.com/news/business-48760799
Up to 20 million manufacturing jobs around the world could be replaced by robots by 2030, according to analysis firm Oxford Economics.

Ok guys as exciting as AI is, lets get back on to the topic. There are other threads to discuss such things, or make a new one.

 

[-Fran-]

https://www.tomshardware.com/uk/reviews/amd-threadripper-3970x-review

The Threadripper 3970X and 3960X delivered devastating threaded performance in their respective price ranges, often trouncing Intel's most exotic silicon. Intel's Xeon W-3175X is ill-suited to take on the comparatively power-sipping Threadripper processors on a power efficiency basis, not to mention pricing. Just for comparison's sake – the overclocked W-3175X pulled 768 watts under load, while the overclocked Threadripper 3970X peaked at 356 watts while often providing more performance in threaded workloads. That math is easy.

Oof.

And from AnandTech's Ian:

I have never used the word ‘bloodbath’ in a review before. It seems messy, violent, and a little bit gruesome. But when we look at the results from the new AMD Threadripper processors, it seems more than appropriate.

Double oof.

Cheers!

 

[cdrkf]

The last 3 years of computer hardware feels like the early 90's with successive CPU releases trouncing the previous generation. I think the most amusing bit is AMD aren't even finished yet, with the 3990X already confirmed for early 2020 release. Intel simply have no answer at the moment on the desktop (or server) side of things.

On the bright side for Intel they are still looking strong in the laptop space.

 

[aldaia]

Anyone else noticed the gap between the 32 core 3970X and the 64 core 3990X?
Is that an implicit confirmation for 48 core 3980X ?

 

[cdrkf]

Anyone else noticed the gap between the 32 core 3970X and the 64 core 3990X?
Is that an implicit confirmation for 48 core 3980X ?

Well due to the construction we know a 48 core is feasible, so I would say it's highly likely. I also wonder if the range will get fleshed out a bit further down the road- there are lots of possible sku's they could construct. I think a 16 core is likely on the TRX4 socket (for users who want the memory bandwidth and PCIe lanes over a 3950x, like they did with the 8 Core part on TR4). I'm not sure what other core counts really make sense, maybe some options with less cache? I guess it depends on what type of defects they are getting in the chiplets....

 

[Eximo]

Hmm, one 3980X or 8 Ryzen 3600... Depends on the price of the I/O section and the target price. 8 motherboards vs one, though I can't imagine there is much money in selling B450 chipsets. Probably why they are waiting on the 8 CCX configs, get more holiday sales of consumer chips.

 

[aldaia]

Hmm, one 3980X or 8 Ryzen 3600... Depends on the price of the I/O section and the target price. 8 motherboards vs one, though I can't imagine there is much money in selling B450 chipsets. Probably why they are waiting on the 8 CCX configs, get more holiday sales of consumer chips.

It could also be 6 Ryzen 3700. You can get 48 cores either using 8x 6-core dies or 6x 8-core dies.
In fact they use both options for EPYC. 7552 seems to be 6 dies, while 7642 clearly is a 8 die component.

 

[cdrkf]

It could also be 6 Ryzen 3700. You can get 48 cores either using 8x 6-core dies or 6x 8-core dies.
In fact they use both options for EPYC. 7552 seems to be 6 dies, while 7642 clearly is a 8 die component.

The advantage of the latter offering 64mb more cache and an extra 2 memory channels

 

[aldaia]

The advantage of the latter offering 64mb more cache and an extra 2 memory channels

More cache, yes. Memory channels, no. All Epyc are 8 channel no matter how many dies or cores.

 

[cdrkf]

More cache, yes. Memory channels, no. All Epyc are 8 channel no matter how many dies or cores.

Ah yes- I was thinking of the previous generation.... on the new design the memory channels are on the IO die in the middle... doh!