AMD Ryzen

[Keyrock]

I'm sure Intel will recover, they have way too much money and way too many talented people.  We've seen them pick themselves up from blunders before.  I am reminded of the Pentium 4 space heater fiasco where Intel, almost blindly, chased the all mighty clock speed and ran face first into the 4 GHz thermal brick wall (fittingly their recent flirtations with 5GHz seem eerily familiar).  Intel went back to the drawingboard and developed a design that emphasized IPC.  Granted, at that time, Intel was able to use FUD and strongarm tactics to maintain a stranglehold on the market despite having an inferior product, until they brought a superior product to the market.  This time, while I have no doubt Intel will again resort to FUD and strongarm tactics, I doubt those tactics will be anywhere near as successful and Intel will almost certainly lose a decent chunk of market to AMD.  This, if it happens, will be good news for consumers, as AMD competing at a closer to even level with Intel should mean innovation and price wars.

[Zoraptor]

It's been years since the 'xx nm' really meant something. You'll find very few to no structural features on current chips which actually fall into the range of some tens of nm. Now they mostly just go to smaller numbers to indicate every kind progress in manufacturing.

 

Yeah, when it comes right down to it there's a lot of stuff on a processor which simply doesn't scale any more. They are still getting decent density improvements on the new processes at least.

 

I'm curious what Patents AMD currently has in play with Zen, because AMD might be able to gain a vicegrip on their advantages. Intel's future might lie in the courts at this point. I guess AMD also being a big player in the GPU market marks a secondary blow to Intel who has spun their wheels with their integrated graphics.

 

Intel have equivalents to ccx structure and infinity fabric ('mesh') as some sort of infinity fabric like approach is pretty much essential for large core counts due to ringbus interfaces starting to choke latency wise as core counts go up. Intel and AMD have a cross licensing agreement anyway which I think covers pretty much everything and is kind of essential with AMD owning x64 and Intel owning x86. As keyrock says, the big advantage of the ccx approach is scalability so Intel's consumer chips having ringbus is fine when they were 4 core sky/kabylake but struggles when you start getting to 8 core whisky(?)lake. The ccx approach also is way more efficient for manufacturing, AMD has a low error rate anyway with Zen but with ccx you can just plug combinations together to get anything from 'super' Epyc/ Threadripper chips right down to an r3 1200; and 'defective' ccx can also (mostly) still be plugged together and sold.

 

Intel's specific move into graphics is probably more about the top end machine learning type stuff than consumer and iGPUs- a lot of that used Intel CPUs originally but has shifted to graphics cards especially nVidia as they are better suited to it. Intel actually have at least one gen of iGPU that haven't been released yet due to the cannonlake problems, but fundamentally most people who buy and use an iGPU just want anything that will output Excel or whatever, if they care about games or workstation stuff they have a dedicated graphics card.

[Keyrock]

with AMD owning x64

 

Kind of ironic as Intel were the first to 64-bit, but their Itanium platform, while great at executing any 64-bit code, was EXTREMELY slow at executing 32-bit code, which it had to emulate.  Itanium would probably be pretty great now that pretty much everything has native 64-bit code (in fact many programs do not have 32-bit code).

Edited June 23, 2018 by Keyrock

[injurai]

Pretty sure most CPUs emulate the 32-bit instruction sets anyways, most of instruction processing is just decoding and decomposing instructions into "risc" equivalents anyways. The standard methods are typically direct implementations of the cutting edge, and supporting legacy through hardware emulation layers.

[MrMoe]

Pretty sure most CPUs emulate the 32-bit instruction sets anyways,

 

The Intel64 architecture emulates 32bit code, but the AMD64 runs it naively and pretty much with very little or no performance loss.

 

https://en.wikipedia.org/wiki/X86-64#Operating_modes

[Zoraptor]

The latest rumour for Zen2 is 8 cores/ ccx, 10-15% (!) IPC gain and 5GHz clocks. I'm... skeptical. If true it would be an utter disaster for Intel as their clock and slight IPC lead would be gone at a stroke, and they'd be potentially competing with threadripper class chips at consumer prices; but I suspect 6 cores/ccx, 5% IPC and up to 5GHz clocks would be more likely given the available information on TSMC's 7nm (with GloFo's being a bit behind both time and process spec wise), but that would still be better and cheaper than Intel's offerings until they can offer maybe 10nm++ which could be years away.

[injurai]

AMD were the ones pushing the 5GHz before that trend broke, if they are better positioned to play at the thermal edge all ontop of their recent fab/arch gains, then yeah Intel should fear.

 

Still I think Intel has the prototypes to win, just their production pipeline has probably been caught a bit behind where they otherwise need to be.

[Zoraptor]

They have some good potential chips on paper, but with 14nm being in its 5th year and with the only way to improve it seemingly being to add cores Intel desperately needs their 10nm process working properly, and by their own admission it's likely to take 2 revisions to get practical performance above their current optimised 14nm process- and with their competition now being AMD and TSMC/ GloFo 7nm and not primarily their own chips that's a major problem. The 10nm fiasco is also gumming up any IPC improvements they may have from improvements to the Core architecture, since they haven't (yet and probably won't with 9 series if the rumours are true and it is just more cores at identical clocks and IPC) or cannot bring them to the working 14nm process.

 

But yeah, the medium term positive for Intel is that when they do get 10nm working they may well get a burst of IPC improvements along with the increased density/ cores/ potential clock speeds/ energy efficiency from the smaller process.

[injurai]

What's the relationship between AMD and TSMC/GloFo that allows them to be ready for 7nm and Intel not?

[Zoraptor]

AMD used to own Global Foundries but spun it off, and they used to have an exclusive contract with them but bought out of it a couple of years ago. So AMD outsources its production and can use TSMC/ GloFo or anyone else's compatible fabs if needed as demand, cost and performance dictates. Intel's production is in house so they have and exclusively use their own processes; historically this has been an advantage as they have typically been ~2 years ahead of the competition tech wise. 10nm was due well before the (equivalent, despite the size difference in the name) TSMC/ GloFo 7nm process but still hasn't really arrived in any meaningful way, so the usual situation is reversed.

 

Theoretically Intel could outsource production too, but that would be crippling to morale/ prestige and require some significant redesigns. The technicalities of which are well beyond my expertise though.

[Keyrock]

The latest rumour for Zen2 is 8 cores/ ccx, 10-15% (!) IPC gain and 5GHz clocks. I'm... skeptical. If true it would be an utter disaster for Intel as their clock and slight IPC lead would be gone at a stroke, and they'd be potentially competing with threadripper class chips at consumer prices; but I suspect 6 cores/ccx, 5% IPC and up to 5GHz clocks would be more likely given the available information on TSMC's 7nm (with GloFo's being a bit behind both time and process spec wise), but that would still be better and cheaper than Intel's offerings until they can offer maybe 10nm++ which could be years away.

8 cores is pretty mucha given as the current top of the line Ryzen chips are already 8 cores/16 threads. As for the increase in IPC, 10-15% is optimistic but certainly not unreasonable. If true, Intel is set to lose more ground until they can bring a true new architecture to the table.The Core architecture is really long in the tooth.

 

As Zoraptor mentioned, Intel's in-house chip manufacturing facilities have been their main advantage (other than strong arm tactics, FUD, and shady anti-competitive deals that stifled the competition's ability to bring products to the table) for decades. Intel has been 2 to 3 years ahead of AMD on almost every die shrink and process improvement... Until now. 10nm has been an unmitigated disaster for them and it's now many years behind schedule and AMD is set to actually beat them to the punch on a die shrink.

 

This is the worst position Intel has been in since the Pentium 4 space heater fiasco.

Edited July 28, 2018 by Keyrock

[kirottu]

First Look: AMD Threadripper 2000 Series

[M4xw0lf]

Tests of the TR 2950X (16c/32t) and 2990WX (32c/64t) will be out next monday, but of course the improvements we're going to see are clear from comparing the 1800X and 2700X (at least for the 2950X - the 2990WX comes with the small boon of double the cores ^^).

Edited August 6, 2018 by M4xw0lf

[injurai]

I wonder if we'll see anything at QuakeCon

[Keyrock]

The Threadripper 2 benchmarks are out and these things are monsters.  The 2950X and 2990WX OBLITERATE anything Intel has on offer right now, and the 2950X does it at a much lower price point than Intel's top offering in the category while still mopping the floor with it.  For us gamers, though, these chips are overkill and not a value proposition as they offer few, if any, benefits for gaming at a much higher price point than consumer grade chips.  The best value proposition for gamers is the 2920X.  The 4 extra cores over AMD's consumer grade top offering isn't a big deal for gaming, but what is nice is that you still get the 64 lanes of PCIE 3.0, like its more expensive cousins, and at a price point that's not as ridiculously higher than consumer grade chips.  For big production workloads, encoding, stuff like that, with properly extremely high-threaded workloads, though, the high-end Threadripper 2 chips are monsters that wipe the floor with anything Intel has out now in the same category, and, quite frankly, I don't think Intel has anything they can put out at that price point that can even attempt to compete with the 2990WX in the next couple years, until they develop a new architecture that can scale like AMD's Infinity Fabric or better.

[Zoraptor]

No 7nm from Global Foundries now, so Zen 2- and the 7nm graphics chips- will be exclusive to TSMC. Bit of a shame as contrary to what I said above GloFo's 7nm was actually tracking a bit better than TSMC's performance wise but apparently was looking to be a lot more costly and in the end too costly.

[Malcador]

That is going to hurt IBM a lot, apparently.

[Zoraptor]

Yeah, might mean no Power 7nm chips for IBM. GloFo's 7nm was largely based on IBM's own 7nm work.

[Zoraptor]

Intel has launched their new i9- for the low cost of one kidney*. Wouldn't normally be relevant for here, but...

 

 
Not like the other two points are exactly stellar either. 9570 was as mainstream as the 9900k, and soldering is returning something that should never have gone from the top end.

 

*here you could buy 2x 2700x or a threadripper and MB (albeit a cheap one) for the same price.

[injurai]

Imo, AMD's pricing encourages more frequent upgrades. Which means it would be easier for AMD to move their top-end stock. Intel's recent models are more about brokering SKU's for dividend reasons. If they were cheaper I'd be inclined to humor them more.

 

Right now Intel's release cycle is in a sticky situation with their Process-Architecture-Optimization. New fabs are great and all but what you really want is the new arch. New archs can be flawed so you really want the improvement. But it's the new fabs where you see some huge jumps in transistor count. So where is the best time to purchase? All have their own risks. The major thing is whenever you do upgrade, it makes sense to have gone through a whole cycle, so if you bought on arch last time, upgrade on arch again. But this is an awful way to have to shop as a consumer. The price forces you to wait at least two/thirds of a cycle and it's not worth it until the 3rd or beyond. Then even when you do upgrade pricing is a racket because the market is flooded with sub-par chips.

[Zoraptor]

AMD certainly offers a more value for money long term upgrade proposition with Zen 2 still coming, plus their aggressiveness of pricing is way beyond Intel's in general. Counterintuitively though at the moment Intel has probably encouraged more upgrades- if you bought a mid range or above Ryzen at launch (1600 or greater) there's little reason to upgrade to either new Intel or AMD offerings (yet), if you own a year+ old i5 you're likely to already be getting limits from their 4 threads, and even if you bought an i7 you could now get double the cores/ threads from AMD or Intel.

 

But yeah, Intel's 10nm problems have resulted in 4 'generations' of Skylake with no IPC improvement. All the IPC improvements were set for 10nm designs which still aren't able to be produced.

[Zoraptor]

Ryzen 3000 series (and Navi) rumours...

 

 

Only a few things I can see there that look a bit suspicious. It seems unlikely AMD would go for both 6 and 8 core ccxs since the beauty of their current design is that basically everything from Epyc to Ryzen 3 uses the same 4 core/ ccx design which makes it both incredibly efficient in terms of (lack of) wastage while being the epitome of KISS design. Then again two complex types makes things more flexible as well, so maybe. Also the integrated graphics would be likely to hit performance limits from using system RAM under current designs though maybe cache can ameliorate it enough. And the supposed Navi improvements are at the top end of expectations.

 

OTOH, calling their Navi cards RX 3060, RX 3070 etc is 100% the sort of thing AMD would do after their chipset namings for Ryzen, and up to 16c/32t 5+ Ghz Ryzen 3850X at 500USD with IPC advantage over Intel is exactly the sort of thing I absolutely want to be true.

[Jozape]

I don't think the 6 cores and 8 cores per chiplet means that is how many cores there are per CCX. Each chiplet would surely have 1 or 2 CCXs (8 or 4 cores per CCX respectively), and then 2 cores are disabled to get salvaged chiplets with 6 cores. Just like current Ryzens.

[Zoraptor]

4 cores/ ccx makes most sense since it's simplest and has fewest intra ccx connections and 8 would have a lot, but unless they planned for it back when they designed AM4 I'd question if they could go to a full chiplet design- with 2 4 core ccx- while staying with AM4. If they did plan for that I would be very impressed, and AMD did put a lot of stuff into the CPU package which would usually go into the chipset on the motherboard. That would also give an easy way to get 12 cores, threadripper had dummy ccx in gen 1 so you could get 12 core with 3x4 and one dummy.

 

I guess if they went the 8 core/ ccx route they'd have more failed chips to pad out the lower grade SKUs as well as 7nm probably being a bit less reliable as well; my skepticism comes from the extremely low relative failure rate for Zen1. We got a lot of 8 core 1600s here due to them running out of 'bad'/ partially failed chips.

Edited December 8, 2018 by Zoraptor

[Jozape]

4 cores/ ccx makes most sense since it's simplest and has fewest intra ccx connections and 8 would have a lot, but unless they planned for it back when they designed AM4 I'd question if they could go to a full chiplet design- with 2 4 core ccx- while staying with AM4. If they did plan for that I would be very impressed, and AMD did put a lot of stuff into the CPU package which would usually go into the chipset on the motherboard. That would also give an easy way to get 12 cores, threadripper had dummy ccx in gen 1 so you could get 12 core with 3x4 and one dummy.

 

I guess if they went the 8 core/ ccx route they'd have more failed chips to pad out the lower grade SKUs as well as 7nm probably being a bit less reliable as well; my skepticism comes from the extremely low relative failure rate for Zen1. We got a lot of 8 core 1600s here due to them running out of 'bad'/ partially failed chips.

Well, Intel is already offering 8-core CPUs that don't employ the CCX concept. It is more complex so probably more expensive to implement, but I don't think it would affect yields much. Plus the 7nm chiplets are approximately 1/3 the area of the 14nm chips, so that will offset much of the loss in yields of an immature 7nm process.

 

That said Voldemort has retracted his claim on the IO chip; Now he is claiming that these Ryzen 3 products use 7nm chiplets only. More likely that they have slightly larger 7nm chips with IO integrated for PCs, at maybe 1/2 the area of 14nm chips each. This also makes more sense with the combination with Navi which I think would have IO integrated because of bandwidth and energy requirements. Infinity Fabric 2 will do what, 100 GB/s between chips? That may be enough for an integrated GPU, but not for any discrete GPU nowadays.