[Rumor] Zen Can OC To 5GHz on Air

[patrickjp93]

1 minute ago, Lord Nicoll said:

That's because it's GPU accelerated, which I don't think will be running off RISC machine due to PCI-e and compatibility, I'm guessing it will be using Xeons for that reason. 

No, it uses Power 9 by IBM. Read please.

[Lord Nicoll]

2 minutes ago, patrickjp93 said:

No, it uses Power 9 by IBM. Read please.

Hmm, I'd be interested to know how many CPU cores against GPU computes units are being used, it seems the CPUs are only there to arrange info for the GPU, again a Xeon or other CISC machine would destroy with similar specs, and for powerdraw, I don't think these people care, when you're putting that much into a PC, it's gonna eat a hell of a lot power no matter what. 

 

[patrickjp93]

3 minutes ago, Lord Nicoll said:

Hmm, I'd be interested to know how many CPU cores against GPU computes units are being used, it seems the CPUs are only there to arrange info for the GPU, again a Xeon or other CISC machine would destroy with similar specs, and for powerdraw, I don't think these people care, when you're putting that much into a PC, it's gonna eat a hell of a lot power no matter what. 

 

In general IBM can stack up to 32CPUs into a coherent node, while Intel only can stack 8. Integrators like Cray can arrange them into 16 for a heavy price.

 

Also, considering with 512-bit SIMD each Power 8 flagship CPU can deliver 1.2TFlops in its own right, you're just wrong!

 

And no, did you not read the benchmarks by Anandtech? Power 8's 2nd best is already 61% faster than the 2699V4 at best and only 7% behind at worst with a geometric mean being 41% better. The 8890 has a lower clock speed than the 2699.

[Prysin]

20 minutes ago, patrickjp93 said:

No, my announcement is this:

 

Bow before your new moderator (as of tomorrow).

well if you become a mod, that isnt going to last very long

[Macimoar]

Honestly, I don't care if the 5GH on air is true or not. I just really appreciated the concept of putting it in binary and then not mentioning it at all in the article. I'm not sure if it was actually the author who put it there or something an engineer told him to put in it; but I don't care, the world needs more cool stuff like this.

[Lord Nicoll]

1 minute ago, patrickjp93 said:

In general IBM can stack up to 32CPUs into a coherent node, while Intel only can stack 8. Integrators like Cray can arrange them into 16 for a heavy price.

 

Also, considering with 512-bit SIMD each Power 8 flagship CPU can deliver 1.2TFlops in its own right, you're just wrong!

 

And no, did you not read the benchmarks by Anandtech? Power 8's 2nd best is already 61% faster than the 2699V4 at best and only 7% behind at worst with a geometric mean being 41% better. The 8890 has a lower clock speed than the 2699.

Those are all work loads that benefit and are written for RISC machines, that's like saying the RX 480 is better than the GTX 1060 because it beat it in Ashes of the Singularity, that are countless other benchmarks that show Power8 falling behind Broadwell EP (not Haswell EP). For the vast majority of workloads power8 falls behind, for workloads that are very common Intel, or possible even the new Zen server systems.
Intel can fit 8 24 core/48 thread CPUs into a single system. Yes the E7-8890 V4 is a little slower, because it has more cores and the architecture is better, you quoted IPC earlier but apparently have completely ignored the E5-2699 v3 is last gen, only has 18 cores and only scales to 2 sockets. The E7 and E5 is much like I5 and I7, the E7 is a higher end much more powerful chip. You don't know a lot about this stuff, that much is clear. broadwell also has AVX2, which is 256 bit, so it isn't that far behind, the new soon to be released Skylake EP (the server versions) will havd 512 Bit support. 

[patrickjp93]

8 minutes ago, Lord Nicoll said:

Those are all work loads that benefit and are written for RISC machines, that's like saying the RX 480 is better than the GTX 1060 because it beat it in Ashes of the Singularity, that are countless other benchmarks that show Power8 falling behind Broadwell EP (not Haswell EP). For the vast majority of workloads power8 falls behind, for workloads that are very common Intel, or possible even the new Zen server systems.
Intel can fit 8 24 core/48 thread CPUs into a single system. Yes the E7-8890 V4 is a little slower, because it has more cores and the architecture is better, you quoted IPC earlier but apparently have completely ignored the E5-2699 v3 is last gen, only has 18 cores and only scales to 2 sockets. The E7 and E5 is much like I5 and I7, the E7 is a higher end much more powerful chip. You don't know a lot about this stuff, that much is clear. broadwell also has AVX2, which is 256 bit, so it isn't that far behind, the new soon to be released Skylake EP (the server versions) will havd 512 Bit support. 

No there aren't, or if there are, it comes down to single-threaded, poorly vectorized code. Of course, no one buys a 24-core Xeon for single-threaded performance.

 

E7 is not much more powerful. In fact the 2699 in any generation is more powerful, but you can't hook as many up per node because Intel disabled the QPI links. Also, where did I say Haswell? It beats BROADWELL in most tasks. The problem is the lack of software optimized for Power and Power 8's excessive purchase price.

 

AVX2 is half as powerful as IBM's 512-bit SIMD. Even with half the the core count, Power 8's superior clock speeds and lower thread comm overhead decimate Broadwell in many supercomputing tasks.

 

https://www-356.ibm.com/partnerworld/wps/servlet/download/DownloadServlet?id=7JQ7IVVYmttiPCA$cnt&attachmentName=Broadwell_EX_10x_per_Dollar_Claim_v1_3.pdf&token=MTQ4MzA4NzI0NjMyMQ==&locale=en_ALL_ZZ

 

106 different compute tasks, and IBM beats Intel Broadwell EX in 94 of them.

 

http://global.sap.com/solutions/benchmark/sd2tier.epx

And another 30, just for good measure.

 

My specialization as a master's student was HPC and heterogeneous computing. I know more than most, and I believe that includes you given the evidence.

[Lord Nicoll]

1 minute ago, patrickjp93 said:

No there aren't, or if there are, it comes down to single-threaded, poorly vectorized code. Of course, no one buys a 24-core Xeon for single-threaded performance.

 

E7 is not much more powerful. In fact the 2699 in any generation is more powerful, but you can't hook as many up per node because Intel disabled the QPI links. Also, where did I say Haswell? It beats BROADWELL in most tasks. The problem is the lack of software optimized for Power and Power 8's excessive purchase price.

 

AVX2 is half as powerful as IBM's 512-bit SIMD. Even with half the the core count, Power 8's superior clock speeds and lower thread comm overhead decimate Broadwell in many supercomputing tasks.

 

https://www-356.ibm.com/partnerworld/wps/servlet/download/DownloadServlet?id=7JQ7IVVYmttiPCA$cnt&attachmentName=Broadwell_EX_10x_per_Dollar_Claim_v1_3.pdf&token=MTQ4MzA4NzI0NjMyMQ==&locale=en_ALL_ZZ

 

106 different compute tasks, and IBM beats Intel Broadwell EX in 94 of them.

2699 is really not enough, there are =two xeons with that name, 

E5 2699 v3 (Haswell EP)

E5 2699 v4 (Broadwell EP)

 

Intel disabled the QPI on the lower end ones because they don't make sense and often can't be scaled for various reasons like they use the lesser silicon that might have defects that have been disabled or lazed off.  IBM's 512-bit SIMD and AVX2 are basically names for the same thing, but the same disadvantages hinder both. Clock speeds don't matter when compare completely different system types, RISC needs a high clock speed to get decent MIPS performance, because it's RISC not CISC. Take an I7 2700k and i7 6700k and clock them both at 3GHz and watch the i7 6700k destroy the i7 2700k. The reason there isn't much software for power8 is because RISC is pretty specialised and only works well for somethings. I'm not going to read that article because of obvious risk of bias coming from IBM, don't forget IBM make lots of servers with Intel Xeon CPUs for a reason.

[patrickjp93]

2 minutes ago, Lord Nicoll said:

2699 is really not enough, there are =two xeons with that name, 

E5 2699 v3 (Haswell EP)

E5 2699 v4 (Broadwell EP)

 

Intel disabled the QPI on the lower end ones because they don't make sense and often can't be scaled for various reasons like they use the lesser silicon that might have defects that have been disabled or lazed off.  IBM's 512-bit SIMD and AVX2 are basically names for the same thing, but the same disadvantages hinder both. Clock speeds don't matter when compare completely different system types, RISC needs a high clock speed to get decent MIPS performance, because it's RISC not CISC. Take an I7 2700k and i7 6700k and clock them both at 3GHz and watch the i7 6700k destroy the i7 2700k. The reason there isn't much software for power8 is because RISC is pretty specialised and only works well for somethings. I'm not going to read that article because of obvious risk of bias coming from IBM, don't forget IBM make lots of servers with Intel Xeon CPUs for a reason.

V4!

 

No! The 2699s just get reduced QPI lanes because Intel has product segmentation to do. It's not about lesser silicon. It's not about defects either.

 

No, IBM's SIMD can do twice as much work in the same period of time.

 

MIPS has nothing to do with actual performance. It's a worthless metric no one pays attention to today, not even ARM.

 

IBM likes money and has a huge cloud infrastructure and integration team (it builds supercomputers). Intel has no such infrastructure, but someone has to play against Amazon and Google for a piece of the cloud computing pie.

 

Not reading the article b/c you think it's biased? What, are you 10? Pick apart the bias if you think it's there.

[Lord Nicoll]

1 minute ago, patrickjp93 said:

V4!

 

No! The 2699s just get reduced QPI lanes because Intel has product segmentation to do. It's not about lesser silicon. It's not about defects either.

 

No, IBM's SIMD can do twice as much work in the same period of time.

 

MIPS has nothing to do with actual performance. It's a worthless metric no one pays attention to today, not even ARM.

 

IBM likes money and has a huge cloud infrastructure and integration team (it builds supercomputers). Intel has no such infrastructure, but someone has to play against Amazon and Google for a piece of the cloud computing pie.

 

Not reading the article b/c you think it's biased? What, are you 10? Pick apart the bias if you think it's there.

Something written by a rival company saying how their CPUs are better? Yeah definitely won't have bias. RISC and CISC aren't the same, stop thinking they are, and  SPARC M7 kicks Power8's ass and possibly Broaswell EP but hasn't seen wide use yet. A RISC CPU will compute differently, and will be better at certain workloads. Virtualisation is an important part of modern computing and supercomputer clusters, Intel has better support for it too.

[patrickjp93]

4 minutes ago, Lord Nicoll said:

Something written by a rival company saying how their CPUs are better? Yeah definitely won't have bias. RISC and CISC aren't the same, stop thinking they are, and  SPARC M7 kicks Power8's ass and possibly Broaswell EP but hasn't seen wide use yet. A RISC CPU will compute differently, and will be better at certain workloads. Virtualisation is an important part of modern computing and supercomputer clusters, Intel has better support for it too.

And you're not smart enough to see through it if it's there?

 

They basically are today. Everyone has MuOps be it Intel, AMD, Sparc, Power, MIPS64, or ARM. Everyone has superscalar out of order processing, and everyone has MIMD instructions at this point. Calling any of the major architectures RISC at this point is just stupid.

[WereCat]

These were the same rumors with 6700k, 7700k.... will believe when they release.

[Sir Asvald]

This thread. 

 

 

 

 

 

My head 

 

[patrickjp93]

12 minutes ago, Abdul201588 said:

This thread. 

 

 

 

 

 

My head 

 

Not sure if picture failed to load or if you're saying the thread's empty.

[Gibbons]

https://translate.google.it/translate?hl=it&sl=auto&tl=en&u=http%3A%2F%2Fwww.cpchardware.com%2Fcpc-hardware-n31-precisions-elucubrations%2F

 

Quote

We hide some easter eggs regularly in the magazine. In March, we had coded "Intel GPU = AMD" in a binary string of the CPU page of the purchase guide. Virtually no one had seen the information at that time when it made a big noise 6 months later. In short, we are teasing. The presence of the chain of the current issue, which decodes "ZenOC @ Air = 5G" in this issue people talking on forums for 2 days. Being unmasked, we owe you some details. First, we did not summarize a test in a few bits. If we had been able to test ourselves overclocking, we would have told you openly in the preview. In spite of everything, we know with almost certainty that the CPU that we used for the tests actually came close to the 5 GHz with an (huge) air-dissipator. The I / O multiplier is not clamped at this time and is configured in steps of 0.25x. One heart, however, was active; The Motherboard VRMs seemed at that time too unstable to test with all of the cores.  Other Ryzen ES are currently in the hands of overclockers and you should not delay to learn more: a demonstration of overclocking could occur at the CES if good results are achieved.

 

It's real but only 1 Core active.

 

Reminder - it's google translate so that is why there is some weirdness in the text. Heart = Core and (huge) air-dissipator probably a Noctua D15 equivalent.

[Fetzie]

23 minutes ago, patrickjp93 said:

Not sure if picture failed to load or if you're saying the thread's empty.

They're saying the thread has gone over their head

 

The OC is impressive, but I'm guessing they disabled a couple of cores (or maybe 6 or 7) or they disabled SMT. That's usually how the pre-release overclocks are reached.

[Sir Asvald]

34 minutes ago, patrickjp93 said:

Not sure if picture failed to load or if you're saying the thread's empty.

Bain failed to load. the stuff you were talking about, IBM and Intel. Went over my head.

[patrickjp93]

7 minutes ago, Abdul201588 said:

Bain failed to load. the stuff you were talking about, IBM and Intel. Went over my head.

There are nerds, and then there are NERDS. Don't feel bad. No one outside of HPC has a reason to know the basics, let alone the ins and outs of the industry.

[Simon771]

8 hours ago, patrickjp93 said:

No. The 6800K, 6850K, and 6900K are all 8-core dies, with the 6800/50 having 2 disabled. The 6950X is a 12-core die with 2 cores disabled. The mainstream chips are quad cores and dual cores with varying amounts of cache and hyper threading disabled. The mainstream quads are roughly 150mm sq. in size. The 8-core HEDT is roughly 220.

Disabled or removed connections to those cores?

If i7 6800k just have 2 cores disabled, I'm pretty sure someone would find a way to enable it.

So do you have any info if it's only disabled or Intel removed (with laser probbably) connections to those 2 cores?

 

Also as far as it goes for Zen at 5GHz on all cores ... yeah if it will do 4,2GHz when overclocked, that will be pretty big step. At least if it does have IPC like Intel.

[patrickjp93]

8 minutes ago, Simon771 said:

Disabled or removed connections to those cores?

If i7 6800k just have 2 cores disabled, I'm pretty sure someone would find a way to enable it.

So do you have any info if it's only disabled or Intel removed (with laser probbably) connections to those 2 cores?

 

Also as far as it goes for Zen at 5GHz on all cores ... yeah if it will do 4,2GHz when overclocked, that will be pretty big step. At least if it does have IPC like Intel.

They're fully "fused off" meaning the electrical connections have been broken via laser, but they're not cut off the die. If "disabled" is the wrong way to say it in general, apologies.

[Fetzie]

15 minutes ago, Simon771 said:

Disabled or removed connections to those cores?

If i7 6800k just have 2 cores disabled, I'm pretty sure someone would find a way to enable it.

So do you have any info if it's only disabled or Intel removed (with laser probbably) connections to those 2 cores?

 

Also as far as it goes for Zen at 5GHz on all cores ... yeah if it will do 4,2GHz when overclocked, that will be pretty big step. At least if it does have IPC like Intel.

Since people unlocked extra cores on their phenoms, AMD and Intel laser-cut the traces to the disabled cores.

[Simon771]

7 minutes ago, patrickjp93 said:

They're fully "fused off" meaning the electrical connections have been broken via laser, but they're not cut off the die. If "disabled" is the wrong way to say it in general, apologies.

Ah well that's what I was expecting. Was just hoping that there was some way to enable those 2 cores. But yeah Intel isn't so dumb anymore lol

Just now, Fetzie said:

Since people unlocked extra cores on their phenoms, AMD and Intel laser-cut the traces to the disabled cores.

Yeah I was 99% sure they are doing that ... same as Nvidia has done with 1070. Basicly they laser-cut some cips from 1080 and make it a bit slower.

[Fetzie]

2 minutes ago, Simon771 said:

Ah well that's what I was expecting. Was just hoping that there was some way to enable those 2 cores. But yeah Intel isn't so dumb anymore lol

Yeah I was 99% sure they are doing that ... same as Nvidia has done with 1070. Basicly they laser-cut some cips from 1080 and make it a bit slower.

They can also be chips where the cut-off cores weren't fully functional in the first place, due to manufacturing flaws.

[Simon771]

3 minutes ago, Fetzie said:

They can also be chips where the cut-off cores weren't fully functional in the first place, due to manufacturing flaws.

Wasn't that the case with AMD 3 cores CPUs. Not sure which generation, but I remember some weird CPUs with only 3 cores, and after I googled around a bit, I found out that they meant to make 4 core CPU, but failed somewhere in proces so they just disabled 4th core and make it effective 3 core CPU.

I hope I'm not mistaken.

[leadeater]

5 hours ago, patrickjp93 said:

-snip-

IBM is like the true god father of super computing/scale out computing. Sure Cray and SGI used to be way up there long time ago but SGI and Cray are starting to come back now, which is nice. Not that it matters who builds what anyway.

 

My dad used to service SGI Skyhawk tactical flight simulators for the air force back in the 90's, actually think they were Israeli software. Hard to remember was some time ago.

 

The IBM Power CPUs are fantastic and lets not forget it wasn't that long ago Intel CPUs were the laughing stock of super computing and AMD was in a massive second place just behind IBM Power. GPUs weren't in super computers and wasn't even being considered and Intel Xeon Phi wasn't even in anyone's imagination yet let alone on a road map. We aren't even talking that large of a time span either, technology has moved forward very quickly in some areas recently.

 

https://www.top500.org/list/2003/06/

 

Also IBM Cell, see IBM can tell jokes..... 

 

@Lord Nicoll

Oh and power is extremely important, heat. Large scale cooling is complex so the difference between 10MW and 12MW could make or break planning wise.