Speculation: Spring refresh for Ryzen

[Vattila]

As I've posted elsewhere, I think AMD fell short of their frequency targets for the Ryzen 3000 series. A later launch than expected, statements relating to frequency challenges, Max Boost redefinition, boost issues, 3900X shortages, and now the 3950X delay from September to November, do all corroborate this. I guess they aimed for 5 GHz, but fell ~8% short.

However, TSMC is firing on all cylinders. Amongst a number of process roadmap announcements, they now offer N7P as a refinement of their N7 process on which the Zen 2 chiplet is built. N7P is compatible with N7 design rules, so it should provide a fairly simple and cost-effective opportunity to optimise existing designs.

With that in mind, is there a chance that AMD will do a spring refresh based on a faster stepping of the CPU chiplet?

 

[Yotsugi]

We need a meme with a banknote with Lisa's face on it

It's called $AMD share.

 

[NostaSeronx]

Are you saying/implying that there will be no replacement for the Ryzen 3xxx series in 2020?

I wouldn't know, but if there is a Zen2+, why wouldn't Ryzen prefer a Zen3+(6nm+/5l-euv) over Zen3(7nm+/4l-euv). An EPYC-focus and Instinct-focus will not lead AMD to the pit of mobile-only of 2012 to 2016.

So you are thinking a Cortex A72 -> A73 type situation, slight IPC loss with significant perf/watt gain?

Height is to frequency, as width is to IPC. Where frequency is lost, IPC can get a huge boost. With the race for density, IPC-focused architecture is preferred over Freq-focused.

Example: Zen2 FPU does 2x256-bit FMA, Zen3 FPU can probably also do 2x256-bit FMA. However, IPC-driven architecture would also allow for 4x128-bit FMA on Zen3. 128-bit/256-bit in Zen3 wouldn't affect power, since total of 128-bit flops and total of 256-bit flops are the same.

 

[naukkis]

Example: Zen2 FPU does 2x256-bit FMA, Zen3 FPU can probably also do 2x256-bit FMA. However, IPC-driven architecture would also allow for 4x128-bit FMA on Zen3. 128-bit/256-bit in Zen3 wouldn't affect power, since total of 128-bit flops and total of 256-bit flops are the same.

So you are saying that whole SIMD is useless as combining things to wider arrays to be executed with single instruction instead of multiple ones won't save any power at all. I don't think so.....

 

[Zucker2k]

"5GHz": Round 2. LOL

Sorry, I couldn't help it. I do think AMD has a better chance of hitting 5GHz with 6 and/or 8 cores than 16, simply because the latter will take double, if not more, effort (power, heat, chiplet resources) just so they can award that elusive distinction to the flagship desktop processor. They need to learn from Intel on this; give the 8 core chip the fastest chiplets (at all costs) because the attractiveness of the 16 core is in its core count, value, and performance.

 

[Vattila]

I do think AMD has a better chance of hitting 5GHz with 6 and/or 8 cores than 16

With AMD's Precision Boost algorithm and chiplet design, this is no longer the case. As they now state, Max Boost Clock is the maximum momentary speed a single core can reach while running a "bursty single-threaded workload". For example, for the 12-core 3900X, which houses two salvaged 6-core CPU chiplets, they just have to find a single chiplet that can reach 4.6 GHz on one core.

The problem is that AMD has difficulty reaching the 4.7 GHz needed for 3850X (hence the delay). Obviously, seeing the scarcity of 3900X, a very low proportion of the manufactured chiplets is capable of reaching these speeds. The absolute quantities will of course improve as AMD ramps production volume, but it seems likely that a refined stepping is needed to comfortably bin chiplets beyond the current top speeds.

 

[Insert_Nickname]

The problem is that AMD has problems reaching the 4.7 GHz needed for 3850X (hence the delay). Obviously, seeing the scarcity of 3900X, a very low proportion of the manufactured chiplets is capable of reaching these speeds. The absolute quantities will of course improve as AMD ramps production volume, but it seems likely that a refined stepping is needed to comfortably bin chiplets beyond the current top speeds.

There are perhaps a very few golden sample chiplets that could reach 5GHz, but perhaps too few for an actual product based on them. They also might just hang onto them for TR3. We'll see.

 

[Vattila]

the attractiveness of the 16 core is in its core count, value, and performance.

Good point. And, if the rumour of the 3900 is true —12C/24T at 65 W and lower clocks — it looks like AMD will fill out their line-up with SKUs like this over time. However, the flagship SKUs of course do benefit from being able to offer great Max Boost as well, allowing AMD to beat the "no compromises" drum.

 

[moinmoin]

Plus, the performance could be really good, thought they might want to implement an "AVX2 offset" like Intel does, in that case, just to keep power limits in check for "heavy" FPU workloads.

Just as static turbo tables static offsets are also a thing of the past. PB2 handles all that dynamically already.

Honestly I expected more of a backlash for AMD already having introduced a non-tweakable offset essentially through the backdoor using PB2, with instances where frequency can be pushed below the rated base frequency with heavy FPU workloads.

 

[tomatosummit]

First I've read on N7p here, but it looks like it's just an extra year of manufacturing refinement and tweaks for N7. As N6 looks like it's coming later than N7+ then N7p is much more likely to be used for a refresh part. It's not all about clock speeds though, obviously the ghz are a sore point for amd right now with intel's counter campaign hitting them hard enough to actually demand a response (I hope it's not just another modern company being overly concerned about what twitter posts etc).
There are other thing to upgrade on ryzen such as the IO die, getting IF over 2ghz for memory clocks, on package memory. Previous rumours point towards r4000 having 7nm io for example

Good point. And, if the rumour of the 3900 is true —12C/24T at 65 W and lower clocks — it looks like AMD will fill out their line-up with SKUs like this over time. However, the flagship SKUs of course do benefit from being able to offer great Max Boost as well, allowing AMD to beat the "no compromises" drum.

Trickling out the lower tdp parts is what intel has been doing for a few years now, normally due to supply, see the 8700k and 9900k launch availability issues where they launched the top end part with low availability and the defunct 8400/9400 dies then released more skus as time goes on.
I wonder how that relates to the supply problems they have currently. A lower clocked 12/16 core cpu at launch would have sold gangbusters just like the 1700 did, 4,2ghz turbo is easy to achieve, the performance is still ample at stock and "easily" overclocked. If the supply was really down to clock speeds a slower sku would be an easy fix unless the ryzen department is totally on the "gaming-ghz vs intel and nothing else matters boat". It makes me think there's another bottleneck somewhere, epyc demand or even somewhere in the assembly process. They had to develop a new packaging technique for r3000 and who knows the logistics behind that and maybe even epyc is taking up those resources too. IO die manufacture is another unknown as well.

 

[Yotsugi]

Previous rumours point towards r4000 having 7nm io for example

Big, fat, guaranteed no.

 

[BTRY B 529th FA BN]

I'd like to see a higher baseline clock. A baseline clock that beats Intel at their turbo clocks in all benchmarks.

 

[beginner99]

No refresh and straight to zen3 / ryzen4000. As has been the plan on roadmaps for years. Refresh only makes sense if zen3 would be delayed a lot. Also I thought 6nm was the direct sucessor to 7nm and 7nm+ actually requires too many chances from 7nm to make it worthwhile since this would be consumer-only refresh.

All they need to do, as far as I understand, is to set a sufficiently low base clock. The Precision Boost algorithm will handle the rest, adjusting frequency to keep within TDP and temperature limits.

I'm with larry here. An AVX offset would help a lot with manual OC. Currently with manual OC, you have to limit yourself to a frequency that works with AVX code which can be couple 100 mhz lower.
Setting that offset might also help in normal situations, if 1 or more cores are running AVX code and others are not, you could shift priority to non-avx code changing avx offset. (less heat from cores running avx will give more headrom to other cores).

 

[Yotsugi]

An AVX offset would help a lot with manual OC

Manual OC is dead.
Welcome to the bleeding foundry edge.

 

[beginner99]

Manual OC is dead.
Welcome to the bleeding foundry edge.

I don't disagree but even in non-oc scenarios it could help with heat and noise or actually improve performance. I mean even x265 only uses AVX at some areas. While encoding you will have some cores running AVX code and others not. Lowering AVX offset will reduce power use of that AVX code giving more headroom to non-avx part. theoretically this could increase performance. Theoretically.

 

[Atari2600]

Manual OC is dead.
Welcome to the bleeding foundry edge.

Yep - there is virtually no good reason for manually overclocking in this day and age.

[ePeen reasons excluded]

 

[moinmoin]

I don't disagree but even in non-oc scenarios it could help with heat and noise or actually improve performance. I mean even x265 only uses AVX at some areas. While encoding you will have some cores running AVX code and others not. Lowering AVX offset will reduce power use of that AVX code giving more headroom to non-avx part. theoretically this could increase performance. Theoretically.

If that really ever happens that only means AMD needs to tweak the algorithm PB2 uses, not add a static offset band aid.

 

[Topweasel]

Yep - there is virtually no good reason for manually overclocking in this day and age.

[ePeen reasons excluded]

There really hasn't been a good reason for overclocking since Sandy Bridge anyways. I am fine with this bringing in the final nail and basically knowing with no effort that its at the silicon limit. Being good at overclocking is really just evidence that the manufacturer left performance on the table to either give you an incremental upgrade later or because they couldn't be bothered to deal with the binning (why sell a 5.2GHz 9900k when they can get all 8c16c dies to run at 5GHz and just bin most of them for locked 9900's?). The most famous overclocker only existed because while Intel developed the chip for economy sake it was actually better and faster then their performance CPU's. They really had to limit the clock speeds and bus speed to keep the $50-$75 from over running a $500 chip.

 

[Zucker2k]

Yep - there is virtually no good reason for manually overclocking in this day and age.

[ePeen reasons excluded]

Yet people on the current AMD platform are buying higher priced overclocking boards left, right, and center to run "stock,"?. Now AMD is maxing the silicon from the factory for benchmarking purposes, (cough, cough boost-gate), now all of a sudden it's okay for overclocking to die. There's always been a significant grey area between stable and the edge of stability for enthusiasts to exploit to their advantage. Where's that on AMD anymore? Factory overclocks should not be celebrated by enthusiasts! Intel is sure following this "stock" benchmarking shenanigans trend with the release of the 9900KS. Even so, I still expect more overclocking headroom on this chip than there is on Ryzen 2 at the moment. The all core overclock on the 9900k, for example, is 4.7GHz, which is crazy high already yet there's still anywhere from 300MHz (easily) to 600MHz reliable and stable room on top of that for those who could to exploit, cooling allowing.

 

[Markfw]

Yep - there is virtually no good reason for manually overclocking in this day and age.

[ePeen reasons excluded]

Well, I have to disagree. While my OC is small, I reduced my heat AND a small OC at the same time. Stock was 4050 and 85c. OC'ed and undervolted is 4100 and 65c.

 

[Atari2600]

Yet people on the current AMD platform are buying higher priced overclocking boards left, right, and center to run "stock,"?

Read again. MANUAL overclocking.

Now AMD is maxing the silicon from the factory for benchmarking purposes, (cough, cough boost-gate), now all of a sudden it's okay for overclocking to die.

Boostgate what? The clock rates that AMD have improved on with their most recent bios?

So performance would actually be **up** from the reviews. That boostgate?

There's always been a significant grey area between stable and the edge of stability for enthusiasts to exploit to their advantage.

Did you really expect large multinational companies to continue to give you a free lunch forever?

 

[Atari2600]

Well, I have to disagree. While my OC is small, I reduced my heat AND a small OC at the same time. Stock was 4050 and 85c. OC'ed and undervolted is 4100 and 65c.

You are very much a niche case and are more like a mini server farm that isn't so concerned with stability than home user that doesn't care about heat so much or an actual server farm where stability is king!

 

[Markfw]

You are very much a niche case and are more like a mini server farm that isn't so concerned with stability than home user that doesn't care about heat so much or an actual server farm where stability is king!

Not quite sure I follow you. Stability and temps are king in my house,. Is that what you were saying ? Uptime on linux boxes is virtually forever. It would be on windows if windows would quit patching everything, and rebooting.

 

[Elfear]

Yet people on the current AMD platform are buying higher priced overclocking boards left, right, and center to run "stock,"?. Now AMD is maxing the silicon from the factory for benchmarking purposes, (cough, cough boost-gate), now all of a sudden it's okay for overclocking to die. There's always been a significant grey area between stable and the edge of stability for enthusiasts to exploit to their advantage. Where's that on AMD anymore? Factory overclocks should not be celebrated by enthusiasts! Intel is sure following this "stock" benchmarking shenanigans trend with the release of the 9900KS. Even so, I still expect more overclocking headroom on this chip than there is on Ryzen 2 at the moment. The all core overclock on the 9900k, for example, is 4.7GHz, which is crazy high already yet there's still anywhere from 300MHz (easily) to 600MHz reliable and stable room on top of that for those who could to exploit, cooling allowing.

For years now we've had diminishing returns with overclocking. With Sandy Bridge, you could expect a sizable 1100-1300Mhz increase from boost clocks on air cooling which equates to a 30-35% increase (3.7Ghz --> 4.8-5.0Ghz). With Coffee Lake (9900k), a 300Mhz increase over boost clocks is the average overclock on air. That's a paltry 6% uplift. Ryzen 2 is even lower but when you're talking a few percentage points, does it really matter which one has more headroom?

As for AMD enthusiasts buying overclocking boards, they aren't needed when running stock unless you're talking a really low-end board matched with a 3900X. A decent B450 board will handle a 3900X just fine.

 

[ondma]

I don't know. I mean, if there were legitimately a way for AMD to get TSMC to fab Zen2-based CPUs, that would ACTUALLY clock at 5Ghz, part of me would be leaping for that chance, just to one-up Intel in yet another product dimension (clock freq.). It might potentially really shut-up Intel fans / AMD detractors for good, about the so called "frequency wars" (now that AMD has brought the "core wars" smackdown to Intel). Plus, the performance could be really good, thought they might want to implement an "AVX2 offset" like Intel does, in that case, just to keep power limits in check for "heavy" FPU workloads. (Like BOINC / PrimeGrid, as I am a fan of running.) A 125W TDP on those chips wouldn't be a bad thing, I don't think. It wasn't so very long ago, that AM2/AM2+ CPUs, like the Athlon X2 5600+ and 6000+, and the Phenom II X4/X6 CPUs were rated at 125W TDP. None of them died, to my knowledge.

So it is OK for AMD fans to brag about "winning" the "core wars", but intel should "shut up" about their frequency advantage? No double standard there, right?

 

[Markfw]

So it is OK for AMD fans to brag about "winning" the "core wars", but intel should "shut up" about their frequency advantage? No double standard there, right?

Their frequency advantage is now every small. But due to IPC, it is almost non existant. Take a stock 9900k and a stock 3800x, and with Intel at a slight frequency boost, the benchmarks on average are equal,.

As far as value, thats where the core wars comes into play. They have more at the same price. 3900x vs 9900k is a perfect example. The 3900x is also on parity with the 9900k in price, I used the 3800x as my example since it also has 8 cores. The 3900x pretty much kills the 9900k except single core tasks.