Intel: Tick-Tock cycle dead, now instead "Process-Architecture-Optimization"

[patrickjp93]

Just now, shdowhunt60 said:

I actually do believe that. Especially for AMD's hardware in particular.

You believe in a provable falsehood. All async compute/shading did was reveal that the programmers weren't correctly assessing what could all be done in a handful of synchronous kernels. By the time you need a queue depth bigger than 16, you have already lost your optimality anyway.

[shdowhunt60]

6 minutes ago, patrickjp93 said:

You believe in a provable falsehood. All async compute/shading did was reveal that the programmers weren't correctly assessing what could all be done in a handful of synchronous kernels. By the time you need a queue depth bigger than 16, you have already lost your optimality anyway.

I wasn't even specifically referring to async compute/shading, but okay.

[Reaver Seijuuro]

24 minutes ago, patrickjp93 said:

The 5820K is cheaper, at least in the U.S.. If you mean the 5930K, well, did you expect an additional 24 PCIe lanes over the 6700K to come for free?

Typo there. Meant the 5820K, that's the one that's 100.00 USD more. So I'm checking if going by the new process, would Skylake be a better option than Kaby Lake.

[patrickjp93]

Just now, shdowhunt60 said:

I wasn't even specifically referring to async compute/shading, but okay.

Then what else could you possibly be referring to? We already know Nvidia maxes out its cards' potentials by having their drivers be in top form and by giving game studios card-specific code. For AMD sure there's room for driver improvement, but not much.

[patrickjp93]

Just now, Reaver Seijuuro said:

Typo there. Meant the 5820K, that's the one that's 100.00 USD more. So I'm checking if going by the new process, would Skylake be a better option than Kaby Lake.

The 5820K is cheaper than the 6700K.

http://pcpartpicker.com/parts/cpu/#k=30,28&sort=d7&page=1&s=13

[Reaver Seijuuro]

3 hours ago, patrickjp93 said:

The 5820K is cheaper than the 6700K.

http://pcpartpicker.com/parts/cpu/#k=30,28&sort=d7&page=1&s=13

Not here.

[Camisto]

I like "tick-tock-opt" better.  Rolls off the tongue better than tick-tock-tack (wtf)

 

Or maybe, tick-tock-top(top it out you know)

[Ansau]

8 hours ago, patrickjp93 said:

The 5820K is cheaper than the 6700K.

http://pcpartpicker.com/parts/cpu/#k=30,28&sort=d7&page=1&s=13

Might be in US, but in most other countries the 6700k is cheaper. Also, 1151 socket can be much more cheaper than 2011-3 if we count motherboard, 2-channel ram and cooling solution.

[suicidalfranco]

9 hours ago, patrickjp93 said:

The 5820K is cheaper than the 6700K.

http://pcpartpicker.com/parts/cpu/#k=30,28&sort=d7&page=1&s=13

one out of topic question, since you seem very knowledgeable on this kind of topic: in virtual environment, does hyper threading make a huge difference? are there any advantages to a VM to from a 2c/2t configuration to a 2c/4t? if yes would it be a worthy upgrade to go from a 6500 to a 6700?

[Tomsen]

10 hours ago, patrickjp93 said:

I'll be vindicated on that with the release of Kaby Lake. Why bother to release a 3rd generation? Clock speed alone? Platform alone? No, there will be minor and few architectural improvements. Since Kaby Lake is supporting AVX 512 when (consumer) Skylake doesn't, there goes the theory about no new extensions.

To have "new" products out. Give it a fancy public codename, "unlock" a few features, and you got a "new" product. You're not going to see any changes in the core microarchitecture (I can only imagine fixing potentials bugs). The core remains the same. For servers, you can expect higher core-counts and more aggressive clock-rate (giving the more mature process node and updated libraries).

 

AVX512 is already supported in skylake hardware. That intel is artificially segmenting their product line is a whole other issue. So, my theory holds, no new extentions, not already supported in hardware.

[patrickjp93]

57 minutes ago, suicidalfranco said:

one out of topic question, since you seem very knowledgeable on this kind of topic: in virtual environment, does hyper threading make a huge difference? are there any advantages to a VM to from a 2c/2t configuration to a 2c/4t? if yes would it be a worthy upgrade to go from a 6500 to a 6700?

If you are running multiple VMs or even a single VM that is doing many tasks of moderate to heavy load at the same time, hyperthreading gives a nice boost. http://www.vmwarebits.com/content/vcpu-and-logical-cpu-sizing-hyper-threading-explained

[patrickjp93]

3 minutes ago, Tomsen said:

To have "new" products out. Give it a fancy public codename, "unlock" a few features, and you got a "new" product. You're not going to see any changes in the core microarchitecture (I can only imagine fixing potentials bugs). The core remains the same. For servers, you can expect higher core-counts and more aggressive clock-rate (giving the more mature process node and updated libraries).

 

AVX512 is already supported in skylake hardware. That intel is artificially segmenting their product line is a whole other issue. So, my theory holds, no new extentions, not already supported in hardware.

Actually AVX 512 isn't possible in consumer Skylake chips at all. The vector units are only 256 bits wide. It's not artificial. The investment only is going into E5/E7 Xeons for now.

[Tomsen]

4 minutes ago, patrickjp93 said:

Actually AVX 512 isn't possible in consumer Skylake chips at all. The vector units are only 256 bits wide. It's not artificial. The investment only is going into E5/E7 Xeons for now.

It is not possible at all, because Intel decided to lock it down.  The hardware IS in place to do 512 vector operations, but Intel DECIDED to artificially lock it down.

Lets talk about it from a R&D perspective then. All the research and design is done. Why only limit your investment on certain SKUs? The hardware is in place on consumer products, that is the point I'm trying to get across. 

[patrickjp93]

54 minutes ago, Tomsen said:

It is not possible at all, because Intel decided to lock it down.  The hardware IS in place to do 512 vector operations, but Intel DECIDED to artificially lock it down.

Lets talk about it from a R&D perspective then. All the research and design is done. Why only limit your investment on certain SKUs? The hardware is in place on consumer products, that is the point I'm trying to get across. 

No, the hardware is not in place. Bitsandchips put Skylake under the electron microscope. AVX 512 isn't even physically supported on consumer Skylake. The vector units (1 int, 1 float) are physically only 256 bits wide.

 

And it's to get the most out of Intel's investment. Why wound future sales by giving away too much too soon?

[Tomsen]

21 minutes ago, patrickjp93 said:

No, the hardware is not in place. Bitsandchips put Skylake under the electron microscope. AVX 512 isn't even physically supported on consumer Skylake. The vector units (1 int, 1 float) are physically only 256 bits wide.

 

And it's to get the most out of Intel's investment. Why wound future sales by giving away too much too soon?

From what I've been hearing, the hardware is well in place. Bitsandchips haven't done any such work of what I know of. That generally chipsworks area of expertise. However, you just link that article that perfectly describes how AVX 512 hardware isn't in place on consumer skylake products.

 

You don't get more out of your investment, by limiting it to fewer markets. Unless, if you are thriving on moving more customers from one market segment to another.

That might be Intels plan, who knows, to move more customers over to the more expensive platform/solution.

If future sales are been wounded by modestly supporting AVX 512, then that is only a big warning, imo, regarding Intel future products.

[patrickjp93]

38 minutes ago, Tomsen said:

From what I've been hearing, the hardware is well in place. Bitsandchips haven't done any such work of what I know of. That generally chipsworks area of expertise. However, you just link that article that perfectly describes how AVX 512 hardware isn't in place on consumer skylake products.

 

You don't get more out of your investment, by limiting it to fewer markets. Unless, if you are thriving on moving more customers from one market segment to another.

That might be Intels plan, who knows, to move more customers over to the more expensive platform/solution.

If future sales are been wounded by modestly supporting AVX 512, then that is only a big warning, imo, regarding Intel future products.

You misunderstand. Intel isn't limiting its markets at all, and it probably understands its own economics better than you do.

 

Again, you misunderstand. For servers, you'd have a good point. For consumerville, the entire problem is software. The amount of untapped potential even in Sandy Bridge for consumer software is huge. And further, most of the older instructions have already been optimized to their theoretical best possible conditions. The moment software starts catching up, the life of Sandy Bridge, Ivy, Haswell, Broadwell, and Skylake will be extended. If software starts moving much more quickly, Kaby Lake provides an incentive to upgrade again. Giving away too much on Skylake is economically stupid.

[Tomsen]

On 23/3/2016 at 5:42 PM, patrickjp93 said:

You misunderstand. Intel isn't limiting its markets at all, and it probably understands its own economics better than you do.

 

Again, you misunderstand. For servers, you'd have a good point. For consumerville, the entire problem is software. The amount of untapped potential even in Sandy Bridge for consumer software is huge. And further, most of the older instructions have already been optimized to their theoretical best possible conditions. The moment software starts catching up, the life of Sandy Bridge, Ivy, Haswell, Broadwell, and Skylake will be extended. If software starts moving much more quickly, Kaby Lake provides an incentive to upgrade again. Giving away too much on Skylake is economically stupid.

I'm still interested in that article that shows consumer skylake been examined. That would settle the argument.

 

Patrick, if my statements are correct (that the hardware is in place, but is locked down), then yes, Intel is by definition, limiting certain markets. Perhaps in the attempt to move certain costumers to the more expensive platform/solution. I'm not arguing whether or not that is the best choice from an economics standpoint, I'm simply responding to your statements. (That the investment was only going into certain SKUs).

 

We also all know, that there is a period between when new hardware capabilities are in place until software start properly supporting it. Delaying the hardware support, will show effects on the software support.
In regards to todays performance, it isn't about optimizing the individual instruction execution path, but rather optimizing common "strings" of instruction to extract as much parallelism as possible, optimize cache and I/O.
Why do you think "when software catches up", that it will be about utilizing wider SIMD instructions? SIMD is one thing, but it is not the holy grace that will solve the recent lack of CPU improvement, as SIMD is fundamentally different than "normal" CPU cores and CPU execution. When the times comes, and software support start catching up, wider SIMD is only going to do so much, but the execution flow will quickly be halted due to other limitations.

[AnnoyedShelf]

On 3/22/2016 at 7:33 PM, DocSwag said:

Well, we have been hearing a lot about Moore's law being dead and stuff recently. I guess this is just another confirmation for that. Intel seems to be officially calling the tick-tock cycle dead as well, and are now moving onto a process architecture optimization cycle, or as I like to call it tick-tock-tack (Sorry if it sounds lame but it makes sense). 

I'm pretty sure Intel said it would be Tick-Tock-Semi Tock already.  There's already evidence of this as Kaby Lake is a Semi-tock.