AMD zen based Opteron slides showcased in a debate at CERN

[Prysin]

3 hours ago, patrickjp93 said:

Not true. They've done iGPU SKUs for big customers like Google before, and they've had compute-capable iGPUs since Ivy Bridge.

having it and marketing it is two different things. ive never seen or heard of intel actually marketing their iGPUs for GPGPU purposes before now. And i think it is because Tesla or Firepro cards has been exponentially more powerful and supported a much wiider range of software.

[patrickjp93]

2 hours ago, Prysin said:

having it and marketing it is two different things. ive never seen or heard of intel actually marketing their iGPUs for GPGPU purposes before now. And i think it is because Tesla or Firepro cards has been exponentially more powerful and supported a much wiider range of software.

In GPGPU, compute power stacks regardless of which accelerators you use. I can run an OpenCL kernel on a Tesla, a FirePro, and Intel's iGPU all at the same time.

[Dabombinable]

6 hours ago, Coaxialgamer said:

pretty sure that was the core 2 quad

Core Duo were effectively 2x updated Pentium III (Dothan) in the same die (Yonah). Core 2 Duo used a different architecture that was based on P6 just like Yonah.

[Coaxialgamer]

Just now, Dabombinable said:

Core Duo were effectively 2x updated Pentium III (Dothan) in the same die (Yonah). Core 2 Duo used a different architecture that was based on P6 just like Yonah.

i heard something about the core 2 quad being made of 2 dual core modules ( full cores, not like the fx stuff)

[Coaxialgamer]

5 hours ago, patrickjp93 said:

Look at the highest price of an E5 Xeon. That will be the approximate cost.

probably more like the price of e7's though

[Dabombinable]

1 minute ago, Coaxialgamer said:

i heard something about the core 2 quad being made of 2 dual core modules ( full cores, not like the fx stuff)

The 65nm core 2 Quad were 2x dual core dies on the same package. Its at 45nm that we had proper quad cores.

[Tomsen]

On 12/2/2016 at 2:34 PM, patrickjp93 said:

Not true. They've done iGPU SKUs for big customers like Google before, and they've had compute-capable iGPUs since Ivy Bridge.

So what are the specs of those SKUs? In what way are they operating? Is Google actually using it for GPGPU?

Also, having few big customers, are a still a starting point. Because those big customers, can do a lot on their end, to get things to work, something that isn't true for the wider market.

[Watermelon Guy]

On 2/10/2016 at 2:30 PM, don_svetlio said:

Still. 32-core Opteron vs 18-core Xeon. I can already smell Linus' vulture-like eyes and grin as he thinks the same

I would imgaine, if AMD is to be believed (lets assume for a moment they are), this cpu should perform 2x better than the 18 core xeon, assuming you can find a program with perfect scaling. 

[Dabombinable]

11 minutes ago, Watermelon Guy said:

I would imgaine, if AMD is to be believed (lets assume for a moment they are), this cpu should perform 2x better than the 18 core xeon, assuming you can find a program with perfect scaling. 

Going off AMD's track record though, that's not likely to happen unless Intel has another problem such as Netburst (rekt by Pentium III and Pentium M consuming a lot less power while running just as fast clocked 1GHz lower).

[patrickjp93]

1 hour ago, Tomsen said:

So what are the specs of those SKUs? In what way are they operating? Is Google actually using it for GPGPU?

Also, having few big customers, are a still a starting point. Because those big customers, can do a lot on their end, to get things to work, something that isn't true for the wider market.

Unknown, but that's why they're custom. And further it's still training for Intel. AMD hasn't had much success at all with their server FirePros, so I'd argue they're on even footing.

[patrickjp93]

43 minutes ago, Watermelon Guy said:

I would imgaine, if AMD is to be believed (lets assume for a moment they are), this cpu should perform 2x better than the 18 core xeon, assuming you can find a program with perfect scaling. 

At the same clock rate*

And don't forget Broadwell E7 goes up to 24 cores and Skylake will push up to 28 with fully custom SKUs for a range of customers.

[Prysin]

33 minutes ago, Dabombinable said:

Going off AMD's track record though, that's not likely to happen unless Intel has another problem such as Netburst (rekt by Pentium III and Pentium M consuming a lot less power while running just as fast clocked 1GHz lower).

thing is though, intel is hitting a wall in terms of performance. they are at 14nm and look how much they improved from their tick to their tock. aka Broadwell to Skylake. They litterally got like 5% better performance...

If AMD hits Haswell performance.

If 32 cores + HT is true

if cache latencies and bus latencies isnt bad

IF the 8 channel memory thing is true

 

Then intel wont have much to show with until they hit 10nm. And speaking of 10nm, they have hinted that Cannonlake will be delayed AGAIN, for the fourth time in a row... meaning unless Kaby Lake increases performance drastically. The projections for ZEN+ should be within 10% of whatever Intel has on the market when it releases.

 

Meaning realistically, Intel will be fighting on power efficiency more then raw performance. And for servers, that is a BIG argument.

5-10w lower power use, for 5-10% higher performance, even if initial purchase price is 1000 USD more per chip, can easily be regained since 24/7 operation draws A LOT of power.... Said chips would probably pay the price difference back within half a year to a year...

[Dabombinable]

3 minutes ago, Prysin said:

thing is though, intel is hitting a wall in terms of performance. they are at 14nm and look how much they improved from their tick to their tock. aka Broadwell to Skylake. They litterally got like 5% better performance...

If AMD hits Haswell performance.

If 32 cores + HT is true

if cache latencies and bus latencies isnt bad

IF the 8 channel memory thing is true

 

Then intel wont have much to show with until they hit 10nm. And speaking of 10nm, they have hinted that Cannonlake will be delayed AGAIN, for the fourth time in a row... meaning unless Kaby Lake increases performance drastically. The projections for ZEN+ should be within 10% of whatever Intel has on the market when it releases.

 

Meaning realistically, Intel will be fighting on power efficiency more then raw performance. And for servers, that is a BIG argument.

5-10w lower power use, for 5-10% higher performance, even if initial purchase price is 1000 USD more per chip, can easily be regained since 24/7 operation draws A LOT of power.... Said chips would probably pay the price difference back within half a year to a year...

That's kind of the problem with AMD. Its always "IF" with their next lot of products.

[Prysin]

10 minutes ago, Dabombinable said:

That's kind of the problem with AMD. Its always "IF" with their next lot of products.

thats the problem for everyone.

 

Intel has been like "yeah, cannonlake early 2016 fo sho my man".... then they were like "Cannonlake 2016 Q3/Q4, ok?" then not long after that "cannonlake 2017 for realz" and now "Cannonlake maaaaaybe 2018...... deal with it (tm)"

 

Do you think Nvidia will be any better with their goals for pascal? 10x compute?
10x next to nothing = barely acceptable

You know, the reason maxwell is so damn efficient, is because all the hardware components to do compute was thrown out of the PCB. Cutting wattage and placing more strain on driver reliability. If Pascal were to be 28nm, it would be less efficient then maxwell due to Nvidia putting the hardware scheduler  back into the PCB alongside a more compute heavy GPU...

 

You aint getting nowere as a company unless you aim big. Thing is, how big?

If you look at AMD the issue is mostly in the marketing dpt. It is obvious that there is a disconnection between marketing and engineering. Their engineers and top management is VERY clear on their issues and doesnt try to hide these. IF you look at a interview with Roy Tailor from AMD some 9 months ago, you would hear him openly admitting that their current CPUs were "less then great".... when you as a top brass of a company says that, the importance of such a statement and the weight behind it means that you are admitting utter defeat.

Top brass CANNOT say negative things, but why? because in business, there is no definition between "slightly bad news" and "bad news". Investors only want good news. Back to Cannonlake, do you think Intels investors are happy about delays? Hell no. Someone somewhere is getting hammered for these delays, imagining anything else is pure idiocy.

AMD however cannot lose investors, they CAN NOT fail. So how do you cope when you CAN NOT FAIL?. You make grand promises, that isnt removed from the possibility of being true.

40% over excavator? Such numbers are NOT out of possibility.

Energy efficient? Not unlikely, 14nm FF+ will bring A LOT of benefits to AMD... remember, FinFET technology allows for much finer voltage control, and intel has had this technology since Sandy or Ivy Bridge at least. Just FinFET alone is a huge deal. 14nm in addition to this will save even more. Remember, AMD is currently at 32nm PLANAR gates. To give you a hint how old PLANAR is... it was invented and put into use around the 60s.... Whilst there has been improvements to this gate type over the decades, the principle has remained somewhat similar, and this brings an issue in regards to leakage current when the gate is closed. FinFET solves this to a much higher degree.

SMT in and of itself is an old ass concept developed by IBM in 1968, fun fact, first CPU that was planned  to be commerically availible with SMT was due in 2004, but the project got scrapped in 2001. It never saw the light of day. One year later, in 2002, Intel released the Pentium 4 (Netburst) which was the first CPU with HT.

 

So, HT is totally likely. 40% Improvement is very likely, energy savings are 100% going to happen (how much?). All that remains is execution.

 

 

[looncraz]

On 2/12/2016 at 2:43 PM, Coaxialgamer said:

i heard something about the core 2 quad being made of 2 dual core modules ( full cores, not like the fx stuff)

Core 2 Quad.

 

Intel also did the same with their original dual cores (Pentium D).

 

[desertcomputer]

Image if it supports two CPUs ... 64 cores....

[Prysin]

Just now, MrUnknownEMC said:

Image if it supports two CPUs ... 64 cores....

AMD is known to feature quad CPU boards. Intel did too until recently i think, when they went with dual CPU boards. You can prolly GET a quad Xeon board, but i havent seen or heard of one for the current skylake/broadwell generation

[patrickjp93]

On February 14, 2016 at 7:16 AM, Prysin said:

thats the problem for everyone.

 

Intel has been like "yeah, cannonlake early 2016 fo sho my man".... then they were like "Cannonlake 2016 Q3/Q4, ok?" then not long after that "cannonlake 2017 for realz" and now "Cannonlake maaaaaybe 2018...... deal with it (tm)"

 

Do you think Nvidia will be any better with their goals for pascal? 10x compute?
10x next to nothing = barely acceptable

You know, the reason maxwell is so damn efficient, is because all the hardware components to do compute was thrown out of the PCB. Cutting wattage and placing more strain on driver reliability. If Pascal were to be 28nm, it would be less efficient then maxwell due to Nvidia putting the hardware scheduler  back into the PCB alongside a more compute heavy GPU...

 

You aint getting nowere as a company unless you aim big. Thing is, how big?

If you look at AMD the issue is mostly in the marketing dpt. It is obvious that there is a disconnection between marketing and engineering. Their engineers and top management is VERY clear on their issues and doesnt try to hide these. IF you look at a interview with Roy Tailor from AMD some 9 months ago, you would hear him openly admitting that their current CPUs were "less then great".... when you as a top brass of a company says that, the importance of such a statement and the weight behind it means that you are admitting utter defeat.

Top brass CANNOT say negative things, but why? because in business, there is no definition between "slightly bad news" and "bad news". Investors only want good news. Back to Cannonlake, do you think Intels investors are happy about delays? Hell no. Someone somewhere is getting hammered for these delays, imagining anything else is pure idiocy.

AMD however cannot lose investors, they CAN NOT fail. So how do you cope when you CAN NOT FAIL?. You make grand promises, that isnt removed from the possibility of being true.

40% over excavator? Such numbers are NOT out of possibility.

Energy efficient? Not unlikely, 14nm FF+ will bring A LOT of benefits to AMD... remember, FinFET technology allows for much finer voltage control, and intel has had this technology since Sandy or Ivy Bridge at least. Just FinFET alone is a huge deal. 14nm in addition to this will save even more. Remember, AMD is currently at 32nm PLANAR gates. To give you a hint how old PLANAR is... it was invented and put into use around the 60s.... Whilst there has been improvements to this gate type over the decades, the principle has remained somewhat similar, and this brings an issue in regards to leakage current when the gate is closed. FinFET solves this to a much higher degree.

SMT in and of itself is an old ass concept developed by IBM in 1968, fun fact, first CPU that was planned  to be commerically availible with SMT was due in 2004, but the project got scrapped in 2001. It never saw the light of day. One year later, in 2002, Intel released the Pentium 4 (Netburst) which was the first CPU with HT.

 

So, HT is totally likely. 40% Improvement is very likely, energy savings are 100% going to happen (how much?). All that remains is execution.

 

 

Broadwell E5 Quad boards tend to only be available directly through Intel. And then there are 8-socket boards for E7 too.

[linuxfan66]

AMD is known to feature quad CPU boards. Intel did too until recently i think, when they went with dual CPU boards. You can prolly GET a quad Xeon board, but i havent seen or heard of one for the current skylake/broadwell generation

I have in a fifty grand fell server(base specs price)