If AMD were to release a new hexcore chip what should its specs and die layout be?

[ShintaiDK]

Its not free either to move the node. A lot of ICs are still made up to 180nm due to design cost.

 

[waffleironhead]

I'd like a 2+2+2 amd cpu.
2 cat, 2 stars, and 2 equipment cores.
Call it the trifecta of awesomeness.

 

[cbn]

If AMD punted on AM3+ and ceased all production (assuming they have not already), then they must figure out what to do with the 32nm wafers for Piledriver FX (and, by logical extension, Richland/Trinity, again assuming they're still fabbing those). They must also figure out what to do with the wafers currently in use to produce chipset ICs for AM3+ boards. For the sake of simplicity, let us assume that they can not merely increase production of 32nm Piledriver Opterons and associated G34 chipset ICs.

What they could do is move FM2+ chipset production onto their 32nm process (more $$$ there), along with AM1 chipset production. It is doubtful that their new Carrizo BGA systems will require many ICs since so much of the board logic is now on the CPU, though fwiw they could move that to 32nm as well. They could punt on the wafers currently used for AM3+, FM2+, and AM1 chipset ICs and pay a penalty.

AMD makes separate ICs for AM1 and Carrizo SOCs?

 

[cbn]

One idea I brought up in the past is for AMD to drop FM2+ and just focus on Vishera/AM1 for desktop and Carrizo/Cat core SOCs for mobile.

Dropping FM2+ (at least temporarily) should at least move some desktop volume away from 28nm and back to 32nm. This especially if increasing volume of Vishera allows AMD to lower prices on FX-6300 and FX-8310, etc.

Meanwhile (and hopefully) Carrizo is a good mobile design so AMD can get back some 28nm volume it would lose by dropping Kaveri on desktop.

As far as the cat cores go, I hope AMD does a few things to tune those up. Maybe even increase the TDP of the AM1 chips so they hit better clockspeeds. (The performance per watt won't look good on paper but at least the chips will fill the need at lower end desktop better).

Then after all that happens (taking maybe 1 year or so), AMD brings back the FM2+ socket and releases the new multicore SR or EX based hexcore with small iGPU along with some new chipset based on 45nm or 32nm. Ideally the new SR or EX based hexcore chip with small iGPU would be compatible with existing FM2+ motherboards.

 

[cbn]

Tape out an iGPU-less 2M design for Steamroller on FM2+ using 28nm planar
Tape out an iGPU-less 3M design (with modified NB, as sm625 pointed out) for Steamroller on FM2+, again using 28nm planar

???

Profit!

And that's the cheapest, shortest-term path I could see for them to actually get 3M Steamroller on FM2+. That assumes no iGPU whatsoever. Add that in there, and things get more expensive.

In an ideal world, I would definitely prefer iGPU-less. Droping die size 25 to 27.5mm2 by removing a 64 stream processor and 30 to 35mm2 by removing a 128 stream processsor iGPU would let AMD lower price of a cache reduced hexcore processor by a good amount! This, in turn, means more budget conscious folks would be interested.

However, I think there would also need to be some type of chipset integrated graphics available (as an option) like we see on AM3+.

P.S. I wonder how long and how much resources it would take for AMD to modify the northbridge for a hexcore?

 

[Enigmoid]

In an ideal world, I would definitely prefer iGPU-less. Droping die size 25 to 27.5mm2 by removing a 64 stream processor and 30 to 35mm2 by removing a 128 stream processsor iGPU would let AMD lower price of a cache reduced hexcore processor by a good amount! This, in turn, means more budget conscious folks would be interested.

However, I think there would also need to be some type of chipset integrated graphics available (as an option) like we see on AM3+.

P.S. I wonder how long and how much resources it would take for AMD to modify the northbridge for a hexcore?

And pretty much eliminate OEM shipments.

6M IGP less chip is not going in laptops, and its not going to sell well in in OEM desktops. That leaves the BYO desktop market.

They would also have to price it below the current price of the 6300 which can be had for <$100.

Even intel, with the resources and volume to ship igp less consumer chips is not. Why would this make sense for AMD?

 

[cbn]

And pretty much eliminate OEM shipments.

6M IGP less chip is not going in laptops, and its not going to sell well in in OEM desktops. That leaves the BYO desktop market.

Like I mentioned, there would also need to be some form of chipset integrated graphics available as an option.

 

[cbn]

They would also have to price it below the current price of the 6300 which can be had for <$100.

Yes, that definitely needs to happen. But with that four module Vishera chip (with whopping 6MB L2 and 8MB L3) at 315mm2 on 32nm, a lowered cache dedicated hexcore on 28nm should weigh in with a much smaller die size.

 

[cbn]

Besides what I wrote in post #54 (<---Hexcore with small iGPU on processor die), AMD could always come out with two dies for AM3+ or whatever follows after that (AM4??):

Fast clocked, cache reduced, hexcore SteamRoller (for desktop gaming enthusiasts) as one die and a much larger higher core count, higher cache, Excavator (with HDL) for the second die aimed at server and a niche class of multi-media enthusiasts. Both without iGPU on the processor die.

Then have two chipsets available, one with integrated graphics and one without.

In any event, I would much rather see AMD focus more on cpu and less on iGPU (with the exception of the niche big core APUs like "Carrizo" for gaming notebooks).

 

[Enigmoid]

Yes, that definitely needs to happen. But with that four module Vishera chip (with whopping 6MB L2 and 8MB L3) at 315mm2 on 32nm, a lowered cache dedicated hexcore on 28nm should weigh in with a much smaller die size.

8350 die is actually 8 MB L2 and 8 MB L3

Besides what I wrote in post #54 (<---Hexcore with small iGPU on processor die), AMD could always come out with two dies for AM3+ or whatever follows after that (AM4??):

Fast clocked, cache reduced, hexcore SteamRoller (for desktop gaming enthusiasts) as one die and a much larger higher core count, higher cache, Excavator (with HDL) for the second die aimed at server and a niche class of multi-media enthusiasts. Both without iGPU on the processor die.

Then have two chipsets available, one with integrated graphics and one without.

In any event, I would much rather see AMD focus more on cpu and less on iGPU (with the exception of the niche big core APUs like "Carrizo" for gaming notebooks).

Still makes no sense. Why tape out a second die for server using steamroller and not excavator (aimed at perf/W)?

 

[cbn]

Still makes no sense. Why tape out a second die for server using steamroller and not excavator (aimed at perf/W)?

The cache reduced Steamroller hexcore die is primarily for gaming desktop and the higher core count excavator die is for multi-thread perf/W.

This assuming Steamroller has higher single thread performance compared to Excavator (with HDL).

P.S. I would assume the higher core count excavator (with HDL) would be at least five or six modules (most likely six modules).

 

[Lepton87]

Six modules but with MCM making a total of 12, otherwise it would be pointless, 24C are really needed to compete with 18C monsters from Intel. Intel has better IPC and HT, so it would still be faster but 18C Xeons cost 4500$ so AMD offering for 1500$ might be actually good, provided it wasn't used with software licensed per core. Maybe they should call a module a core with SMT instead to half the licensing cost? Intel gets away with 2T per core...

 

[DrMrLordX]

AMD makes separate ICs for AM1 and Carrizo SOCs?

I don't really know what kind of board logic is required to implement AM1 or Carrizo. Obviously, neither requires a Southbridge. It may not be anything large or expensive enough that AMD must produce it in a GF fab by themselves.

One idea I brought up in the past is for AMD to drop FM2+ and just focus on Vishera/AM1 for desktop and Carrizo/Cat core SOCs for mobile.

Not happening. Remember that WSA? They have 28nm fab capacity whether or not they want it.

However, I think there would also need to be some type of chipset integrated graphics available (as an option) like we see on AM3+.

That's really up to the motherboard manufacturers. However, AMD has managed to sell quite a few Piledriver FX chips with no iGPU to people using boards with no integrated graphics.

P.S. I wonder how long and how much resources it would take for AMD to modify the northbridge for a hexcore?

Unknown, but the entire process of preparing an iGPU-less 2M or 3M Steamroller chip would take months, at least. Maybe over a year.

Even intel, with the resources and volume to ship igp less consumer chips is not. Why would this make sense for AMD?

AMD has transistor budget problems, and they are currently selling a whole variety of iGPU-less chips (FX chips, 750k, 760k, 860k). It would actually make some sense, if it weren't for fab capacity issues and suchlike.

 

[Blitzvogel]

AMD really needs to converge platforms. If they could design a socket that could handle APU, FX as well as server chips (with variable TDW capacities and memory channels disabled for desktop), I think that would be good for them. Though it would be good for them to segway into quad channel for APUs.

All their mobile stuff should be one socket or BGA type too. If there is a reason to include extra peripheral control not supported by the SoC, it can be tethered on via the mobo. One platform that could for example run Beema, Mullins, or Carrizo would make it easy for OEMs to just swap chips for custom product orders.

 

[DrMrLordX]

That is actually a pretty good idea. A convergence of AM3+, FM2+, and AM1 on the desktop (with the option to or to not include a Southbridge chip for OEMs/integrators) would be . . . well not trivially simple, but still a very good thing for AMD users.

That being said, you'd have enough board variation on such a platform that some boards would barely resemble others using the same socket. As mentioned above, some el-cheapo boards would simply ship without the Southbridge, obviously supporting only Kabini (and/or Carrizo, supposing AMD ever released it for that platform). There would also probably be segmentation based on whether or not the board manufacturer included VGA/DVI/HDMI for a hypothetical APU (some high-end boards intended for non-APU processors might omit this feature as a cost-saving measure). So . . . while a unified socket might be doable, the final incarnation of the socket wouldn't look too much different from the environment we have today.

Overall, it'd make it easier for AMD to keep their platform features up-to-date for all desktop chips, and it would definitely improve the HSF situation for AM1 users that aren't trying to cram their systems into the smallest possible enclosure.

 

[cbn]

Not happening. Remember that WSA? They have 28nm fab capacity whether or not they want it.

Well they will have Carrizo. So Carrizo takes the place on Kaveri on 28nm. Then they also have various cat core SOCs on GF 28nm (AM1, etc).

P.S. There are also reports of AMD moving GPU production from TSMC 28nm to GF 28nm:

http://www.techpowerup.com/208431/amd-to-switch-to-globalfoundries-28-nm-shp-node-in-2015.html

http://www.hardocp.com/news/2015/01/02/amd_moving_28nm_gpu_process_to_globalfoundries#.VLl00nuGXcs

(If those reports of GPU production moving to 28nm are true, then I figure this would be even more reason to cease production on Kaveri go as well as drop the FM2+ socket temporarily)

 

[cbn]

Even though I am glad to see AMD making a significant improvement in the Kaveri product line with the upcoming A8-7650K, we have to realize these three things:

1. Kaveri's CPU throttling under iGPU load. (Because this happens, Kaveri is not quite a good as it looks on paper)

2. Lack of memory bandwidth for the expensive to integrate 512 stream processor iGPU.

3. Increased expense (~$11) of DDR3-2400 kits over 2 x 4GB DDR3 1600 kits. (At one time 2 x4 GB DDR 2400 and 2 x 4GB DDR3 1600 were the same price. But this is no longer the case at Newegg. Furthermore, the processors without iGPU can get by with one 4GB DDR3 stick even greater cost savings....whereas the APUs need dual channel to effectively utilize the iGPU portion). Granted 2 x 2 GB kits are available, but they are slower speed than the 2 x 4 GB kits and much more expensive per GB.

Unfortunately, Kaveri on the desktop is just a very expensive product with several significant performance penalties.

Vishera, on the other hand, is a stronger cpu. The major problem I see with Vishera is the high amount of cache bloating the die which increases the cost to manufacture. But between Kaveri and Vishera, I am still thinking Vishera would be the processor to go with on desktop till AMD offers us a better solution. This especially if AMD can increase volume of Vishera by removing competition with Kaveri. Plus, there is the GF 32nm demand to fill.

 

[Soulkeeper]

I've seen it mentioned that "amount of cache" is a bad thing for AMD in a few threads now.
This kinda bothers me because I know if the products performed better nobody would be saying this.

Nobody complained about the amount of cache on core2 CPUs (upto 12MB L2), P4s, I-series, etc. being too high ...

 

[cbn]

I've seen it mentioned that "amount of cache" is a bad thing for AMD in a few threads now.
This kinda bothers me because I know if the products performed better nobody would be saying this.

Haswell Core i7 has 9 MB of combined L2/L3 cache (1MB of L2 and 8MB of L3) and the Kaveri die only has 4MB of L2 (2MB per module) with no L3 cache. This compared to 16 MB of combined L2 /L3 (8MB L2 and 8MB L3) for the full Vishera die.

Of course, if AMD had the performance lead we would probably be willing to forgive some of the excess cache size (like we currently do with Core i5's and Core i7's excessive iGPU size on desktop).

But since AMD does not have the performance crown, we need them to compete better on price. And IMO, this means removing as much of the least productive parts of the die as possible.

 

[Soulkeeper]

I for one don't want to see 1MB cache per core on AMD CPUs/APUs for another 4yrs
software sizes are only getting bigger

vishera's main issue was their weak L3 performance and other design limitations, not the cache size itself

 

[Ramses]

First I've heard of bad cache size too. I've heard everything else that's wrong with the FX lol, but not that.

 

[cbn]

First I've heard of bad cache size too. I've heard everything else that's wrong with the FX lol, but not that.

The problem I see with that 8MB L3 cache is that it takes up a considerable amount of die size relative to the size of cpu modules themselves.

And I've never been impressed looking at gaming comparisons of AM3+ quad cores (which have L3 cache) to FM2 quad cores (which don't have L3 cache). Even in this gaming comparison here with the Piledriver based Athlon x4 750K @ 4.3 Ghz (no L3 cache) vs. FX-4350 (4.2 Ghz/4.3 Ghz turbo, 8 MB L3 cache) the difference in average gaming performance amounts to only a few percent:

 

[Ramses]

Wonder if it's a holdover from the server cpu stuff?

 

[DrMrLordX]

Well they will have Carrizo. So Carrizo takes the place on Kaveri on 28nm. Then they also have various cat core SOCs on GF 28nm (AM1, etc).

Yes, but they have effectively cancelled Carrizo on the desktop. They could cease mobile Kaveri production (assuming they haven't already) and probably fill out that segment of their wafer supply with Carrizo, but that does not address desktop Kaveri production at all. It's not like they're going to crank out more 5350s.

1. Kaveri's CPU throttling under iGPU load. (Because this happens, Kaveri is not quite a good as it looks on paper)

This is easily defeated in software. Also, I've never actually seen it happen in Linux, which is really interesting. In Windows, yes, it's an issue . . . until you run amdmsrtweaker. Then it isn't.

The "cTDP throttling" that seems to happen to Kaveri over 4.5 ghz is board-related and does not happen on the A88x-Pro or Crossblade Ranger.

Most other throttling is related to thermal margin, which means the chip isn't being cooled well-enough. Some of them just scream "delid me!" due to poor application of TIM under the IHS.

2. Lack of memory bandwidth for the expensive to integrate 512 stream processor iGPU.

Take a look at the GPUPI thread. While that is only one niche benchmark, that bench is totally unaffected by NB speed or system memory speed/latency. I believe AtenRa also did some Luxmark runs that showed that memory bandwidth did have an impact on the OpenCL versions of that particular benchmark. So, at least when it comes to compute functions, the iGPU varies in how much bandwidth it really needs.

3. Increased expense (~$11) of DDR3-2400 kits over 2 x 4GB DDR3 1600 kits.

That's still a tiny price spread for DDR3-2400.

Unfortunately, Kaveri on the desktop is just a very expensive product with several significant performance penalties.

I got my 7700k for $104 and my A88x-Pro for $53. Granted, not everyone is going to get a deal on their board like I did, but . . . it ain't that expensive. It certainly won't be when the 7650k comes out.

Vishera, on the other hand, is a stronger cpu.

2M Vishera isn't, and per module I'll take Kaveri any day of the week. 32m SuperPi ftw!

Wonder if it's a holdover from the server cpu stuff?

I'm thinking so, yes. It helps with cache coherency between sockets in MP configurations and . . . other server-y stuff like that thar.

 

[Yuriman]

There's a good reason for keeping FM1 and FM2 separate. The bigger chips need dual channel ram, which increases motherboard cost. The cat cores would stand no chance of competing with Atom in price.