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

[raghu78]

Nice to wake up today to this kind of news Was a little worried they might not be able to clock these chips high enough but 4ghz on a 16C is awesome!

4 Ghz is going to be for 1-2 threads. All core turbo will most probably be at 3.5 Ghz for 1950x and 3.6 Ghz for 1920x. Still nice to see base clocks at 3.4/3.5 Ghz.

 

[ddogg]

4 Ghz is going to be for 1-2 threads. All core turbo will most probably be at 3.5 Ghz for 1950x and 3.6 Ghz for 1920x. Still nice to see base clocks at 3.4/3.5 Ghz.

That's still pretty impressive especially if it can scale equally well in other applications. Not to mention it will not have any of the x299 issues such as thermals and power consumption.

 

[IEC]

4 Ghz is going to be for 1-2 threads. All core turbo will most probably be at 3.5 Ghz for 1950x and 3.6 Ghz for 1920x. Still nice to see base clocks at 3.4/3.5 Ghz.

1) Turbo on the 32c/64t Epyc chips is up to 12 threads(?) for the maximum turbo frequency
2) XFR on "X" models is +100 over default all-core turbo

This means, at a minimum, the effective all-core clock for 1950X is likely to be 3.6GHz.
This likely also means the effective max turbo clock of 4GHz will be achievable on 4+ threads. Maybe even 6+.

 

[Topweasel]

I am just going Toot my own horn here and say I was right about the clocks. No way AMD clocks these within a 100MHz limit of the dies. These are cherry picked and I feel like 4GHz all core is a reasonable goal.

 

[dnavas]

The scaling in Cinebench R15 from R7 1700X to TR 1950X is almost 100%. Amazing.

First, I'm *really* happy to see a top-end turbo of 4, and I'm hoping to see reviews with OCs of all cores there-ish just so I know what I would need to buy in terms of case and cooler.
That said, the scaling from 1920 to 1950 is not linear -- 25% faster with 33% more cores is a little weird. Cache pressure? Fabric? OS scheduling? [I haven't watched the video, but these numbers are both on the same OS, right?]

 

[USER8000]

Impressive clockspeeds for both parts!

 

[EXCellR8]

1950 is literally twice as good as my 1700... but also twice the processor. If Vega sucks I might need to scrounge up 1000 bucks for one of those suckers

 

[moinmoin]

That said, the scaling from 1920 to 1950 is not linear -- 25% faster with 33% more cores is a little weird. Cache pressure? Fabric? OS scheduling?

TDP is the same for both so of course 1950X is lower clocked as a result.

 

[Tup3x]

I can only imagine what's possible with Ryzen 2 and 7nm... That should be epic.

 

[CatMerc]

I am just going Toot my own horn here and say I was right about the clocks. No way AMD clocks these within a 100MHz limit of the dies. These are cherry picked and I feel like 4GHz all core is a reasonable goal.

I was out of the loop about discussions here. Was anyone expecting more than 4GHz for Threadripper?

I can only imagine what's possible with Ryzen 2 and 7nm... That should be epic.

And that's how EPYC was born.

 

[Topweasel]

I was out of the loop about discussions here. Was anyone expecting more than 4GHz for Threadripper?

No I stated a 4-4.1 Limit on TR and it became I was crazy. That realistically we were going to be limited to 3.7-.3.8. But the Cinebench numbers for the 1950x actually points to an all core turbo of 3.7. So not only is a 4GHz overclock look possible, their choice on dies actually makes it look reasonable.

 

[CatMerc]

No I stated a 4-4.1 Limit on TR and it became I was crazy. That realistically we were going to be limited to 3.7-.3.8. But the Cinebench numbers for the 1950x actually points to an all core turbo of 3.7. So not only is a 4GHz overclock look possible, their choice on dies actually makes it look reasonable.

I don't see why it would be unreasonable to have 4GHz on Threadripper. Heat is spread apart by virtue of having two dies, and the heatspreader is absolutely massive due to coming from EPYC. Heat would not have been an issue, and with some binning, neither would power.

 

[Topweasel]

I don't see why it would be unreasonable to have 4GHz on Threadripper. Heat is spread apart by virtue of having two dies, and the heatspreader is absolutely massive due to coming from EPYC. Heat would not have been an issue, and with some binning, neither would power.

I think part of it was power but considering these boards will be on par if not slightly better than their X299 counterparts, we know 300w should be within reason on the boards.The other issue being spreading the opportunities out on two dies. Which again to me meant sure less of them would get that high, but not unreasonable. Then was the interconnect. I really don't know what that guy was going on about because I just couldn't think of a reason why that might limit overclocks. So the general consensus was 3.7 or so. Which to me seemed way too low (and obviously is if they are fitting a 3.7GHz all core turbo into that power envelope).

 

[raghu78]

I can only imagine what's possible with Ryzen 2 and 7nm... That should be epic.

Ryzen on 14nm+ in 2018 should bring higher clocks. 14nm+ is likely to bring around 15% higher transistor performance than 14LPP. I think 4.1 - 4.2 Ghz base on 8 cores is likely. Max clocks could hit 4.6 - 4.7 Ghz. Zen 2 on 7nm is when we can expect major gains in IPC from micro-architectural improvements. Some of the weaknesses in current Zen core will be addressed such as cache latency and memory latency. GF 7LP brings a 55% area reduction and > 40% performance increase compared to GF 14LPP . The GF 7LP process is designed for 5 Ghz operation. Zen 2 might be the first AMD CPU to hit 5 Ghz turbo clocks at stock.

 

[SlowSpyder]

Zen 2 might be the first AMD CPU to hit 5 Ghz turbo clocks at stock.

Don't forget the FX9590.

 

[nathanddrews]

I think the Threadripper lid is going to be more than 2X the surface area of 2 Ryzen lids, so I would expect that extra surface area to work to its advantage, especially if TR can achieve the same clocks at lower voltages than Ryzen (assumption).

 

[maddie]

Ryzen on 14nm+ in 2018 should bring higher clocks. 14nm+ is likely to bring around 15% higher transistor performance than 14LPP. I think 4.1 - 4.2 Ghz base on 8 cores is likely. Max clocks could hit 4.6 - 4.7 Ghz. Zen 2 on 7nm is when we can expect major gains in IPC from micro-architectural improvements. Some of the weaknesses in current Zen core will be addressed such as cache latency and memory latency. GF 7LP brings a 55% area density increase and > 40% higher performance than 14LPP . The GF 7LP process is designed for 5 Ghz operation. Zen 2 might be the first AMD CPU to hit 5 Ghz turbo clocks at stock.

You seem to be quoting incorrect figures for the 7LP process, or at least writing it in a confusing manner. If you say an area reduction of 55%, you'll be stating it in a better manner although still not fully accurate.

Semiwiki: https://www.semiwiki.com/forum/content/6837-globalfoundries-7nm-euv-update.html

The increase in transistor density is at least 100% for the 14LPP > 7LP transition. Thus > 2X transistors in the same area between the two nodes.

 

[dnavas]

I don't see why it would be unreasonable to have 4GHz on Threadripper. Heat is spread apart by virtue of having two dies, and the heatspreader is absolutely massive due to coming from EPYC. Heat would not have been an issue, and with some binning, neither would power.

I wasn't expecting it, because I don't see my 1800X being stable at 4, the tdp is 95W, the mcm won't be 0w overhead, and the socket was supposed to be 180W (all of which pointed to a lower ceiling). Add to that the motherboard VRM debacle on X299, and I really didn't have high expectations.
That said, I was talking all-core OC, and we still don't really know about that :shrug:

 

[Topweasel]

I wasn't expecting it, because I don't see my 1800X being stable at 4, the tdp is 95W, the mcm won't be 0w overhead, and the socket was supposed to be 180W (all of which pointed to a lower ceiling). Add to that the motherboard VRM debacle on X299, and I really didn't have high expectations.
That said, I was talking all-core OC, and we still don't really know about that :shrug:

But the VRM issue is happening at nearly 400w. It's why on the reverse end I warned people to not get to hopeful for large clocks on the x299.


As for all core clocks. The Cinebench numbers suggest a 3.7 GHz all core turbo. Which implies a lot of headroom for getting it up to 4.

 

[raghu78]

You seem to be quoting incorrect figures for the 7LP process, or at least writing it in a confusing manner. If you say an area reduction of 55%, you'll be stating it in a better manner although still not fully accurate.

Semiwiki: https://www.semiwiki.com/forum/content/6837-globalfoundries-7nm-euv-update.html

The increase in transistor density is at least 100% for the 14LPP > 7LP transition. Thus > 2X transistors in the same area between the two nodes.

I meant 55% area reduction (edited my post) .

https://www.semiwiki.com/forum/cont...alfoundries-discloses-7nm-process-detail.html

GF 14LPP
CPP - 78nm MMP - 64nm . Cell Area = CPP X MMP = 78 x 64 = 4992

GF 7LP
CPP - 56nm MMP - 40nm Cell Area = CPP X MMP = 56 x 40 = 2240

7LP vs 14LPP = 2240/4992 * 100 = 44.8% area (55% area reduction)

 

[french toast]

Am I correct in thinking threadripper will be able to be better cooled due to the larger surface area of the die?

 

[scannall]

Am I correct in thinking threadripper will be able to be better cooled due to the larger surface area of the die?

That, and being soldered instead of glued.

 

[Topweasel]

Am I correct in thinking threadripper will be able to be better cooled due to the larger surface area of the die?

Not just surface area of the Die but by by having two dies with spacing, heat generated by them has an opportunity of being absorbed by the cooler before it contributes to the rest of the die (or other die in this case). Think reverse reheating lasagna in a microwave, splitting it up helps. The Heat spreader is also huge which should help.

 

[Aenra]

glued

Got a whooole new meaning since yesterday, hasn't it ^^
Will never say "glue" in just the same way again..

On topic, am glad to see these clocks. Now on to my 'Problem n2' (cautiously optimistic if i'm to be honest, but still).

 

[StinkyPinky]

How many cores is in use when it turbos to 4GHz?

Could the 1920x in theory overclock better because some cores are disabled? (i presume it's just a 1950x with some cores disabled...)