AMD Ryzen Threadripper Detailed - Why Intel HEDT is in Trouble

[trparky]

by the time they can ramp up clocks on Zen2 and the new fab, intel is going to have something new out too...

No doubt but there would be less of a gap knowing Intel's recent "lets throw them a bone" strategy.

 

[bug]

That's the thing, I don't think AMD Ryzen's issue is IPC. People have said and even AMD has confirmed that Ryzen's IPC is about what Haswell's IPC numbers were which really isn't all that bad. The key difference is that most Intel chips today are clocked higher. And as you, @bug have said, clock rate multiplied by IPC equals what I refer to as IPS or Instructions Per Second. The higher the IPS the faster the chip is perceived to be to the end user.

To increase the perceived performance of the chip you can do one of three things...
1. Increase IPC while keeping the clock rate the same.
2. Increase the clock rate while keeping IPC the same.
3. Increase both.

If AMD can improve upon both Ryzen's IPC and the clock rate it will dramatically increase the perceived performance of the processor. I figure that until Ryzen is fab'ed on the 7nm process node that Global Foundries was talking about a couple of days ago higher clock rates on Ryzen won't be possible.

4. Lower the IPC and increase clock speed (Netburst, Bulldozer).
5. Increase IPC and lower clock speed (Palomino).

They're all reasonable actions in various circumstances. Hell, you may even lower both if that allows for a killer price.

 

[EarthDog]

No doubt but there would be less of a gap knowing Intel's recent "lets throw them a bone" strategy.

lololol!!!!

So, that aside, you are talking about something that cant be released right now..and is likely not even taped out. This is more than throwing a bone, being that far away...

 

[trparky]

I'm referring to Intel's recent miniscule performance increases that they've been giving us over the last couple of years. 5% here, 7% there. That's what Intel's been giving us lately.

Now maybe with AMD back in the game Intel will actually put out something decent but I'm not holding my breath.

 

[bug]

I'm referring to Intel's recent miniscule performance increases that they've been giving us over the last couple of years. 5% here, 7% there. That's what Intel's been giving us lately.

Now maybe with AMD back in the game Intel will actually put out something decent but I'm not holding my breath.

Define decent. Because I have a feeling it means different things to different persons.

 

[trparky]

Like we used to see, huge improvements like in the 15 to 20% improvements; leaps and bounds.

 

[bug]

Like we used to see, huge improvements like in the 15 to 20% improvements; leaps and bounds.

It's not that simple anymore. If they throw 15-20% more cores at you, does that make the CPU 15-20% faster?

What you need to realize is x86 is almost 40 years old. x86_64 is almost 15. The architecture has been optimized again and again. I doubt there's much left to squeeze out of it.
All improvements we see today are all because more transistors are shoved onto a CPU, not because significant pipeline redesigns. And even shoving more transistors becomes increasingly more difficult.

Yes, I'm 100% sure Intel has been slacking as a result of AMD lagging behind, but at the same time, I'm 95% sure the x86 days you're talking about are over.

 

[Steevo]

If that is the case then great, interesting to see where pricing falls then.

AMD is also getting 99.9% yield on Ryzen dies, compared to 70%ish percent Intel is getting, making each die 30% relatively cheaper.

 

[trparky]

So you're saying that the glory days of x86 and its constant dramatic improvements are behind us. Damn.

 

[Fluffmeister]

AMD is also getting 99.9% yield on Ryzen dies, compared to 70%ish percent Intel is getting, making each die 30% relatively cheaper.

This makes me very happy, the amount of money Intel spend on coffee and biscuits just isn't fair.

 

[EarthDog]

AMD is also getting 99.9% yield on Ryzen dies, compared to 70%ish percent Intel is getting, making each die 30% relatively cheaper.

they were just past 80% in mid may which is a great yield.
https://www.google.com/search?q=ryz...droid-verizon&sourceid=chrome-mobile&ie=UTF-8


99.9% also seems incredibly unrealistic for these complex dies anyway...

 

[Steevo]

they were just past 80% in mid may which is a great yield.
https://www.google.com/search?q=ryz...droid-verizon&sourceid=chrome-mobile&ie=UTF-8


99.9% also seems incredibly unrealistic for these complex dies anyway...

https://www.overclock3d.net/news/cp...portedly_able_to_use_99_9_of_ryzen_cpu_dies/1

That is what I thought, but apparently the lower power light used during production on the mask is getting better yields as it has a cleaner result than Intel which seem to have issues with edge deterioration from higher power or intensity light damaging the mask or resist layer.

 

[EarthDog]

Ok, so it's 80% 8 cores...and apparently 99% of total yields. Impressive.

 

[trparky]

Newer and better instruction sets should help improve performance. We've seen several new instruction sets that improve upon the current architecture by including CPU instructions that reduces the work required to reach the end result. For instance, before the various forms of the SSE instruction set a program may be required to execute four or even as many as six individual operations whereas with SSE one operation can do the same job. The unfortunate part about this is that many programs are compiled for the lowest common denominator in the sense that they are compiled so as it allow it to run on just about any old processor. Yes, this increases your market share because you can then run that program on just about anything but that holds back programs in terms of being able to really use what modern processors can do.

An example of this is Firefox, it's compiled to be able to run on just about anything including an old Pentium 4 whereas Waterfox (which is based upon the same source code that Firefox is compiled from) uses all of the tricks that modern processors can do thus improve performance by making it so that it requires less CPU cycles to do the job.

Backwards compatibility is holding us back. At some point the industry has to say "Fuck it" and leave the old shit behind.

 

[goldstone77]

Newer and better instruction sets should help improve performance. We've seen several new instruction sets that improve upon the current architecture by including CPU instructions that reduces the work required to reach the end result. For instance, before the various forms of the SSE instruction set a program may be required to execute four or even as many as six individual operations whereas with SSE one operation can do the same job. The unfortunate part about this is that many programs are compiled for the lowest common denominator in the sense that they are compiled so as it allow it to run on just about any old processor. Yes, this increases your market share because you can then run that program on just about anything but that holds back programs in terms of being able to really use what modern processors can do.

An example of this is Firefox, it's compiled to be able to run on just about anything including an old Pentium 4 whereas Waterfox (which is based upon the same source code that Firefox is compiled from) uses all of the tricks that modern processors can do thus improve performance by making it so that it requires less CPU cycles to do the job.

Backwards compatibility is holding us back. At some point the industry has to say "Fuck it" and leave the old shit behind.

They has to be some incentive for companies to want to do that. Hardware is so far past software no one cares as long as it works. This will change when A.I. starts designing CPU's and programming. So, you won't have to wait much longer for much better efficiency. A.I. is learning at the rate of 8 geniuses lifetimes a day now. Within the next 10 years we should see amazing things from A.I.

 

[bug]

https://www.overclock3d.net/news/cp...portedly_able_to_use_99_9_of_ryzen_cpu_dies/1

That is what I thought, but apparently the lower power light used during production on the mask is getting better yields as it has a cleaner result than Intel which seem to have issues with edge deterioration from higher power or intensity light damaging the mask or resist layer.

So about 80% of defect free cores, which is in line with Intel. And between 80 and 99.9 it's sightly flawed, repurposed cores. Probably still in line with Intel. All in all, great to see so little silicon thrown away.