The Ryzen "ThreadRipper"... 16 cores of awesome

[Atari2600]

I think otherwise. Ryzen 1700 is cannibalizing AMD itself.

You have to remember that the number of overclockers in the wider market is miniscule. Absolutely miniscule.

If the overclocking of the 1700 c.f. the 1700X affected the product sales by any more than 5% for either CPU I'd be absolutely utterly astounded.


Enthusiast are <1% of the market. Its likely <10% of enthusiasts would overclock.

 

[IEC]

Still waiting for Threadripper.

Personally I think 16 core X model will end up with 1700X clocks, all core turbo and all but I'd like to be pleasantly surprised with a 2+ core XFR of 4.1GHz+, but who knows.

Keep in mind 2-core XFR turbo of 4.1GHz is possible with the 1800X that uses only 1x Zeppelin die. There is technically nothing stopping AMD from using 2x well-binned Zeppelin dies and ending up with a 16c/32t Threadripper capable of 4.1 GHz XFR... on four cores. That could actually be very interesting for lightly-threaded workloads like a lot of DX11 games.

 

[tamz_msc]

Still waiting for Threadripper.


Keep in mind 2-core XFR turbo of 4.1GHz is possible with the 1800X that uses only 1x Zeppelin die. There is technically nothing stopping AMD from using 2x well-binned Zeppelin dies and ending up with a 16c/32t Threadripper capable of 4.1 GHz XFR... on four cores. That could actually be very interesting for lightly-threaded workloads like a lot of DX11 games.

Actually the 32-core EPYC can sustain single-core Turbo on up to 12 cores, depending on the workload. Would love to see something similar in TR.

 

[IEC]

Actually the 32-core EPYC can sustain single-core Turbo on up to 12 cores, depending on the workload. Would love to see something similar in TR.

I would be pleasantly surprised if it was capable of XFR turbo of 4GHz+ on 6+ cores. That would make it very competitive with X299 low core count CPUs at stock frequencies. I'm expecting a bit less.

 

[inf64]

I would be pleasantly surprised if it was capable of XFR turbo of 4GHz+ on 6+ cores. That would make it very competitive with X299 low core count CPUs at stock frequencies. I'm expecting a bit less.

From Hardware.fr, performance at 3Ghz:

Games SKL-X is ~5% faster than Zen.

Apps, SKL-X is ~11% faster than Zen.

Zen is super-competitive IPC wise and even more so in perf./watt and perf./$. This will extend to HEDT and TR, even more so than with "regular" Ryzens. Intel's only real advantage is that they can clock(auto/TB and manually) higher and that's it.

 

[ultima_trev]

BROADWELL-E BETTER CLOCK FOR CLOCK THAN SKYLAKE-X??!!

Well, hopefully Intel can find a way to rectify this...

 

[TheGiant]

You have to remember that the number of overclockers in the wider market is miniscule. Absolutely miniscule.

If the overclocking of the 1700 c.f. the 1700X affected the product sales by any more than 5% for either CPU I'd be absolutely utterly astounded.


Enthusiast are <1% of the market. Its likely <10% of enthusiasts would overclock.

That why I am saying it. AMD has a attractive 8C product that is enough for lots of people and they wont buy the 1700X because of it. As if they don't like money
As for AMD's market research, I wouldn't bet on it. As I see it AMD is very strong and clever technical wise but commercial not really..

I am looking forward to buy 12C TR. SKL-X disappointed me a lot. Let's wait for TR. I bet it will be excellent.

 

[moinmoin]

IMO they should release the 1700X as 1700 and sell it for minimum 400USD and the actual 1700 shouldn't even exist.

Thanks to binning 1700 with its lower TDP is actually more efficient at the stock frequencies it runs at than the higher clocked 1700X and 1800X would be when downclocked. And AMD does position the 1700 as the most efficient 8c chip on the market. Compared to that 1700X and 1800X (especially the latter) are more like factory overclocked chips for users that prefer higher stock clocks instead overclocking a 1700 themselves. So rather than questioning the 1700 it's actually the 1700X that's positioned in an odd middle ground.

 

[Topweasel]

I think otherwise. Ryzen 1700 is cannibalizing AMD itself. IMO they should release the 1700X as 1700 and sell it for minimum 400USD and the actual 1700 shouldn't even exist. I bet it wouldn't hurt their sales. The one who wanted to buy ryzen would buy it even for 400 USD without problems.
ThreadRipper should be made by AMD as something really exclusive, the product itself doesn't sell. Even if it is really good.

My point is you can't cannibalize any of the AMD products the way you are thinking. Just look at the price of the EPYC 8C CPU it's more expensive than a 1700 or 1700x (you can probably get even a 1800x cheaper) what it is, is ~ the price of 4 Ryzen R5 1400's.

Outside the king of the hill 32c EPYC AMD's pricing seems simple. A fully working Zeppelin is worth X, a 6c Zeppelin is worth Y, a 4c Zeppelin and lower is worth Z. From there they then add price for higher factory clock settings. The 1700 doesn't cannibalize 1800x options because you can't assume anyone who purchased a 1700 would have paid full price for a 1800x. It is just as likely that a 1700 purchaser if given only the choice between a 1800x and 1600x would get the 1600x.

The Intel pricing model should not be our focus when looking at pricing of AMD's products. Just because Intel tends to go out of their way to screw over enthusiasts with single configuration overclockable chips and inventing a HEDT segment where they can seriously ratchet up the margins.

If you think that this is something new to AMD think again. Look at their video card lineup. They have for the longest time had 3-4 times the price tiering of Nvidia. While Nvidia had one top end, one mid high, one mid, one mid, one low, and one basic. AMD has almost allways had 2-3 using the top chip, 2-3 using the middle chip, 2-3 using the low end chip. This only changed at the 300 series as they started to get far enough behind Nvidia that they really couldn't tier them out like that due to the compressed price/perf that their top chip sat at. It didn't make sense to offer sku's priced so close together, between the 460-480 would have been like 80 dollars that isn't enough room for 9 variations.

 

[Aenra]

Until we know the final voltages, theorizing on a TR capable of attaining 1800X speeds is at best premature.
That IF isn't free; and the voltages required for each and every trip are accounted, more so on this platform, what with all the extra lanes (for which IF is again responsible of). Two 1800X separate dies, are one thing. Two 1800X linked dies, another.

And even this is flawed, as the stepping's changed (B2 now, we have yet again zero data). It would be grand if they managed it but at this point, presuming it's a possibility? Like i said, premature at best.

Edit: I do hope this happens though, i'd get me one and nevermind the price or core numbers. I'd even pay the extra cost, for cores i'd fully disable, just so i could have 1800X performance with quad band + more available lanes (which frankly, i'd probably leave alone yet again, give headroom to IF).

 

[Topweasel]

Until we know the final voltages, theorizing on a TR capable of attaining 1800X speeds is at best premature.
That IF isn't free; and the voltages required for each and every trip are accounted, more so on this platform, what with all the extra lanes (for which IF is again responsible of). Two 1800X separate dies, are one thing. Two 1800X linked dies, another.

And even this is flawed, as the stepping's changed (B2 now, we have yet again zero data). It would be grand if they managed it but at this point, presuming it's a possibility? Like i said, premature at best.

Edit: I do hope this happens though, i'd get me one and nevermind the price or core numbers. I'd even pay the extra cost, for cores i'd fully disable, just so i could have 1800X performance with quad band + more available lanes (which frankly, i'd probably leave alone yet again, give headroom to IF).

No one is stating that you can regularly hit 1800x over clocks or even be able to OC to 1800x speed. But the science on it is pretty sound. We know the over clocks are silicon limited at ~4ghz. We know what the power usage is at 4ghz (~120w). We can pretty confident that the socket is capable of 200w or more given headroom that current x99 and x299 boards support. The is little reason to think that a sustained 4GHz overclock shouldn't be possible and that people should regularly be able to hit 3.7 or 3.8GHz.

Things that can negatively impact performance. I don't see the interconnect between dies being an issue. Might not be free on a power sense, but AMD would be silly for using an interconnect that could work on the upper end their clock speed and if the refresh clocks higher it would screw them over (though the roadmap does imply that EPYC and TR may not get the refresh and will wait till Zenver2). Silicon lottery. For a given overclock you need two dies to be able to hit those numbers. For power usage I don't think it will be significant. 2 dies means higher chance one can't. This implies a larger percentage hits a wall lower than their R7 counterpart. Heat. While Ryzen runs relatively cool. OCing two dies will still require tons of cooling, can it truly clear that much cleanly.

But as I said it was simple science. We know how a Zeppelin die performs. Whether it's One, Two, or Four it should be relatively similar. This is in the same way we (and I did predict it) the SL-x power usage and limits were easy to see ahead of times. TIM or no TIM it was always going to be an issue on that. 4 cores, 8 cores 18 cores the major characteristics of SL is easy to measure. If a 7700 takes 90w to hit stock numbers, a 8 core would need ~180w, a 16 core 360w. SL-X muddies it a bit by the changes to the core config. But it still held mostly true. There is a reason people thought they could hit 4.5GHz on a 7900 and that is holding true and for those same reasons to think you can get 3.7-3.8 with some regularity on TR16.

 

[TheGiant]

My point is you can't cannibalize any of the AMD products the way you are thinking. Just look at the price of the EPYC 8C CPU it's more expensive than a 1700 or 1700x (you can probably get even a 1800x cheaper) what it is, is ~ the price of 4 Ryzen R5 1400's.

I understand that. But those are more technical arguments.

I am saying that AMD set the price standard for 8C too low. When you do that, you set the expectation for next generations. AMD will have a difficult time to increase the 8C zen2 model price even if it performs excellent.
I really want to see AMD get financially better. They are not here to please the enthusiast crowd with prices..

TR needs to be prices properly high. It gets all the properties needed- high MT performance, acceptable power, reasonable priced boards, upgradeable platform...AMD has THE product now !

 

[DrMrLordX]

4.0-4.1 GHz Threadripper may be a bit much, but it would not surprise me if overclockers got Threadripper to 3.6 GHz with relative ease.

 

[teejee]

Until we know the final voltages, theorizing on a TR capable of attaining 1800X speeds is at best premature.
That IF isn't free; and the voltages required for each and every trip are accounted, more so on this platform, what with all the extra lanes (for which IF is again responsible of). Two 1800X separate dies, are one thing. Two 1800X linked dies, another.

And even this is flawed, as the stepping's changed (B2 now, we have yet again zero data). It would be grand if they managed it but at this point, presuming it's a possibility? Like i said, premature at best.

Edit: I do hope this happens though, i'd get me one and nevermind the price or core numbers. I'd even pay the extra cost, for cores i'd fully disable, just so i could have 1800X performance with quad band + more available lanes (which frankly, i'd probably leave alone yet again, give headroom to IF).

 

[teejee]

I'm pretty sure AMD did their market research. If they thought they could get away with what you propose I'm sure they'd done it. They're not that clueless.

The way I see it the issues are first that we don't really know if there's a difference in quality between the 1700 and 1700x chips we buy. Suppose they're all tested, and the 1700 chips tend to do less well at higher frequencies as far as power consumption goes. It could be entirely reasonable for AMD to then rebrand those as lower powered lower speed 1700 chips, and keep the ones that can run a higher default frequency at lower power draw 1700x. In other words do we really know they're all "the same"? From what unscientific impression it seems that the 1700 chips aren't overclocking as high with the same power consumption as the "x" chips.

Secondly, not everyone wants to tinker with OC. So, there might be a case to make for making a tier of chips that run at different speeds by default, and by default then charging different amounts, regardless of the previous possible issue.

Lastly, you could look at it from a different perspective; instead of the 1700 being underpriced for what it is perhaps it's the "x" chips that are overpriced. Perhaps AMD did their research and found out that most people would totally consider buying a 1700 for X dollars, and that some people then would happily pay more for an "x" chip even though that was a less good deal. And in turn they might have found that people actually would not buy as many chips if the pricing was as you suggested as a whole, and that they'd lose revenue that way.

yy%j9ygdxrfxttxtggtgtägg08~t
dyt0

Sent from my LG-D855 using Tapatalk

 

[Aenra]

@Topweasel you mix Intel boards, 4core CPU wattages..? I don't even know where to start from

Since it would take me a page or two to address everything (and i really shouldn't, as a good part of your examples are entirely unrelated to this), let us just agree to disagree.

As already stated, i do share your wishes, but currently they are just that, wishes. I'd personally be much more careful about reaching conclusions. Needless to say, to each their own.

 

[Aenra]

It occurs to me i may have been misunderstood, perhaps it's my fault, should have expanded further..
Rephrased then:

I used the word 'performance' for a reason.
CCX to CCX (same die), 4 cores at each complex, at same speed, is one thing.
Chip to Chip (different die), 4 cores at each die, same speed as above, is or can be another.

Problem 1), attaining the clocks themselves.
Problem 2), attaining relative performance; hence my mentioning the Fabric in the second paragraph. They could all run at 1800X speeds, both scenarios above, but still suffer in the second scenario (MCM) if the die-to-die latency is too high.
2) i) We have no knowledge of whether this is possible yet or not, especially since it's a new stepping, no data.
2) ii) Even if it was, theoretically/in their labs, there's still a chance the cost (power) entailed is so high that AMD has had to compromise with lower performance. Again, we cannot know in advance, divination doesn't work here ^^

 

[moinmoin]

Problem 1), attaining the clocks themselves.
Problem 2), attaining relative performance; hence my mentioning the Fabric in the second paragraph. They could all run at 1800X speeds, both scenarios above, but still suffer in the second scenario (MCM) if the die-to-die latency is too high.
2) i) We have no knowledge of whether this is possible yet or not, especially since it's a new stepping, no data.
2) ii) Even if it was, theoretically/in their labs, there's still a chance the cost (power) entailed is so high that AMD has had to compromise with lower performance. Again, we cannot know in advance, divination doesn't work here ^^

Your problem 2 is mainly a software issue. The die to die communication will be presented as NUMA to the OS, so all software not already NUMA aware won't scale beyond a die anyway (Geekbench is already making quite a mess of this). The whole NUMA stuff is not new and accounted for in server/datacenter usage, but it will be quite the change and challenge on desktop.

 

[dnavas]

We know what the power usage is at 4ghz (~120w).

Umm, all core 4Ghz is a lot closer to twice that. I'm regularly seeing 160W @ 3.9, and if I run actual stress tests, 200W. I'm not even stable under stress at 4Ghz, and after about 5-6 hour encodes, 3.95G has a tendency to black-screen. If TR boards ship with crap VRM heat spreaders, it's not going to matter what the chips hit anyway.

Of the markets AMD is trying to hit, I think this one is the hardest. The platform has an enviable advantage when it comes to pcie lanes, but (likely) a serious OC clock deficit. My expectation is that the 16-core TR will be ~= 12-core Skylake-X (OCd), but that's just a guess, of course.

 

[Dresdenboy]

Your problem 2 is mainly a software issue. The die to die communication will be presented as NUMA to the OS, so all software not already NUMA aware won't scale beyond a die anyway (Geekbench is already making quite a mess of this). The whole NUMA stuff is not new and accounted for in server/datacenter usage, but it will be quite the change and challenge on desktop.

Moin moin, NUMA unawareness should only cause some inefficiencies due to an additional hop to reach a different socket's or die's physical mem range. Similar to Hypertransport.

 

[Zor Prime]

4.0-4.1 GHz Threadripper may be a bit much, but it would not surprise me if overclockers got Threadripper to 3.6 GHz with relative ease.

I would figure 3.6 base minimum they'd be able to sort out. My 1700 does 3.65 without bumping volts. CPU-Z shows 1.2v under full load (if it's being accurate).

 

[Topweasel]

I understand that. But those are more technical arguments.

I am saying that AMD set the price standard for 8C too low. When you do that, you set the expectation for next generations. AMD will have a difficult time to increase the 8C zen2 model price even if it performs excellent.
I really want to see AMD get financially better. They are not here to please the enthusiast crowd with prices..

TR needs to be prices properly high. It gets all the properties needed- high MT performance, acceptable power, reasonable priced boards, upgradeable platform...AMD has THE product now !

I still have a problem understanding what your issue is. Do you know what AMD's market penetration is? Do you know what manufacturing profit AMD is getting from these chips? Do you know what AMD needs to make to be profitable?

If AMD priced their chips in line with Intel end segmented their chips the way Intel did. No one would purchase them. It the 16c was priced at 1.5k who would purchase that over a 12c SL-X. AMD has managed to stay solvent on almost max ASP of like $150. They are going to be averaging $250-$300, for 6c+ capable dies. They have a portfolio that is going to let them sell 90%+ of their dies.

Selling those dies is much more important that some using an EPYC instead of a TR or vice versa, or buying a 1700 instead of 1800x. I'll say it again, don't assume (you know they say about that) that 1700 sale would have been a 1800 if it didn't exist. AMD has and will always target every pricing envelope and revels in offering the Bang for the buck package.

If you really think about the existence of the 1800x and it's $500 pricing actually increases the attractiveness of Ryzen and the 1700 in general. Keep the 1700 at $330 and remove the 1800x and I am willing to bet the 1700 doesn't sell as well. Offering the 1700 while the 1800x exists makes the 1700 look like a fantastic deal since you can get it and clock it up to a 1800x clock speed. Making your $300 run like a $500 chip. The 1800x made a $300 dollar chip seem like a fantastic deal and if you can't see how that helps and doesn't hurt AMD I don't know if I have anything else to say.

 

[scannall]

My own wild guess is 3.3 Ghz base. 3.6 Ghz turbo. A reasonable guess at overclocking would be 3.7 all core on air. A 'Golden Chip' on water and solid VRM's hitting 4? Maybe, but don't count on it. And yes, I'll buy one when they are released.

 

[Topweasel]

Umm, all core 4Ghz is a lot closer to twice that. I'm regularly seeing 160W @ 3.9, and if I run actual stress tests, 200W. I'm not even stable under stress at 4Ghz, and after about 5-6 hour encodes, 3.95G has a tendency to black-screen. If TR boards ship with crap VRM heat spreaders, it's not going to matter what the chips hit anyway.

Of the markets AMD is trying to hit, I think this one is the hardest. The platform has an enviable advantage when it comes to pcie lanes, but (likely) a serious OC clock deficit. My expectation is that the 16-core TR will be ~= 12-core Skylake-X (OCd), but that's just a guess, of course.

Is that from the wall or CPU because I was reading 4.1 at 1.35 was closer to 140w CPU only. The upper end if the overclock is going to be a little less predictable because of the heavy difference in voltage needed to sustain the speed. Binning plays a heavy part in it as well a 1700 die could hit 3.9 but it's probably going to need more voltage and cooling than a die used for the 1800x.

I guess what my point was though I might have gone a bit high for the clocks is the 3.7-3.9 overclock should be a regular thing with the higher end 16c, if you get lucky with the die paring could still hit the higher marks of the 1800x. The boards shouldn't be a limit because even with the increased power usage of the dies at the speed and of the two dies as a whole is still more efficient that SL-X. Considering the similarity of the boards between the platforms I have confidence that the VRMs will hold up. I also believe the interconnects on the MCM will not impact the clocking chances measurably.

It's going to be glorious. Because for once AMD will have the clock advantage. Those MCC SL-X are going to be lucky if they could hit 4GHz at 400w.

 

[stockolicious]

Until we know the final voltages, theorizing on a TR capable of attaining 1800X speeds is at best premature.
That IF isn't free; and the voltages required for each and every trip are accounted, more so on this platform, what with all the extra lanes (for which IF is again responsible of). Two 1800X separate dies, are one thing. Two 1800X linked dies, another.

And even this is flawed, as the stepping's changed (B2 now, we have yet again zero data). It would be grand if they managed it but at this point, presuming it's a possibility? Like i said, premature at best.

Edit: I do hope this happens though, i'd get me one and nevermind the price or core numbers. I'd even pay the extra cost, for cores i'd fully disable, just so i could have 1800X performance with quad band + more available lanes (which frankly, i'd probably leave alone yet again, give headroom to IF).

AMD ran a demo of TR and the scaling looked impressive - i have not seen many benches of EPYC so i'm hoping when TR comes out allot of folks post the result fast.
I am not one for predictions as i thought the Mets would be good this year but i did watch the EPYC launch and there were allot of "major" players in the server/cloud markets. The important part wasn't that they were in attendance but they were presenting "Products" These big players would not be dropping their draws on products if the scaling wasn't great. Those big customers do tons of validation work way ahead of time so having products launching soon really gives an indication that TR is going to be pretty hot. IMHO.