AMD Vishera Packs Quad-Channel DDR3 IMC, G34 En Route Desktop?

[btarunr]

AMD might be a little sore that its "Zambezi" FX processor family based on its much-hyped "Bulldozer" architecture didn't quite meet the performance expectations of a ground-up new CPU architecture, but it doesn't want to take chances and build hype around the architecture that succeeds it. From various sources, some faintly-reliable, we have been hearing that the next-generation of high-performance desktop processors based on "Piledriver" architecture, codenamed "Vishera", will pack five modules or 10 cores, and will be structured essentially like Zambezi, since Piledriver is basically a refinement of Bulldozer architecture. The latest leak comes from the Software Optimization Guide for AMD 15h family (read here), which was picked up by CPU World while most of us were busy with CES.

CPU World compiled most of the features of what it suspected to be AMD referring to its future processors based on the Piledriver architecture, that's "Vishera" (desktop high-performance), "Terramar" (high-density server), and "Sepang" (small-medium business server) parts. The three are not the first chips to be based on Piledriver, AMD has a new mainstream desktop and notebook APU in the works codenamed "Trinity", which is en route for a little later this year. Trinity basically has an identical CPUID instruction-set as Vishera, Terramar, and Sepang, confirming their common lineage compared to today's "Bulldozer" architecture. The most catchy detail is of Vishera featuring 4 DDR3 channels.



The plot thickens where "HyperTransport Assist feature" is listed as being supported on Vishera. HT Assist is a feature found on AMD's enterprise socket G34 processors, which facilitates better inter-die communication between the two dies of a typical socket G34 Opteron processor. The G34 (LGA1972) package is a multi-chip module of two quad-core, six-core, or four-module dies, which combined have four DDR3 memory channels, and a number of HyperTransport links to communicate with neighbouring sockets and the system's chipset. Could this be the first indication that AMD wants to take on Intel LGA2011 HEDT (high-end desktop) using Vishera chips in the G34 package? It will be a while before we find out.

Apart from using common silicon between client and enterprise platforms, AMD does have a history of colliding the two.

Source: CPU World

[Hustler]

Well unless Vishera's IPC increases by about 50% over Bulldozer, they may as well not bother, no matter how many cores they shove into it to try and disguise it's shortcomings.

[TheMailMan78]

So does this mean AMD will change their sockets finally?

[repman244]

Originally Posted by Hustler

Well unless Vishera's IPC increases by about 50% over Bulldozer, they may as well not bother, no matter how many cores they shove into it to try and disguise it's shortcomings.

50% ? I doubt it, more like 15-20%, but that's just my guess.

[meirb111]

instead of adding more cores they must improve preformance per clock or this is all going
to be another amd joke , a deadly joke to the future of amd's cpu market

[Hustler]

Originally Posted by repman244

50% ? I doubt it, more like 15-20%, but that's just my guess.

Given that Bulldozer's IPC is between 5-10% slower than the Phenom II, increasing Vishera's IPC by 15-20% will only make it about 10% faster than Phenom II (on average)...clock for clock.

Like i said, unless you approach an IPC increase of 50% or so, it's going to be seen as a failure, and even with a 50% increase, that will still only bring it up to par with a 2500k (perhaps a little faster).

[suraswami]

so like 939 days we can see Optys for desktop too?

[TheLaughingMan]

I disagree. Bulldozer's issues are boil down to single thread/single core performance and issues with code using older instruction sets. Those are a result of less than optimal core design due to their switch to an automated layout system.

They need to increase L2 cache speed, improve memory throughput more, and bring up low thread count/single core performance by at least 35%. The latter being the most important. Bulldozer is a great chip when all cores are being utilized and scales well. When you through tasks at it that only use 1 thread, or half the available threads with older instruction sets, it chokes.

I have no issues gaming on my Bulldozer except for the whole Deus Ex HR BSOD my rig due to BIOS issues. That is GIGABYTE's fault mostly. What is AMD's fault is when it was working on the older BIOS versions, my FX-8150 came in well behind my 1100T cause the game (for me) ran on 1 thread/core. Insert BF3 (runs at least 4 threads) and the FX-8150 will go toe to toe with a i7 2600. Its that strange behavior and wild swings in performance due to thread count that need to be addressed, not bring quad-channeled memory to desktop space.

[theoneandonlymrk]

im bored of hearing how AMD need to trounce intel with their next cpu, what rubbish, it needs to be good enough to play games and surf plus a bit of transcodeing ,now and again for me and hence BD as it is would do so PD/vishera weva dosnt require 50%(daydreaming improvement),ive given up playing most of the games that run shit on it anyway (single threaded's so last decade),

all these people winging about BD better be folding or cryunching 24/7 as if your pc sits doing nowt while you workin then is only on to game and surf and your still moaning about BD your not right in the head, its like a cyclist moaning about car insurance going up wtf is the point

do these moaners have a CRAY in their bedroom NO, just an intel 2600K ,no doubt wasted on em

[Steevo]

Originally Posted by TheLaughingMan

I disagree. Bulldozer's issues are boil down to single thread/single core performance and issues with code using older instruction sets. Those are a result of less than optimal core design due to their switch to an automated layout system.

They need to increase L2 cache speed, improve memory throughput more, and bring up low thread count/single core performance by at least 35%. The latter being the most important. Bulldozer is a great chip when all cores are being utilized and scales well. When you through tasks at it that only use 1 thread, or half the available threads with older instruction sets, it chokes.

I have no issues gaming on my Bulldozer except for the whole Deus Ex HR BSOD my rig due to BIOS issues. That is GIGABYTE's fault mostly. What is AMD's fault is when it was working on the older BIOS versions, my FX-8150 came in well behind my 1100T cause the game (for me) ran on 1 thread/core. Insert BF3 (runs at least 4 threads) and the FX-8150 will go toe to toe with a i7 2600. Its that strange behavior and wild swings in performance due to thread count that need to be addressed, not bring quad-channeled memory to desktop space.

BIOS BSOD on a game? ORLY......


The BD architecture is completely memory dependent, thats why the tests with higher speed memory fared better than those with memory speeds equaling the Intel systems. Intel has at least 50% of their lead due to much better branch prediction and internal data management, the easiest way to make up for this is more cache at faster speed, and more memory bandwidth as it cuts down the number of idle cycles in soft fault.

[badtaylorx]

is this saying that vashira is going to be 4-4170 bulldozers on a single chip using hyper transport to aid it???

[TheLaughingMan]

Originally Posted by Steevo

BIOS BSOD on a game? ORLY......

The BD architecture is completely memory dependent, thats why the tests with higher speed memory fared better than those with memory speeds equaling the Intel systems. Intel has at least 50% of their lead due to much better branch prediction and internal data management, the easiest way to make up for this is more cache at faster speed, and more memory bandwidth as it cuts down the number of idle cycles in soft fault.

In short for those that don't want to read. I want AMD to not take the easy way out and refine their new architecture, not hide behind tricks to drive competition. They need to turn Bulldozer/Piledriver into something software developers want to utilize to optimize software due to its design. If they could say because are chip is built this way, you could do this and that to reduce instruction fetches or executions or whatever to make your software run faster.

Good enough, fine for what i use it for, I am a gamer and it doesn't matter should not be statements we use to defend the purchase of AMD. They need to first refine the architecture to remove weaknesses and level out its performance across the board. Why? Because not every piece of software is going to follow the more threads the better trend. Some stuff don't need more than 1 or 2 threads and Bulldozer should run those just as well as a program that uses 6 or 8 threads.

Once they have done that they need something, anything they can market as, "With AMD you can do this task much faster" I don't even thing it matters what that one thing is at this point as long as they have a benefit to pick them over Intel to market. Intel has Quick Sync on top of being the big dog. AMD needs to be able to say, "Our chip is more than good enough. It is just as good across the board as Intel and with AMD Super Awesome all single thread applications run 35% faster thanks to blah blah blah."

In short, then need to stop trying to build a faster sports car and concentrate on improving/innovations on their entire vehicle line up from the cheapest compact to the most task specific 3.25 ton truck.

[ensabrenoir]

Wonder if microsoft has already begun working on a patch....

[DigitalUK]

Originally Posted by meirb111

instead of adding more cores they must improve preformance per clock or this is all going
to be another amd joke , a deadly joke to the future of amd's cpu market

AMD bulldozer is no joke when it comes to serious multi thread and gets even better as you push it, only problem is single thread stuff and whos buying an 8 core cpu for single thread these days.

[Dent1]

Originally Posted by meirb111

instead of adding more cores they must improve preformance per clock or this is all going
to be another amd joke , a deadly joke to the future of amd's cpu market

Increasing performance per clock is ideal, but considering adding two extra cores will still compensate and bring performance more inline to Intel. The multi threaded performance will be off the chain. And remember Piledriver is going to be upto 30% faster ontop of this.

As far as I'm concerend this is a good thing, if the price is right.

Why should AMD dumb down their architecture, when the problem is obviously the sofware developers, more specifically the games developers.


Edit:

I'm curious to find out whether quad channel memory will improve performance. Many have said AMD in general lacks memory bandwidth. We shall see. Does that mean we have to move away from AM3+ to enjoy quad channel?

Originally Posted by DigitalUK

piledriver better be on am3+ as its the reason i picked up a 990fx board, i was about to get a 8150p for myself but was holding off as there was rumours of another BD Revision coming and PD is very soon.

Piledriver is definitely on AM3+, its on the old roadmap.

But will the quad channel Piledriver be on AM3+ or is that on the new socket?

[DigitalUK]

piledriver better be on am3+ as its the reason i picked up a 990fx board, i was about to get a 8150p for myself but was holding off as there was rumours of another BD Revision coming and PD is very soon.

[TheGuruStud]

All you need to know about bulldozer is SSE4.1, AVX and FMA.

Intel has always had ICC make AMD use SSE2 (along with unoptimized libraries), now it's more important than ever.

Read up how bulldozer performs when using those instructions

[Steevo]

Wow, it doesn't matter if it can make you look like superman with those instructions if no one uses them for anything that matters.



Adding two more cores and the energy consumption will raise by that factor, and there is only so much heat it can stand before electromigration destroys it, before the power supply pins start to detach as the solder softens, before the cooling solution no longer provides the thermal dissipation needed to cool the package.

[faramir]

Originally Posted by repman244

50% ? I doubt it, more like 15-20%, but that's just my guess.

With refinement of 32 nm manufacturing process at Global Foundries and AMD's better understanding of operation of Bulldozer at 4+ GHz (it is their first chip clocked that fast aftrerall) they are more likely going to be able to reach the targeted clockspeeds so some IPC gains and some clockrate gains could translate into significant gain.

Granted, AMD needs to improve IPC rate of their new architecture in order to be able to keep up with Intel but the architecture itself was devised to be a "speed demon" so it makes sense to expect higher clock rates.

[meirb111]

Originally Posted by DigitalUK

AMD bulldozer is no joke when it comes to serious multi thread and gets even better as you push it, only problem is single thread stuff and whos buying an 8 core cpu for single thread these days.

The reason why bulldozer is a joke is because on the same clock with the same number of cores it has worse preformance than a phenom ii not better how can a new 32nm cpu with the same number of cores be worse than 45nm cpu, amd had to be really stupid to do something like this.

[DigitalUK]

rubbish im on a 1090t @4ghz 3ghz Nb htt 2600 and i built a system for a customer the other day 8120p on cheaper 970 chipset overclocked to 4ghz on stock volts no other tweaks and the BD smoked mine , the main tests were 3dMark11 abit of memory testing etc. no single thread tests tho.

[meirb111]

Originally Posted by digitaluk

rubbish im on a 1090t @4ghz 3ghz nb htt 2600 and i built a system for a customer the other day 8120p on cheaper 970 chipset overclocked to 4ghz on stock volts no other tweaks and the bd smoked mine , the main tests were 3dmark11 abit of memory testing etc. No single thread tests tho.

use only 6 cores vs 6 cores to compare like fx-6100 vs 1090

[WhiteLotus]

I want to know what IOMMU is.

Anyone?

[DigitalUK]

its for VMWare etc to access graphics i believe?

[xaira]

how exactly do they plan to fit a g34 socket and 4 channels of memory on a single atx board?