IBM unveils their Power9 chip

[Mondozei]

Read it here.

What do you guys think?

 

[HiroThreading]

Simply amazing...

I sometimes wish we saw the POWER architecture scale down to the consumer level (yes I know that we would also need to get around the x86 domination issue when it comes to software).

I'm very surprised to see PCIe 4.0 on the POWER9 chips though. I thought the specs had yet to be finalised.


Sent from my iPhone using Tapatalk

 

[SAAA]

Zen: Daddy!

Being serious, the engineers working on this are amazing. How in the world do you get close to twice the performance than previous gen. at the same clock when you are at such a high level already?

The new modular design that allows execution resources to be split among one or two cores is interesting too, sounds a lot like reverse multi-threading if they could do that on the fly one day rather than offer separated processors. For now I'm sure it's just the same die with fused off parts that can't be enabled but the concept is there, maybe with Power 10 they'll introduce it.

 

[XavierMace]

Simply amazing...

I sometimes wish we saw the POWER architecture scale down to the consumer level (yes I know that we would also need to get around the x86 domination issue when it comes to software).

I'm very surprised to see PCIe 4.0 on the POWER9 chips though. I thought the specs had yet to be finalised.


Sent from my iPhone using Tapatalk

It was at the consumer level for many, many years in a computer called a Macintosh.

 

[krumme]

Intel must make sure not to be stuck between a high volume low cost zen server product and this traditional brute dude. As they have a lot to lose.

Today the compettition is shaping up at serverside and there is real choice. Tell what you need and there is a optimal product for it. From fat power to slim arm.

How does the future need shape up? What is same what is different?

 

[Lepton87]

Being serious, the engineers working on this are amazing. How in the world do you get close to twice the performance than previous gen. at the same clock when you are at such a high level already?

As impressive as it is they made the comparison to a previously under-performing SMT8 mode that's why they got such a huge speed-up. Even so I'm still mighty impressed. It's going to trash Intel Xeons performance-wise. Even power8 chips are still very competitive with Xeons with twice the number of cores that were released two years later that's because power8 chips were made for very high dual and quad threaded performance(2nd thread brings 50% more performance and 4 threads about 100% HT gain is around 20% in Xeons) but 8 threaded mode seemed to be not quite done.
It seems now even SMT8 gets enough resources.

The SMT-8 mode can sometimes be a step too far for some applications, as 4 threads are now dividing up the resources of each issue queue. There are more signs that SMT-8 is rather cramped: instruction prefetching is disabled in SMT-8 modus for bandwidth reasons. So we suspect that SMT-8 is only good for very low IPC, "throughput is everything" server applications. In most applications, SMT-8 might increase the latency of individual threads, while offering only a small increase in throughput performance. But the flexibility is enormous: the POWER8 can work with two heavy threads but can also transform itself into a lightweight thread machine gun.

 

[Abwx]

https://www.computerbase.de/2016-08/ibm-power9-prozessor/

The chip is manufactured at GF....

 

[MongGrel]

Looks IBM is looking to give INTEL a fight at enterprise level again.

Be interesting to see where it goes.

 

[mnewsham]

https://www.computerbase.de/2016-08/ibm-power9-prozessor/

The chip is manufactured at GF....

The simple fact it's 14nm basically confirms that anyway.

 

[frozentundra123456]

Intel must make sure not to be stuck between a high volume low cost zen server product and this traditional brute dude. As they have a lot to lose.

Today the compettition is shaping up at serverside and there is real choice. Tell what you need and there is a optimal product for it. From fat power to slim arm.

How does the future need shape up? What is same what is different?

Right now Zen is 0 volume and totally unknown in cost, performance, and power usage. And considering the availability issues with Polaris, even if it is a competitive product, I dont think they have the resources to ramp up high volume. Seems to me Intel has a lot more to worry about vs IBM than Zen.

 

[HiroThreading]

It was at the consumer level for many, many years in a computer called a Macintosh.

True, but that was quite a while ago. POWER chips have changed a lot since then (heck, x86 chips have changed a lot since then).

I mean, it would be nice to see a derivative of the POWER7/8/9 in the HEDT or 1S/2S market as an alternative to x86.

 

[itsmydamnation]

Right now Zen is 0 volume and totally unknown in cost, performance, and power usage. And considering the availability issues with Polaris, even if it is a competitive product, I dont think they have the resources to ramp up high volume. Seems to me Intel has a lot more to worry about vs IBM than Zen.

So that's different to power 9 how...........
IBM's far bigger issue is software, not performance. The lack of enterprise software on Power is the first and biggest hurdle, AMD has no such problems........

 

[SAAA]

As impressive as it is they made the comparison to a previously under-performing SMT8 mode that's why they got such a huge speed-up. Even so I'm still mighty impressed. It's going to trash Intel Xeons performance-wise. Even power8 chips are still very competitive with Xeons with twice the number of cores that were released two years later that's because power8 chips were made for very high dual and quad threaded performance(2nd thread brings 50% more performance and 4 threads about 100% HT gain is around 20% in Xeons) but 8 threaded mode seemed to be not quite done.
It seems now even SMT8 gets enough resources.

Yeah 24-core Broadwell is going to be severely outperformed if this monster delivers.
Anyway it will be more interesting to see how Skylake-E compares, there were no news at HotChips but if Intel finally decides to diverge server core from the mobile variant there's a chance of competition. You won't see Power 9 at 4.5W TDP and that's hurting Skylake performance a bunch: imagine what they could achieve with twice the caches, eDRAM/HMC on every processor, 4-way HT and tweaks that don't aim at 2:1 performance/power increase every time. Server workloads are too different from single users needs today, there will be a split sooner or later.

 

[NTMBK]

So if a single thread can't possible use all of the resources of the SMT8 core... in what way is it still a core? Is this not just a big licensing dodge, so instead of paying Oracle licensing for 32 SMT4 cores you pay for 16 SMT8 "cores"? Is this not like if AMD had called a Bulldozer module a core?

 

[Phynaz]

So that's different to power 9 how...........
IBM's far bigger issue is software, not performance. The lack of enterprise software on Power is the first and biggest hurdle, AMD has no such problems........

What lack of enterprise software?

 

[The Stilt]

Would be nice to see what would happen to Polaris and Zeppelin, if they were made on the same IBM developed 14nm HP process (now GlobalFoundries produced) instead of the 14nm LPP

 

[Glo.]

Would be nice to see what would happen to Polaris and Zeppelin, if they were made on the same IBM developed 14nm HP process (now GlobalFoundries produced) instead of the 14nm LPP

According to BitsAndChips all of them are produced in the same GloFo fab nr 8.

 

[itsmydamnation]

What lack of enterprise software?

Just so you know i spent the last year working in an environment with more redhat on power and aix on power then x86. They have power8, as well as z13. The power decision was purely political the actual enterprise architects wanted to move it all to redhat on x86.

IBM does a good job of getting open *ix stuff cross compiled and it mostly works. But anything Microsoft (obviously) , any cloud services automation (azure stack, Vrealize, etc), there are lots of issues with middle-ware even for stuff that is well supported on power like Hana. Your standard document archive solutions (objective, trim, etc), unified communications,etc the list just goes on and on, if it wasn't midnight i might indulge you more.

Honestly in an average enterprise workplace how much as a percentage of x86 do you think you could drop in replace with power? Now how much of a standard enterprise could you drop in replace with Zen?

 

[el etro]

This Power to Power comparison likely compares SMT8 Power8 versus SMT8 Power9, accounting huge SMT yield increase. Not that is not incredible to increase performance by this margin in just two years.

 

[el etro]

Would be nice to see what would happen to Polaris and Zeppelin, if they were made on the same IBM developed 14nm HP process (now GlobalFoundries produced) instead of the 14nm LPP

Hey the Stlit, do you know if AMD has any access to the 14HP process?

 

[The Stilt]

Hey the Stlit, do you know if AMD has any access to the 14HP process?

No idea.
Wouldn't matter much for existing designs anyway, since they are all designed for 14nm LPP

 

[Abwx]

Since 14nm HP was part of the IBM/GF deal it s likely that IBM agreement is necessary if they want to use it for another customer, and i m not sure that IBM would agree to give some help to the concurrent ISA, be it AMD..

 

[SarahKerrigan]

So if a single thread can't possible use all of the resources of the SMT8 core... in what way is it still a core? Is this not just a big licensing dodge, so instead of paying Oracle licensing for 32 SMT4 cores you pay for 16 SMT8 "cores"? Is this not like if AMD had called a Bulldozer module a core?

SPARC T2/T3 used a microarchitecture with hard-partitioned thread groups too, and nobody insisted that it had to be licensed as two cores. Like T2 and T3, P9-SMT8 shares an L1 cache between both thread groups, as well as certain functional units (crypto, and iirc the DFU.)

Of course it's obviously designed to maximize compute per core license, but whether that makes it less of a core is debatable.

 

[HiroThreading]

Yeah 24-core Broadwell is going to be severely outperformed if this monster delivers.
Anyway it will be more interesting to see how Skylake-E compares, there were no news at HotChips but if Intel finally decides to diverge server core from the mobile variant there's a chance of competition. You won't see Power 9 at 4.5W TDP and that's hurting Skylake performance a bunch: imagine what they could achieve with twice the caches, eDRAM/HMC on every processor, 4-way HT and tweaks that don't aim at 2:1 performance/power increase every time. Server workloads are too different from single users needs today, there will be a split sooner or later.

I've been thinking about this for a while too. It's pretty obvious that the one-design-fits-all approach to x86 architecture design is reaching its limits for Intel. Performance targets for the server market are hampering some of the goals of mobile, and the power limits of mobile are constraining high end server performance. Sooner or later, Intel will have to split the design of mobile/desktop and servers in two.

I suppose this was one of the reasons behind the Itanium -- just that it was perhaps ahead of its time.

But yeah, a design optimised for high performance (ultra wide, large amounts of cache) at TDPs north of 150-200w would be very interesting. With a wide enough execution engine, Intel could probably do SMT4 or SMT8?

Just so you know i spent the last year working in an environment with more redhat on power and aix on power then x86. They have power8, as well as z13. The power decision was purely political the actual enterprise architects wanted to move it all to redhat on x86.

IBM does a good job of getting open *ix stuff cross compiled and it mostly works. But anything Microsoft (obviously) , any cloud services automation (azure stack, Vrealize, etc), there are lots of issues with middle-ware even for stuff that is well supported on power like Hana. Your standard document archive solutions (objective, trim, etc), unified communications,etc the list just goes on and on, if it wasn't midnight i might indulge you more.

Honestly in an average enterprise workplace how much as a percentage of x86 do you think you could drop in replace with power? Now how much of a standard enterprise could you drop in replace with Zen?

But isn't most of what you mentioned due to the vastly different markets that POWER and x86 target?

 

[stuff_me_good]

No idea.
Wouldn't matter much for existing designs anyway, since they are all designed for 14nm LPP

I feel you bro. I just wonder how AMD engineers must feel when they can see first hand how their great design works now, while being able to see also how much better it could do with proper process, but not being able to anything to bring that great design with proper process to the market?