Intel Could Upstage EPYC "Rome" Launch with "Cascade Lake" Before Year-end

[HTC]

4 of these on a board ? May God have mercy on whoever has the job of designing cooling solutions.

Won't a wraithripper be enough?

Wait: the base may actually be too small ...

 

[hat]

This raises an interesting issue. Even if this 48 core CPU drew 400w, that's still less than 9 watts per core. Rome may be even more efficient. Are we getting stuck on this high TDP figure and stopping short of the benefits it comes with? Fourty eight cores on a single chip! And Rome is looking even better! What kind of power draw would we be looking at for the same number of slower cores 5 years ago? 10 years ago?

Won't a wraithripper be enough?

Wait: the base may actually be too small ...

These are server chips, which are often housed in cases that have hurricane force winds blowing through them thanks to multiple Delta fans. The cooler itself, if not liquid like @FordGT90Concept is suggesting, likely wouldn't even have a fan on it, because the airflow going through the case would already be enough. Nobody had a problem with the those "passive" AMD cards (Vega 20, I believe?) we had an article on a while back because of this, but everyone's crapping their pants over a high power draw CPU? These ain't going in your grandma's Dell, and neither is Rome.

 

[efikkan]

Do you honestly expect Intel to launch a 48 core CPU with 410W TDP? Even disabling HT, TDP will still be around 350W, or so we're guessing.

The clock speeds would have to be significantly lowered to enable a "more friendly" TDP but that would be a problem VS Rome 64c / 128t so i seriously doubt Intel would do that.

I guess it depends on the clock speeds, e.g. 8176M at 2.1-3.8 GHz consumes 165W.
Considering these dual Cascade Lake-SP dies are on 14nm++, a total TDP of ~300W could be feasible, which is not a problem to cool.

As I said before, mainframes should move towards liquid cooling because liquid is easier to manage in clusters. We don't have TDP figures for Rome either--it might be liquid cooled as well.

Do you mean like racks hooked up to a loop for the whole rack, or many small AiOs?

The answer why liquid cooling is rarely used this way is reliability, serviceability and damages when accidents occur.

4 of these on a board ? May God have mercy on whoever has the job of designing cooling solutions.

Racks with 8+ Titan GPUs are cooled just fine, so I don't see the problem.

 

[ShurikN]

4 of these on a board ? May God have mercy on whoever has the job of designing cooling solutions.

And power delivery

 

[FordGT90Concept]

Do you mean like racks hooked up to a loop for the whole rack, or many small AiOs?

By the rack. Servers slot into the liquid system.

The answer why liquid cooling is rarely used this way is reliability, serviceability and damages when accidents occur.

All of those problems can be addressed by design and they'll save loads of money in cooling. Come to think of it, it could easily be a phase-change refrigerant system so not only does it make CPUs cold, it can cool the environment/circuitry too. Almost all of the heat is moved outside of the facility. Humidity becomes more of a problem than heat.

 

[R0H1T]

8180? 205w each.


We're talking about a $20,000 processor here. The only Xeons without Hyperthreading are <$500. If they couldn't offer a 48c processor without Hyperthreading, they simply wouldn't offer the processor at all and instead, choose a smaller package to MCM (e.g. 16c/32t 8153).

The only 24c/48t packages for candidates in the 48c/96t processor are golden samples. They've likely been binning them for over a year now.

These chips are clearly a niche of a niche, the best selling chips would still probably be the 28 core platinum. This thing might have worse perf/W (not sure) whilst having less throughput per core, if HT is disabled. Even if we leave the HT part out, the regular Xeon will probably give more perf/W & offer arguably a simpler system to deal with. The chip would also need a new socket &/or be on BGA.

The chips likely doesn't go beyond 2S anyway, looking at the slides it is possible that they're disabling HT since thread count is missing ~

This raises an interesting issue. Even if this 48 core CPU drew 400w, that's still less than 9 watts per core. Rome may be even more efficient. Are we getting stuck on this high TDP figure and stopping short of the benefits it comes with? Fourty eight cores on a single chip! And Rome is looking even better! What kind of power draw would we be looking at for the same number of slower cores 5 years ago? 10 years ago?



These are server chips, which are often housed in cases that have hurricane force winds blowing through them thanks to multiple Delta fans. The cooler itself, if not liquid like @FordGT90Concept is suggesting, likely wouldn't even have a fan on it, because the airflow going through the case would already be enough. Nobody had a problem with the those "passive" AMD cards (Vega 20, I believe?) we had an article on a while back because of this, but everyone's crapping their pants over a high power draw CPU? These ain't going in your grandma's Dell, and neither is Rome.

Rome is more efficient, undoubtedly. The TDP value is important in this case because big data/cloud spend more money on power/cooling than the hardware itself.

I guess it depends on the clock speeds, e.g. 8176M at 2.1-3.8 GHz consumes 165W.
Considering these dual Cascade Lake-SP dies are on 14nm++, a total TDP of ~300W could be feasible, which is not a problem to cool.


Do you mean like racks hooked up to a loop for the whole rack, or many small AiOs?

The answer why liquid cooling is rarely used this way is reliability, serviceability and damages when accidents occur.


Racks with 8+ Titan GPUs are cooled just fine, so I don't see the problem.

They are also bigger chips, than Xeon XCC dies, literally massive. The heat density is also insane for Intel Xeons, remember the OCed 18 core sipping more than 1000W
https://www.techpowerup.com/237313/...pu-6-1-ghz-on-all-cores-consumes-up-to-1000-w
The point being smaller chips are generally hard(er) to cool, especially monolithic dies.

 

[FordGT90Concept]

The chips probably don't go beyond 2S anyway...

It's going to be a "Xeon Platinum" which is 8-way.

 

[R0H1T]

It's going to be a "Xeon Platinum" which is 8-way.

That's regular Cascade Lake SP, AP are "special" parts.
https://www.anandtech.com/show/13535/intel-goes-for-48cores-cascade-ap

 

[efikkan]

By the rack. Servers slot into the liquid system.

All of those problems can be addressed by design and they'll save loads of money in cooling. Come to think of it, it could easily be a phase-change refrigerant system so not only does it make CPUs cold, it can cool the environment/circuitry too. Almost all of the heat is moved outside of the facility. Humidity becomes more of a problem than heat.

Your proposal gives way too many points of failure, and spills will cause terrible damage.

Many datacenters have this already solved by watercooling the air going through the rack, so kind of the opposite way people watercool their computers, so there is one huge closed cooling system and water never gets close to the electronics.

But let's not spin too far off topic.

 

[FordGT90Concept]

That's regular Cascade Lake SP, AP are "special" parts.
https://en.wikichip.org/wiki/intel/microarchitectures/cascade_lake

I think the highest TDP they'll have is 240w and manufacturers like Cray could pack 8 on to a single PCB.

 

[R0H1T]

I think the highest TDP they'll have is 240w and manufacturers like Cray could pack 8 on to a single PCB.

Possibly, Intel does make custom chips for large clients. With 14nm++ they can probably squeeze another 100MHz across 28 cores with same TDP, or get even higher clocked cores & slightly worse TDP.

Intel atm still holds all the aces in this sector, Rome can't come soon enough for AMD. Even with Rome & then Milan it'll take years before AMD is ubiquitous in this space.

 

[Deleted member 158293]

This is why AMD is emphasizing TCO (Total Cost of Ownership), not just the cost of the CPU itself.

 

[Valantar]

But Intel HAS to show something and, apparently, Intel's CPU is "made" to work with a 4 socket system, so it will have 192 cores VS AMD 128c / 256t in a dual socket system: depending on the speeds involved, it's possible Intel's CPUs could "outrun" AMD's in many workloads, so it can still be competitive, but it will be so @ double the sockets.

I think the highest TDP they'll have is 240w and manufacturers like Cray could pack 8 on to a single PCB.

These CPUs reportedly won't go beyond 2S, as hardware-wise they're nothing more than two CPUs in a single package (meaning a 2S system with these is 4S in practice, and a single one is a 2S system glued together). 8 of these on a board? Even with 1DPC, that's 96 memory slots. That and the CPUs/sockets would fill whatever motherboard you managed to throw together regardless of form factor, negating the possibility for any PCIe or power delivery. Yeah, that's not happening.

 

[hat]

Your proposal gives way too many points of failure, and spills will cause terrible damage.

Many datacenters have this already solved by watercooling the air going through the rack, so kind of the opposite way people watercool their computers, so there is one huge closed cooling system and water never gets close to the electronics.

But let's not spin too far off topic.

Water is dangerous, especially on that scale. There's a rather intricate water cooling system where I work. Granted, that's for heavy machinery and not servers, but still... lots of points of failure. Wanna guess how many huge leaks I've seen?

 

[Prima.Vera]

Too late intel too late!
AMD 7nm
AMD Intel

Well, to be fair, especially if you chech the details in deep, Intel's 10nm process is basically the same as Samsung's and TSMC's 7nm process. The Asian companies just play with the lower nrs for marketing purposes, but the specialists know better

 

[hat]

You're forgetting one major difference. Samsung's and TSMC's 7nm works. Intel's 10nm...

 

[FordGT90Concept]

Rome is able to achieve higher density and lower power consumption because it's on TSMC 7nm. Cascade Lake fails to impress simply because it's made on Intel's archaic (relatively speaking) 14nm.

 

[DeathtoGnomes]

AMD 64-core. kabashes Intel.