AMD takes the fight to Intel with Ryzen CPU

[DanceswithUnix]

Depending on your definition of 'decently fast', you might be able to do it now with the intel small socket. A B150 motherboard comes in just under £60, and if you don't want to upgrade to a K-series or run multiple graphics cards it should have all the basics that you need in a motherboard. You can't upgrade it to have more that 4 cores, but for sub-4k gaming I can't see any glaring issues. Ryzen will probably have better OC support lower in the range, but a cheap AM4 board won't be too different from a cheap intel board & OC'ing a CPU doesn't get much in modern games.

You can get an FM2 motherboard for £40 and a dual core APU for about £20. Upgrade to something i7 class is fine, i5 would be enough, but FM2+ APUs struggle to match an i3 so dead end.

Intel starts at about £40 for a motherboard, but a Pentium is about £60 so not exactly impulse purchase.

The AM4 socket uses an SoC, so the motherboard is really simple. Hopefully there will be some at the AM1 cost level, and AMD have some Excavator APUs ready to roll on AM4 which could be a cheap way in to the platform.

[scaryjim]

... a Pentium is about £60 so not exactly impulse purchase. ...

Ebuyer has a s1151 Celeron at £40 ... be very interesting to see a benchmark comparison with, say, the A4 6300 (@ £33 currently the cheapest FM2 APU on ebuyer). The AMD setup would be a shade cheaper, but the Intel offers more potential for upgrading...

[Xlucine]

I was talking about the APU itself on an interposer with the HBM2 acting as system RAM. It would also mean,that you could remove the need for a dGPU. Imagine something like a MacBook PRO for example which ships with a dGPU?? HBM2 allows 8GB per package.

You'd remove the need for a dGPU, but not much else. It would be really neat, but the size difference wouldn't amount to any real change in battery capacity or laptop size. Looking at the PCB in this teardown, I'd guesstimate you'd save an area about 75mm^2 or so. if that GPU and PCB come to 3mm thick that'd only be about 10-20cc, and it's an awful lot of expense for such a small volume. For comparison, I estimate the battery volume to be about 150cc[1], so an interposer would make very little difference to the battery you could fit inside - maybe 10% improvement?

[1]: The teardown reports a capacity of 76Whrs, which corresponds to 0.274MJ. Wiki claims a range of 0.9-2.63 MJ/L for li-ion, so picking a middleish value of 2 gives about 140cc

[CAT-THE-FIFTH]

You'd remove the need for a dGPU, but not much else. It would be really neat, but the size difference wouldn't amount to any real change in battery capacity or laptop size. Looking at the PCB in this teardown, I'd guesstimate you'd save an area about 75mm^2 or so. if that GPU and PCB come to 3mm thick that'd only be about 10-20cc, and it's an awful lot of expense for such a small volume. For comparison, I estimate the battery volume to be about 150cc[1], so an interposer would make very little difference to the battery you could fit inside - maybe 10% improvement?

[1]: The teardown reports a capacity of 76Whrs, which corresponds to 0.274MJ. Wiki claims a range of 0.9-2.63 MJ/L for li-ion, so picking a middleish value of 2 gives about 140cc

It would massively reduce the size of the PCB:

https://d3nevzfk7ii3be.cloudfront.ne...JjicDXj1a.huge





The second picture is of a Fury series card with an AMD Fiji chip. That is around 600MM2 with 4 HBM stacks.

You need to consider the Zen CPUs are an SOC with no graphics onboard.

I expect a Zen APU to be around 160MM2 to 250MM2 depending on the number of cores,so the interposer would be significantly smaller. It would also only need a single 8GB HBM2 stack at most,or two stacks if they want to offer 16GB system RAM.

That means the two chips on that PCB would be down to one,on the interposer and there would be no space needed for the GDDR5 and DDR4 soldered on the board.



They could probably shrink the PCB down to not much larger than the area in the black square. The MacBook has a weirdly shaped PCB as it has to accomodate two fans for both chips - I suspect you coul go back to using a normal shaped PCB too,and it would be cheaper too,since it would simply have much less logic onboard.

Then,you have the other consideration that the cooling system will be far less bulkier too.



The system needs to cool an Intel CPU AND an AMD GPU.

HBM2 also consumes less power than GDDR5- I expect you would need a simpler VRM section on the motherboard too,which would reduce space too.

Cutting the PCB and cooling system in half,would yield significantly more space for a battery. Plus with only one heat source to cool it should make for a less complex cooling system.

Now look at the 12" MacBook PCB:

https://d3nevzfk7ii3be.cloudfront.ne...lspAPXA6p.huge



That orange part is the memory chips. Imagine if that RAM was on the top package,that could easily be where the SSD is located.

HBM2 has so much bandwidth,it might even help the CPU perform better than normal too,and the bandwidth is massive. One stack of 1GHZ HBM2 would give 256GB/sec which is more than what an RX470 has,so they could easily downclock it to drop power consumption. AMD could also push 1024 shaders in the IGP but at a lower clockspeed.

But put that in context of a RX460 that only has 112GB/sec.

That would make it possible to have an UltraBook with dGPU level graphics performance.

[DanceswithUnix]

You'd remove the need for a dGPU, but not much else.

You remove the need for all those power hungry off-chip buffers on the ram and the APU, allowing that power budget to go into CPU cores and shaders, or in the race to the bottom cheaper cooling and a smaller battery.

AMD were supposed to be looking at this for HPC use according to leaks from a year ago, so I guess it is down to how modular they make their APU slices to go on the interposer. A part populated HPC interposer could still save space over discrete components, and the interposer is passive so no power requirement on an empty bit.

http://wccftech.com/amd-exascale-het...en-cores-hbm2/

Note that Intel does have a similar technology for embedding stacks of ram in-package for Altera FPGA use. Someone will do it.

[Xlucine]

It would massively reduce the size of the PCB:

https://d3nevzfk7ii3be.cloudfront.ne...JjicDXj1a.huge

[IMG:https://d3nevzfk7ii3be.cloudfront.net/igi/PiKnJNeJjicDXj1a.huge[/IMG]

[IMG:http://images.anandtech.com/doci/9421/PCB2.jpg[/IMG]

The second picture is of a Fury series card with an AMD Fiji chip. That is around 600MM2 with 4 HBM stacks.

You need to consider the Zen CPUs are an SOC with no graphics onboard.

I expect a Zen APU to be around 160MM2 to 250MM2 depending on the number of cores,so the interposer would be significantly smaller. It would also only need a single 8GB HBM2 stack at most,or two stacks if they want to offer 16GB system RAM.

That means the two chips on that PCB would be down to one,on the interposer and there would be no space needed for the GDDR5 and DDR4 soldered on the board.

[IMG:http://i.imgur.com/T917Keb.png[/IMG]

They could probably shrink the PCB down to not much larger than the area in the black square. The MacBook has a weirdly shaped PCB as it has to accomodate two fans for both chips - I suspect you coul go back to using a normal shaped PCB too,and it would be cheaper too,since it would simply have much less logic onboard.

Then,you have the other consideration that the cooling system will be far less bulkier too.

[IMG:https://d3nevzfk7ii3be.cloudfront.net/igi/cBFfrfQPrPBFgV1s.huge[/IMG]

The system needs to cool an Intel CPU AND an AMD GPU.

HBM2 also consumes less power than GDDR5- I expect you would need a simpler VRM section on the motherboard too,which would reduce space too.

Cutting the PCB and cooling system in half,would yield significantly more space for a battery. Plus with only one heat source to cool it should make for a less complex cooling system.

Now look at the 12" MacBook PCB:

https://d3nevzfk7ii3be.cloudfront.ne...lspAPXA6p.huge

[IMG:https://d3nevzfk7ii3be.cloudfront.net/igi/CXMhriTlspAPXA6p.huge[/IMG]

That orange part is the memory chips. Imagine if that RAM was on the top package,that could easily be where the SSD is located.

HBM2 has so much bandwidth,it might even help the CPU perform better than normal too,and the bandwidth is massive. One stack of 1GHZ HBM2 would give 256GB/sec which is more than what an RX470 has,so they could easily downclock it to drop power consumption. AMD could also push 1024 shaders in the IGP but at a lower clockspeed.

But put that in context of a RX460 that only has 112GB/sec.

That would make it possible to have an UltraBook with dGPU level graphics performance.

The problem is that the pcb isn't the main driver of laptop volume, the battery is. Using the measurement tools in adobe pdf, and setting the scale so that that pcb was 340mm wide (total MBP width is 350mm, I eyeball that to be about 1cm thinner), gives a total pcb area of 230cm^2. Staying with a 3mm PCB & chip width*, that's only 70cc - half the size of the battery. So if the interposer lets you shrink the pcb by 50%, that's only a 25% improvement in battery life. Whether that's worthwhile depends on the cost of the interposer vs the cost of including the tracks for RAM & VRAM, and I'd be surprised if that worked out in the interposer's favour.

The cooling system wouldn't be any different. Currently the 15" MBP has a single fat heatpipe connecting the CPU, GPU and the two fans - with CPU & GPU next to each other or sharing a die there'd be no reason for this to change. The chips would still produce similar amounts of heat, so you'd need similar amounts of fin surface area. While HBM does produce less heat than GDDR5, the GDDR5 is currently not connected to the heatpipe so this wouldn't change the cooling system. Extra memory bandwidth won't let you get away with downclocking the GPU - if a task is shader limited, memory bandwidth won't help. For a similar reason, limited memory bandwidth doesn't force the designer to run the GPU faster - if memory bandwidth is a limit, then more shaders won't help. More memory bandwidth also won't help a modern cpu: http://hexus.net/tech/reviews/ram/97...-32gvr/?page=4 As hexus found, 25% higher memory bandwidth only corresponds to 2% better performance in CPU benchmarks

(Image links borked on purpose for ease of scrolling)
*I'm really pleased with how accurate the 3mm guess turned out. I got around to applying my micrometer to a wifi card I had lying around, and it measured 3.01 +/- 0.005mm!

[chinf]

Staying with a 3mm PCB & chip width*, that's only 70cc

Perhaps I'm limited by access to 2007-era technology, but isn't the volume within the case rather larger than just the PCB & chip width? I don't think notebooks are constructed quite the same as smartphones (yet). I'm willing to take a donation of a recent laptop to be proved wrong though

[DaMoot]

Are those actual, real, 8 logical cores? Or mini cores?

[kalniel]

Are those actual, real, 8 logical cores? Or mini cores?

I think most people would say they're 8 real cores, but I'm not sure how you define a 'mini core'.

[DanceswithUnix]

Are those actual, real, 8 logical cores? Or mini cores?

They are 8 full blown cores, about as powerful as you can get so not mini by any measure.

Are you thinking of the PS4/XBOne which have 8 mobile cores? Those cores are quite small, and completely different.

[DaMoot]

Are those actual, real, 8 logical cores? Or mini cores?

I think most people would say they're 8 real cores, but I'm not sure how you define a 'mini core'.

AMD has gotten itself in hot water for a couple years now over using mini cores (not full logical cores) and lumping those numbers in with the total core count. People were quite upset to find that their octocore was only really a dual core.

AMD's mini cores were/are things like FPUs or associative math units, meaning that the amount of work performed by discrete, independent cores was much lower in reality than what they marketed. In contrast with Intel which uses only full, independent logical cores that can all operate independently.

If this new architecture is going to be full cores, I'm quietly rooting for AMD again. Maybe they can claw their way back from the jaws of death! Come on, AMD, do another K6-2, Athlon, or Duron-type release! Fast, cheap, honest and embarrassing to the competition!

[DaMoot]

They are 8 full blown cores, about as powerful as you can get so not mini by any measure.

Are you thinking of the PS4/XBOne which have 8 mobile cores? Those cores are quite small, and completely different.

Not physical core size. I was asking/commenting regarding logical cores.

[kompukare]

AMD has gotten itself in hot water for a couple years now over using mini cores (not full logical cores) and lumping those numbers in with the total core count. People were quite upset to find that their octocore was only really a dual core.

A bit hyperbole, surely?
The Bulldozer derived CPUs share their FPUs in a CMT design*.
*https://en.wikipedia.org/wiki/Bulldo...hitecture)#CMT
Basically, they are based on modules, where each two integer cores, shares one FPU. Intel's HT (HyperThreading) is called SMT:
https://en.wikipedia.org/wiki/Simult...multithreading
and it only shares some parts (duplicate registers etc.).
Actually, CMT should/does have better scaling than SMT/HT.
The stats are something along this line: HT vs non-HT (i5 vs i7 for example) gains around 15%. CMT vs non-CMT (running more than one thread per module.gains around 30% or more.
However, since AMD's Bulldozer design has far worse IPS than Intel's Core design, it usually doesn't matter. In the end while it scaled well, it looks like it was rather wasteful of die space.

Anyway, Zen does away with all that, and instead does a SMT type design. We will have to wait for full independent testing, but compared to Bulldozer it seems AMD have a far better design which seems to be able to compete with Intel without being a vastly larger die. As has been pretty evident recently, AMD could really use a profitable CPU line to hire more people as a lot of their recent stuff seems to have been delayed (probably) due to not having enough manpower to validate and finish multiple lines at the same time even when the basic design was done.