Overclock.net banner

1 - 20 of 46 Posts

·
there is no easy way out
Joined
·
3,255 Posts
Discussion Starter #1
SOURCE


Quote:
What if I told you that the world's fastest desktop processor available today was never originally planned? That's right, in a revealing interview with Forbes, AMD'ers Sarah Youngbauer and James Prior revealed that the company's pride and joy, Ryzen Threadripper, was developed by engineers in their spare time.
 

·
Joined
·
2,327 Posts
Please tell me some rebels at RTG are secretly working on a monster GPU architecture to take revenge on Nvidia...
 

·
there is no easy way out
Joined
·
3,255 Posts
Discussion Starter #5
Quote:
Originally Posted by Dennybrig View Post

Nice article but i think that this news appeared on Forbes first (i think you should link to the Forbes article.
Yeah this is from forbes via WCCFTech. I've included the WCCFTech since we are supposed to read the full story from there and there you will obviously going to know that it is taken from Forbes. I think it doesn't matter much since Forbes as the original source is already implied when you reach WCCFTech.
 

·
Registered
Joined
·
2,725 Posts
It's somehow surprising that someone thought about doing an intermediate chip between Epyc and Ryzen, using exactly the same technology that both of those have?

Stop the presses.
 

·
PC Evangelist
Joined
·
47,394 Posts
Lol. TR is nothing but a Epyc CPU lol. Its like saying Intel never planed the 18-Core i9. They did not.
 

·
there is no easy way out
Joined
·
3,255 Posts
Discussion Starter #8
Quote:
Originally Posted by ZealotKi11er View Post

Lol. TR is nothing but a Epyc CPU lol. Its like saying Intel never planed the 18-Core i9. They did not.
But making an exclusive socket for desktop HEDT targeting the bests of Intel's hardware was undoubtedly a bold and surprise move from AMD, all of a sudden. Specially when everyone was considering the best AMD could do was Intel SB-IvyB level performance, also when they had this much little money. Imo, Threadripper was "more than a surprise". The 'surprise' was 40% >52% IPC improvement(at stock memory frequency/timings).
 

·
Registered
Joined
·
2,725 Posts
Quote:
Originally Posted by sumitlian View Post

But making an exclusive socket for desktop HEDT targeting the bests of Intel's hardware was undoubtedly a bold and surprise move from AMD, all of a sudden. Specially when everyone was considering the best AMD could do was Intel SB-IvyB level performance, also when they had this much little money. Imo, Threadripper was "more than a surprise". The 'surprise' was 40% >52% IPC improvement(at stock memory frequency/timings).
Yet, aside from the dummy chip claim, nothing else is unique to TR.

This makes TR a no-brainer product. If AMD's management didn't think of it then maybe they shouldn't have admitted that.
 

·
Registered
Joined
·
844 Posts
TR was a major surprise to Intel for sure, they were ready to counter the 8 core Ryzen with their 6 core i7 series, but AMD actually beat Intel's top of the line 10 core for the months it takes for them to release the 18 core monster.
 

·
Registered
Joined
·
2,725 Posts
Quote:
Originally Posted by Nightbird View Post

TR was a major surprise to Intel for sure, they were ready to counter the 8 core Ryzen with their 6 core i7 series, but AMD actually beat Intel's top of the line 10 core for the months it takes for them to release the 18 core monster.
If TR was a surprise to Intel then that's because they didn't have enough corporate espionage to understand Ryzen and Epyc.

AMD was already using Infinity Fabric for Ryzen, despite the latency drawback. It was already trying to entice consumers with more cores and threads, despite the single thread/IPC deficit versus Intel. The Zen core was already designed around reduced size and being united by fabric to give more cores and threads versus higher IPC, especially when it comes to AVX2/512. Given that Epyc is also about more cores and threads for less money, and uses Infinity Fabric, there is literally nothing special about TR at all (other than the dummy chip implementation, perhaps). It's the obvious step between Ryzen and Epyc. Dead obvious.
 

·
there is no easy way out
Joined
·
3,255 Posts
Discussion Starter #12
Quote:
Originally Posted by Nightbird View Post

TR was a major surprise to Intel for sure, they were ready to counter the 8 core Ryzen with their 6 core i7 series, but AMD actually beat Intel's top of the line 10 core for the months it takes for them to release the 18 core monster.
Exactly, that is what nobody is seeing, AMD shattered plans of Intel, made Intel do architecture change for a usual segment(Intel Enthusuast segment) that hasn't happend in a decade, and this is the significance. Also when is the last time we heard Intel launched three different sockets for desktop (including upcoming CoffeeLake) in a year !?
 

·
Registered
Joined
·
2,725 Posts
This story sounds like PR from AMD. Drum up excitement about TR.

"It's special. It's rare. You weren't going to get access to it. But, hey, look at your luck."

"Big bad management didn't get in your way. Real engineers making real products that the serious enthusiasts want. And, hey, look at how nice and organic AMD's management actually is. We listen to people."

The first thing I thought of when I found out about Epyc was... I wonder how much better the performance will be on the desktop versus Ryzen due to quad channel RAM. I wonder if it would be possible to disable cores to get clocks up. Then, I figured it would cost too much and the boards wouldn't facilitate overclocking. Plus, the form factor would be for server cooling.

But, yeah, if a nobody like me can think up TR immediately, upon hearing about Ryzen and Epyc, then it shouldn't require millions of dollars of development money - or genius engineering passion.
 

·
there is no easy way out
Joined
·
3,255 Posts
Discussion Starter #14
Quote:
Originally Posted by superstition222 View Post

If TR was a surprise to Intel then that's because they didn't have enough corporate espionage to understand Ryzen and Epyc.
There are unlimited perspectives of seeing this release. This is one of them.
tongue.gif

Though that doesn't mean I don't believe you.
Quote:
Originally Posted by superstition222 View Post

AMD was already using Infinity Fabric for Ryzen, despite the latency drawback. It was already trying to entice consumers with more cores and threads, despite the single thread/IPC deficit versus Intel. The Zen core was already designed around reduced size and being united by fabric to give more cores and threads versus higher IPC, especially when it comes to AVX2/512. Given that Epyc is also about more cores and threads for less money, and uses Infinity Fabric, there is literally nothing special about TR at all (other than the dummy chip implementation, perhaps). It's the obvious step between Ryzen and Epyc. Dead obvious.
Latency drawback looks real in artificial benchmarks, and even technically it does exist at default core settings, I agree.
But how many real world workloads have you seen that show any significant performance decrease in single or multithreading applications due to that latency ?
I've not even heard of even any perceptible performance degrade duo to that latency claim. May be the problem doesn't exist that much in the real world as it is advertized just because of some scores of synthetic benchmarks.

Also people seem to neglect memory bandwidth efficiency of Zen cores. It probably applys to Threadripper as well, because both Ryzen's stock memory bandwidth (2400 MHz) and OC (3466) shows ~95% efficiency in memory writes. I think that is a little bit better than competition. Than means relatively reliable and uninterruptedly multitaksing performance. And this is a fact you can check anywhere.

There simply not any market for AVX512 till now for desktop HEDT, you can't justify Intel's purchase only because of that. Good reasons will be a little bit higher IPC as well as more IPS if somebody wants absolute best regardless of price, then go for 18 core one. Only things beside a little IPC (not considering IPS since OCed effciency of Skylake-X is not that good and sometimes is unreliable) are if you literally do H.265 transcoding most of the time and it does utilize AVX2 very much and Skylake-X does come with 16 more 256 bit registers(per core) as compared to Kaby Lake. Other than that Threadripper too is indeed is an option for consumers, and all this was revealed to us, from first news to actaul products within 6 months(I don't exactly remember).
 

·
Waiting for 7nm EUV
Joined
·
11,527 Posts
Quote:
Originally Posted by superstition222 View Post

It's somehow surprising that someone thought about doing an intermediate chip between Epyc and Ryzen, using exactly the same technology that both of those have?

Stop the presses.
Quote:
Originally Posted by superstition222 View Post

This story sounds like PR from AMD. Drum up excitement about TR.

"It's special. It's rare. You weren't going to get access to it. But, hey, look at your luck."

"Big bad management didn't get in your way. Real engineers making real products that the serious enthusiasts want. And, hey, look at how nice and organic AMD's management actually is. We listen to people."

The first thing I thought of when I found out about Epyc was... I wonder how much better the performance will be on the desktop versus Ryzen due to quad channel RAM. I wonder if it would be possible to disable cores to get clocks up. Then, I figured it would cost too much and the boards wouldn't facilitate overclocking. Plus, the form factor would be for server cooling.

But, yeah, if a nobody like me can think up TR immediately, upon hearing about Ryzen and Epyc, then it shouldn't require millions of dollars of development money - or genius engineering passion.
Yeah, this story does seem to be taking people for fools quite a bit. And the thing is, they don't need to drum up Threadripper, they managed to get ahead of Intel, even if in the long run it's going to get tougher as Intel releases the 14, 16 and 18 core models, they will probably hold on to the price / performance crown - and they can always add an extra die and go up to 24 cores to keep bringing the fight to Intel.

Here's what I think that actually happened: they took advantage of their past 10 year history of never having competed with Intel's HEDT platforms to surprise Intel by hiding the sideproject Threadripper as best as they could so that Intel would only hear about it when it was inevitable - that explains the later release of the 12 core and the even later release of the 14, 16 and 18 core versions; Intel obviously only heard about TR very late in the game, when motherboard makers and other entities inevitably had to get on board.

But until then, it was relatively easy to hide TR because it uses the same socket as EPYC and if you look at an engineering sample photo it has the same 4 dies, little would one know that 2 are dummies, and it also uses the same Infinity fabric to connect the dies, none of this is news and the way the information is presented seems like they made up something new, when that's not really the case; they just leveraged the possibilities of the scalable design and re-used the EPYC socket.

They should tone down this kind of PR because they don't need it, the product speaks for itself. Of course they can't say that they outsmarted Intel because they had never signalled going against the HEDT platform, the plan was to bring HEDT-like performance (6900K performance) to the mainstream, so it's easy to see how Intel would believe that AMD would focus their resources on that and EPYC alone.

Was it a side-project with a secretive team? Surely. Was it the romantic thing that they are trying to portray? Not really. But I'll give them that, having the notion that they had a very good chance to outsmart Intel must have been exciting, that's the actual romantic part that they can't admit to, but it's probably what happened.
 

·
Joined
·
2,327 Posts
Quote:
Originally Posted by sumitlian View Post

Exactly, that is what nobody is seeing, AMD shattered plans of Intel, made Intel do architecture change for a usual segment(Intel Enthusuast segment) that hasn't happend in a decade, and this is the significance. Also when is the last time we heard Intel launched three different sockets for desktop (including upcoming CoffeeLake) in a year !?
Yeah, it's like Lisa played Intel like a damn fiddle
biggrin.gif
 

·
Untitled.jpg
Joined
·
970 Posts
Wasn't the intention of the Zen architecture that it would scale up? I have a hard time believing designers/management/directors or whoever would scale it up for Epyc and not realise the same could be done for HEDT.. IDK, if they hadn't even considered HEDT that's the sort of oversight that should probably get someone fired. Employees shouldn't have to be fixing mismanagement in their own time.

Basically if this is true, someone needs firing and those engineers need a raise and promotion.
 

·
Registered
Joined
·
451 Posts
Remember when intel told us Threadripper were just two desktop cpu's glued together?

"The efficiency in manufacturing came from the fact that we didn't need entire wafer runs just to produce Threadrippers, we could use Ryzens and that was the big key in Threadripper's success."

Who's laughing now?!
Quote:
Originally Posted by Lexi is Dumb View Post

Wasn't the intention of the Zen architecture that it would scale up? I have a hard time believing designers/management/directors or whoever would scale it up for Epyc and not realise the same could be done for HEDT.. IDK, if they hadn't even considered HEDT that's the sort of oversight that should probably get someone fired. Employees shouldn't have to be fixing mismanagement in their own time.

Basically if this is true, someone needs firing and those engineers need a raise and promotion.
Simplicity is the reason. Epyc originally had 8 core versions with the full 128 pcie lanes planned. The only difference Threadripper made is that their boards are cheaper to produce because there are less traces.
 

·
Registered
Joined
·
1,116 Posts
Quote:
Originally Posted by sumitlian View Post

There are unlimited perspectives of seeing this release. This is one of them.
tongue.gif

Though that doesn't mean I don't believe you.
Latency drawback looks real in artificial benchmarks, and even technically it does exist at default core settings, I agree.
But how many real world workloads have you seen that show any significant performance decrease in single or multithreading applications due to that latency ?
I've not even heard of even any perceptible performance degrade duo to that latency claim. May be the problem doesn't exist that much in the real world as it is advertized just because of some scores of synthetic benchmarks.

=
You won't see the latency issues become a real problem until you actually start to multi-task, and deal with many clients/operations. You don't hear much about it in the "Real world" because most people are too busy trying to get that last 100 Mhz, or don't run such operations.

Others...just don't care enough.
tongue.gif


Oh, and Intel *still* has the better IMC...by a fair margin.
 

·
Super Moderator
Joined
·
9,195 Posts
Quote:
Originally Posted by tpi2007 View Post

Yeah, this story does seem to be taking people for fools quite a bit. And the thing is, they don't need to drum up Threadripper, they managed to get ahead of Intel, even if in the long run it's going to get tougher as Intel releases the 14, 16 and 18 core models, they will probably hold on to the price / performance crown - and they can always add an extra die and go up to 24 cores to keep bringing the fight to Intel.
See people keep saying that, but I'm not sure they quite understand how bad NUMA can be for some software. With Threadripper and EPYC, memory latency is on the order of 70-80ns for the near memory (i.e. the memory attached to that die's IMC) and 110-130ns for the far memory (i.e. the memory attached to another die on the same package). This varies depending on the clockspeed, but it's 79ns near and 136ns far using the stock DDR4-2400 speed. What this means is that two dies will behave like normal, but a third die will be forced to access the far memory at all times, increasing latency by about 75% at any memory clock. Alternatively, the package could be rewired so two dies can use one memory channel each and the third can use both, but that's still a problem because now two dies can only move half the data at low latencies.

But that's not going to happen. Thermals will be weird once you have dies arranged in an L shape on the package. Performance will get weird because an OS' scheduler needs to be aware of the third die that doesn't have direct memory access and will suffer from latency. Motherboards could be designed to support six memory channels, sure, but I'm not sure the TR4 socket is able to support that. I know it's a modified version of EPYC's SP3 and its eight memory channels, but they're only identical mechanically, not electrically.

And above all, it's just pointless. If AMD really wants to go nuts, they'll release unlocked EPYC CPUs for crazy people.
thumb.gif
 
1 - 20 of 46 Posts
Top