AMD - Zen chitchat

[scaryjim]

Interestingly everyone seems to be assuming that the 2C/3CU part is a fused off version of the full die rather than a separate piece of silicon. That would seem like an odd decision, unless they've got a lot of really buggy dies lying around in a warehouse somewhere....

EDIT: the official AMD release says the 2C variant has 5MB of L2+L3 cache, which would imply a full complement of 4MB of L3 cache. I'm not convinced they'd invest in that much cache for a separate 2C die, so maybe these are salvage parts...

EDIT 2: Aha! You can't find this by searching on the AMD site, but the Ryzen 3 2200U page is live if you know how to construct the url: https://www.amd.com/en/products/apu/amd-ryzen-3-2200u

Sure enough, it has 4MB of L3 cache and a dual channel memory controller, neither of which I'd expect to see on a dedicated 2C die, so I'm pretty sure these are just Raven Ridge salvage. I'm a bit less excited now...

[watercooled]

Given they've used one die for HEDT, workstation and server CPUs, and the only other Ryzen die for all APU purposes, it would be unusually costly to have one dedicated 2C die - stripping two cores likely wouldn't save enough die space to make it worthwhile unless it was a massive market.

[scaryjim]

It's not just two cores though; it's 72% of the IGP shaders.

The last generation of construction cores did split silicon, with a 2 core/3 CU parts targeted at the entry level market and a 4C/8CU part for the mid-range. Drop to half the cores, 3/8 of the shaders and half the memory channels and you're going to save well over half the silicon area - and therefore you're going to half the cost as well. And for Raven Ridge, that goes up to 3/11 of the IGP - an even bigger silicon saving.

The only reason to salvage a part this hard is if you'd otherwise have to bin them, which implies that there's a real problem yielding the IGP on Raven Ridge. There's more than twice as many disabled shaders as active ones. That's .... odd, frankly.

[DanceswithUnix]

The only reason to salvage a part this hard is if you'd otherwise have to bin them, which implies that there's a real problem yielding the IGP on Raven Ridge. There's more than twice as many disabled shaders as active ones. That's .... odd, frankly.

Creating a new die costs millions. So if they save $10 at the foundry per die by making a smaller die, they have to sell a million parts to make that back vs just disabling some blocks.

[watercooled]

It's not just two cores though; it's 72% of the IGP shaders.

The last generation of construction cores did split silicon, with a 2 core/3 CU parts targeted at the entry level market and a 4C/8CU part for the mid-range. Drop to half the cores, 3/8 of the shaders and half the memory channels and you're going to save well over half the silicon area - and therefore you're going to half the cost as well. And for Raven Ridge, that goes up to 3/11 of the IGP - an even bigger silicon saving.

The only reason to salvage a part this hard is if you'd otherwise have to bin them, which implies that there's a real problem yielding the IGP on Raven Ridge. There's more than twice as many disabled shaders as active ones. That's .... odd, frankly.

That was on 28nm - development and mask costs are massively higher on 14nm don't forget. I wasn't aware of how much they'd stripped back the IGP shaders TBH but perhaps they're just not expecting to sell all that many low-end parts? Given Summit Ridge's performance and that people have been apparently getting free 'unlocked' cores, I don't think there's any obvious indication yield is poor.

Aside from that, provided it's cost-effective, having just the one die in production should be good for inventory management - they can meet pretty much any market demand by just binning dies differently.

I do agree it's an unusually heavy binning though...

Edit: That 2200U is the extreme end of the scale, the next one up the 2300U has 4 cores and 6 CPU threads, so a 2C, stripped GPU die wouldn't make sense for one SKU.

AMD typically haven't churned out die variants as aggressively as Intel, and a stripped-back die could backfire if e.g. a competitor were to up the core count of the entry-level stuff in response. You can be almost certain AMD considered it.

[scaryjim]

... Edit: That 2200U is the extreme end of the scale, the next one up the 2300U has 4 cores and 6 CPU threads, so a 2C, stripped GPU die wouldn't make sense for one SKU.

I agree it wouldn't make sense if the only market was consumer entry-level laptops, but a lot of the cat core APUs (and I suspect Stoney Ridge too) went into embedded and industrial systems (not to mention that there were many Stoney Ridge SKUs under the E2, A6 and A9 brandings). It's not just down to cost; smaller silicon with a smaller (or no) L3 cache and a single memory channel should also draw even less power, offering the possibility of getting Zen down to tablet-capable TDPs, at which point having less silicon area is also beneficial. So there's good reasons to have that smaller die. Plus it was on the industrial roadmap 2 years ago, and while they may have had a change of heart it'd be a bit of a surprise.

All that said, it's also a surprise that they would cut down Raven Ridge that far. Perhaps the thought is if you have to disable at least one core then you might as well gut it, because a dual core isn't a good match for a higher CU count? Would seem odd though given most people would be happy to pair a Haswell i3 with an entry level discrete card, and that the performance level Ryzen APUs offer...

[DanceswithUnix]

If you are cutting down to tablet TDP, you want all those PCIe lanes gone as well as most storage options and anything else with a high speed SERDES talking to the outside world. With all that gone you could cut down the number of bumps connecting to the outside world to produce a smaller package.

I'm not seeing a big pull for a cut down AM4, and certainly wouldn't want to give the likes of HP an excuse for making single channel RAM laptops any time soon.

[watercooled]

The embedded silicon is an interesting angle though - I wonder how they'll address that moving forward? AFAIK there's nothing about any future cat cores?

[Ozaron]

Edit: That 2200U is the extreme end of the scale, the next one up the 2300U has 4 cores and 6 CPU threads, so a 2C, stripped GPU die wouldn't make sense for one SKU.

Do you mean 4 cores 4 threads? It has 6 "GPU cores". The thought of half CPU having threading and half not got me googling.. here

[scaryjim]

If you are cutting down to tablet TDP, you want all those PCIe lanes gone as well as most storage options and anything else with a high speed SERDES talking to the outside world. With all that gone you could cut down the number of bumps connecting to the outside world to produce a smaller package.

The speculative industrial roadmap (https://videocardz.com/69428/amd-sno...eat-horned-owl) lists a lot of IO options, but how much of that was meant to be usable simultaneously is questionable. That said, we still don't know how many PCIe lanes Raven Ridge has - it could easily have been cut down to 16, giving x8 for graphics, x4 for chipset, and x4 for storage/USB. I suspect a 2C die would be intended to run without a chipset at all, so a smattering of USB2 and 3, one or two storage ports and a couple of spare PCIe lanes would probably do. Maybe an x8 complex at most.

Curiously those slides list the APUs as having shared L2 cache; I'm pretty sure that wasn't implemented on Raven Ridge.

... I'm not seeing a big pull for a cut down AM4, and certainly wouldn't want to give the likes of HP an excuse for making single channel RAM laptops any time soon.

I certainly don't think the cut down die would ever make it to a socket or desktop (unless they decided to revive the AM1 platform). Embedded and mobile would be the target. As to dingle channel laptops, it's not like HP need an excuse to do that, so why not give them a single channel die anyway

The embedded silicon is an interesting angle though - I wonder how they'll address that moving forward? AFAIK there's nothing about any future cat cores?

No, Zen will replace the cat cores in the low end space. The target TDP for the 2C die is 4W - 15W, nicely covering the space that the cat cores used to play in. Given the clocks we've already seen from Raven Ridge, I imagine the 2C die will clock somewhere between 1.5GHz and 2.5GHz to fill that space. With the significant IPC advantage it has, it should give much better performance in the same power budget.

[CAT-THE-FIFTH]

I wonder if we will get an Athlon II X4 version of the 2200G and 2400G?? If the 2200G is around £90 to £100,that means we should in theory have an overclockable quad core version for under £90 - its going to make the Intel Pentium CPUs look a tad underspecced IMHO OFC!

[watercooled]

Do you mean 4 cores 4 threads? It has 6 "GPU cores". The thought of half CPU having threading and half not got me googling.. here

It was a typo sorry, I meant 4 core, 6 GPU core.

[watercooled]

No, Zen will replace the cat cores in the low end space. The target TDP for the 2C die is 4W - 15W, nicely covering the space that the cat cores used to play in. Given the clocks we've already seen from Raven Ridge, I imagine the 2C die will clock somewhere between 1.5GHz and 2.5GHz to fill that space. With the significant IPC advantage it has, it should give much better performance in the same power budget.

Yeah that's what I was getting at, there must be *something* Zen-based planned for the embedded market given there's apparently no future cat cores.

[scaryjim]

I wonder if we will get an Athlon II X4 version of the 2200G and 2400G?? If the 2200G is around £90 to £100,that means we should in theory have an overclockable quad core version for under £90 - its going to make the Intel Pentium CPUs look a tad underspecced IMHO OFC!

Given the price they're currently listed at, and the fact that the full die is being cut all the way down to 2C/3CU for the Ryzen Mobile, I'd be surprised if they chopped the IGP off to make Athlon-style processors. Not impossible though; I wonder how much TDP you could save by fusing the IGP off in its entirety? We might see Ryzen 5 2400E and Ryzen 3 2200E with a 35W TDP at similar clock speeds to the old 1200 and 1400...?

[ET3D]

I wonder if we will get an Athlon II X4 version of the 2200G and 2400G?? If the 2200G is around £90 to £100,that means we should in theory have an overclockable quad core version for under £90 - its going to make the Intel Pentium CPUs look a tad underspecced IMHO OFC!

If the 2200G is enough to take a lot of low end market share from Intel, I'd expect a Pentium variant of the i3 8100. That would probably pressure AMD into releasing a Ryzen 1 1100 or whatever. Nice to have competition.

[DanceswithUnix]

I wonder if we will get an Athlon II X4 version of the 2200G and 2400G?? If the 2200G is around £90 to £100,that means we should in theory have an overclockable quad core version for under £90 - its going to make the Intel Pentium CPUs look a tad underspecced IMHO OFC!

That would be odd.

Die size of Summit Ridge/Zepplin is 192 mm^2, die size of Raven Ridge APU is 210 mm^2.

So it would be cheaper to heavily fuse off the old Ryzen cpu. ThreadTickler?