Linus Has Some Things to Say about Core i9 and X299

[EarthDog]

Does anyone know if AMD is tying the speed of the Infinity Fabric to the speed of the system RAM in Threadripper and Epyc?

I dont see why it would be any different...do you? Its all Zen...

 

[trparky]

I would have to agree with @R-T-B here, the Infinity Fabric as it exists today sucks. It really needs to have a clock speed independent of the system RAM speed. The Infinity Fabric as it exists today is a bottleneck in terms of performance on the Ryzen chip. Either run it at system RAM speed or get rid of it all together and stitch the cores together as one die as versus having two CCXs.

 

[Mr.Scott]

Every next good thing has to start somewhere.
You would rather step backward in tech than move forward and try to work the problems out?

 

[FR@NK]

Bit lower in most, bit faster in some select few, but let us say (even though falsely) that it is behind in everything, compared to a 6900K, O.K.?

- $1200 for the i7 + $500ish for the Intel mobo (and i'm being generous here) = $1700 total
- $500 for the X + $250 tops for the mobo (not even holding back like above, buying the best this time around) = $750 total

Cant really compared prices since the 6900k came out last year.

Myself and I'm sure plenty of the members in this thread would have paid close to x99 prices if ryzen would have been released last summer. Well released last summer and without all the issues it had on march 2nd.

 

[Aenra]

Cant really compared prices

Give us a break, of course i can. At the time R7 came out, we only had 2 8cores. One from Intel (and only one) and one from AMD (because it was just one.. the 'fail' ones were simply renamed as non-X).
I can and do as such compare prices, just like i compare performance between the two and only between the two; equally fairly

Important this last.. i never compared a Volvo station wagon (8c) to a Nissan Z (4c), i never compared a Rolex to my army G-Shock. I would however compare a Nissan Z with a Mazda MX. Sounds logical, but.. yeah..
So again, warped minds. Take it or leave it, but do smell the coffee guys, it's getting stale already and here we are, still debating on it.

 

[FR@NK]

Car analogies dont translate really well into computer tech....

 

[R-T-B]

Car analogies dont translate really well into computer tech....

This is techpowerup, it never stops us.

 

[cdawall]

Car analogies dont translate really well into computer tech....

They do depending on scenario. That wasn't one.

 

[R-T-B]

They do depending on scenario. That wasn't one.

I'm actually of the opinion they are actually very poor, but then, I don't drive.

 

[dorsetknob]

They do depending on scenario. That wasn't one.

hows this then
"bulldozer was volkswagon with 4 pregnent women passengers" gotta say it carried 8 persons in total

 

[EarthDog]

They do depending on scenario. That wasn't one.

Just QFT.. because I love them.. and most of mine work out fine... at least, for those that understand cars and computers. For the daft, well, not so much.

 

[HTC]

I would have to agree with @R-T-B here, the Infinity Fabric as it exists today sucks. It really needs to have a clock speed independent of the system RAM speed. The Infinity Fabric as it exists today is a bottleneck in terms of performance on the Ryzen chip. Either run it at system RAM speed or get rid of it all together and stitch the cores together as one die as versus having two CCXs.

I think you have it backwards: it's because this "fabric" exists that AMD is able to produce high core count CPUs for cheap, which is why AMD may very well be able to surpass Intel when threadripper and epyc launch.

It's very costly to produce big chips that run @ intended speeds because of the yields. The existence of infinity fabric allows for very big chips to be produced much more easily because they're essentially small chips "glued together" by the fabric.

This approach has it's drawbacks because in it's smaller size (only 2 CCXs), this fabric has no room to "stretch it's legs" but it's advantage is that you won't lose much speed when scaling the core count. This is not possible with Intel's architectures (as well as previous AMD's architectures) because, when the core count increases, you'll be forced to reduce the clocks by a significant amount. If infinity fabric works as intended, which is yet to be seen in larger chips, the reduction in clock speeds will be much smaller.

 

[Frick]

They do depending on scenario. That wasn't one.

Nah, not really. I can't remember a single apropriate car analogy actually. Well they can work if you have lots of chrome in your case and it actually is styled after a classic Cadillac, but technically a car and a computer is two completely different beasts operating on different principles, even though the borders are blurring the more tech we squeeze into the cars. Why not use computer parts to describe computer parts?

 

[cdawall]

Nah, not really. I can't remember a single apropriate car analogy actually. Well they can work if you have lots of chrome in your case and it actually is styled after a classic Cadillac, but technically a car and a computer is two completely different beasts operating on different principles, even though the borders are blurring the more tech we squeeze into the cars. Why not use computer parts to describe computer parts?

Plenty work. I mean more heat is bad, make more hp required cooling goes up, everything is a wear item when racing (overclocking), people love specific brands, people hate specific brands, half the guides posted for fixing them are wrong, repair shops can be crooks.

I mean the list goes on.

 

[Frick]

Plenty work. I mean more heat is bad, make more hp required cooling goes up, everything is a wear item when racing (overclocking), people love specific brands, people hate specific brands, half the guides posted for fixing them are wrong, repair shops can be crooks.

I mean the list goes on.

But they are also unneccesary. Heat = bad is not a car thing. Neither is more work = more cooling. Hating brands is definitely not a car thing. Overclocking is most definitely not like racing.

 

[r9]

How would cherry picked review samples effect reviews results? Would effect overclocks for sure but it wouldn't effect the performance of the chips running at the stock clock speeds.

We dont really need reviews anyways. These chips are similar IPC to current broadwell-e chips. Since we already have a good idea how broadwell-e performs, we can conclude that because skylake-x chips are running faster they will be faster. They also will cost less per core. Skylake-x will be the best performance per dollar jump compared to the previous generation we have ever seen from Intel. I'm sure anyone looking to spend $500 on a processor will gladly shell out the extra Benjamin for the 7820x.

AMD: You're welcome.

 

[EarthDog]

But they are also unneccesary. Heat = bad is not a car thing. Neither is more work = more cooling. Hating brands is definitely not a car thing. Overclocking is most definitely not like racing.

Random mentions of car/pc associations clearly mean nothing when they are not wrapped in context.

Make no mistake about it though, they can be spot on....though weve seen plenty of fails. I dont blame it on the differences between them, but the user dishing it out and the daft/not daft users reading it.

 

[trparky]

It's very costly to produce big chips that run @ intended speeds because of the yields. The existence of infinity fabric allows for very big chips to be produced much more easily because they're essentially small chips "glued together" by the fabric.

OK, I get that but perhaps they should start trying to get the manufacturing to be better when making this silicon so that perhaps in the future the CCXs will be six-core CCXs instead. There's no reason why a CCX needs to only be a quad-core, in fact there's no "quad" in CCX since it means "CPU complex". That way with a six-core CCX they can make the lower-end dual, quad, and six core Ryzen CPUs as one single CCX ditching the Infinity Fabric and keeping it only for those chips that need it due to higher core count.

This approach has it's drawbacks because in it's smaller size (only 2 CCXs), this fabric has no room to "stretch it's legs"

That's because the Infinity Fabric is being gimped by not having more than two CCXs to work with. Like I said above, if they can get more cores on a CCXx there would be no need for the Infinity Fabric on lower-end CPUs since a single CCX would have all of the cores needed for those lower-end CPUs on one single CCX.

but it's advantage is that you won't lose much speed when scaling the core count.

Yes, it seems that as you add more CCXs into the mix (Threadripper and Epyc) the Infinity Fabric scales far better. Perhaps the Infinity Fabric can run at full system RAM speed instead of half of it in the current crop of desktop Ryzen chips.

This is not possible with Intel's architectures (as well as previous AMD's architectures) because, when the core count increases, you'll be forced to reduce the clocks by a significant amount.

Yeah, that I know since I've already seen the next crop of Intel x299-based chips that have lower clock speeds due to higher core count.

If infinity fabric works as intended, which is yet to be seen in larger chips, the reduction in clock speeds will be much smaller.

They really need to get the manufacturing of these chips to be better since apparently the Threadripper and Epyc chips (according to rumors) have much higher out-of-the-box base clock speeds than the current crop of desktop Ryzen chips. The only thing that I can think of is that AMD is saving the very best silicon for these high-end to ultra high-end CPUs while leaving the less-than-best silicon for the desktop CPUs. Getting the manufacturing to be better would solve this thus allowing for the desktop CPUs to clock higher out of the box just like the high-end Threadripper and Epyc chips.

 

[cdawall]

But they are also unneccesary. Heat = bad is not a car thing. Neither is more work = more cooling. Hating brands is definitely not a car thing. Overclocking is most definitely not like racing.

Overclocking is just like racing throw money at it and you can have the best parts doesn't mean you will have the fastest car or computer. I happen to do both...user side is by far the most important.

 

[HTC]

OK, I get that but perhaps they should start trying to get the manufacturing to be better when making this silicon so that perhaps in the future the CCXs will be six-core CCXs instead. There's no reason why a CCX needs to only be a quad-core, in fact there's no "quad" in CCX since it means "CPU complex". That way with a six-core CCX they can make the lower-end dual, quad, and six core Ryzen CPUs as one single CCX ditching the Infinity Fabric and keeping it only for those chips that need it due to higher core count.

For all we know that might be their plan for Zen+.

Yes, it seems that as you add more CCXs into the mix (Threadripper and Epyc) the Infinity Fabric scales far better. Perhaps the Infinity Fabric can run at full system RAM speed instead of half of it in the current crop of desktop Ryzen chips.

You do realize that within the same CCX, the latency is way better then Intel's, right? It's between CCXs that the latency is worse. Perhaps they should have limited the DDR4 RAM speed to 3000 MHz while having the fabric run @ RAM's speed? I'm guessing they chose half RAM because of some limitation @ higher RAM speeds? Dunno: just an hypothesis.

Yeah, that I know since I've already seen the next crop of Intel x299-based chips that have lower clock speeds due to higher core count.

You just have to look @ high cores count xeons and opterons: their speeds decrease as the cores count goes up.

They really need to get the manufacturing of these chips to be better since apparently the Threadripper and Epyc chips (according to rumors) have much higher out-of-the-box base clock speeds than the current crop of desktop Ryzen chips. The only thing that I can think of is that AMD is saving the very best silicon for these high-end to ultra high-end CPUs while leaving the less-than-best silicon for the desktop CPUs. Getting the manufacturing to be better would solve this thus allowing for the desktop CPUs to clock higher out of the box just like the high-end Threadripper and Epyc chips.

I've heard threadripper and epyc will have similar speeds to the 1800X: perhaps just a small decrease. This should only be possible if the fabric works as intended.

Ofc AMD is saving their very best silicon for epyc and threadripper: think along the lines of threadripper with 16c / 32 t using just 140w and running @ 1800X clocks (just a guess)?

 

[trparky]

I've heard threadripper and epyc will have similar speeds to the 1800X: perhaps just a small decrease.

I thought I read somewhere that Threadripper was to start at 4 GHz as an out-of-box clock.

They really need to increase the base clocks on these things, at least a base clock of 4 (maybe 4.2) GHz. We're already pushing the envelope in terms of what this silicon can do since anything past 4 GHz on the current crop of Ryzen CPUs causes the chip to throttle internally due to heat issues.

 

[HTC]

I thought I read somewhere that Threadripper was to start at 4 GHz as an out-of-box clock.

They really need to increase the base clocks on these things, at least a base clock of 4 (maybe 4.2) GHz. We're already pushing the envelope in terms of what this silicon can do since anything past 4 GHz on the current crop of Ryzen CPUs causes the chip to throttle internally due to heat issues.

Seriously doubt this: double the cores means nearly 2 times the wattage (assuming better binned chips that need less power to run @ R7 speeds. By starting @ 4 GHz, the required power would negate it's efficiency. It may very well be overclockable to those speeds, but not out-of-box, IMO.

Check this pic, where highlighted:



AMD doesn't need 4 GHz for 16 cores when Intel has it's 16 core chip @ 2.1 GHz. If they run it @ 3 GHz it should run circles around Intel's 16 core chip.

 

[trparky]

I meant that they need to increase the clock speed on their desktop chips. either that or find a way to increase single-thread IPC. Either bump the speed up or increase IPC. Whatever they choose to do they need to do it fast. There's still many of us that need better single-core IPC performance. If Ryzen v2,0 can somehow make that possible it may just become the Intel killer that we all wanted in the first place.

 

[HTC]

I meant that they need to increase the clock speed on their desktop chips. either that or find a way to increase single-thread IPC. Either bump the speed up or increase IPC. Whatever they choose to do they need to do it fast. There's still many of us that need better single-core IPC performance. If Ryzen v2,0 can somehow make that possible it may just become the Intel killer that we all wanted in the first place.

You do realize that AMD has made a gigantic comeback that pretty much 99.9% of the tech world never thought possible. If threadripper and epyc deliver what we think they'll deliver, that will be nothing short of a miracle with the resources AMD has had lately.

Try not to be too greedy with the 1st Zen generation. Who knows: perhaps the 2nd generation will bring a bump to IPC to an already outstanding architecture!

 

[trparky]

And why can't I be greedy? Why can't I have my cake and eat it too?

There's no question, I know that AMD came from way behind; this I know very much so. But in some circumstances Ryzen v1.0 is not good enough. There's a lot of old software that will never see Ryzen optimizations and that's where more raw IPC is needed.