Discussion Zen 7 speculation thread

[igor_kavinski]

¹) In general computing at least; maybe there are niche applications of such an approach.

Yup. That niche would be gaming, especially ones that can't use more than 8 threads so the CPUs having more than that many threads will have something extra to do rather than have some threads twiddling their thumbs.

 

[Win2012R2]

because it would be a huge waste of energy

Not if it gets CS2 FPS up from 600 to 800.

 

[MS_AT]

One would think that computer enthusiast after spectre would understand branch prediction has side effects.

Leaving security aside processing two paths of the branch is a perfect way to trash caches, including TLB.

Now when it comes to pure math reason why it does not make sense branch prediction accuracy is usually >90%. So 10% work is wasted. Scheme when you process 2 sides of the branch gives you exactly 50% waste.

 

[igor_kavinski]

From Agner Fog:



Assuming that a context switch happens in less than 10 clock cycles, that could allow 50% smoother gaming performance on average.

 

[inquiss]

Maybe but I don't think even Zen 7+ on A14 SPR would necessarily justify DDR6 in 2029. Like DDR5 is going to get faster (already at 8400 MT/s CUDIMM, imagine what CAMM2 DDR5 will bring) between now and the next few years and AMD will probably want something to go against TTL & HML as an interim.

I mean Zen 8 will be 100% DDR6 for sure, just a question if AMD can release it with 1nm CCDs maybe late 2029 somehow or it has to be 2030, or Zen 8 is A14 SPR.

Why not LPDDR as standard by then?

 

[511]

Why not LPDDR as standard by then?

So we want MB with soldered Ram?

 

[inquiss]

So we want MB with soldered Ram?

Why would LPDDR need to be soldered? One does not require the other....

 

[511]

Why would LPDDR need to be soldered? One does not require the other....

I have not seen a single LPDDR in non soldered format

 

[Darkmont]

I have not seen a single LPDDR in non soldered format

LPCAMM2 is a ratified JEDEC standard. It’s in the Lenovo ThinkPad P1 Gen 7

 

[Doug S]

6 GHz was breached but not for mainstream CPUs. That should obviously change in 2026. Maybe 8 GHz by 2030 if we are very lucky?

We won't see 7 GHz by 2030, let alone 8.

 

[Doug S]

I have not seen a single LPDDR in non soldered format

I think most will probably until they transition to LPDDR6 before worrying about CAMM support.

 

[Geddagod]

We won't see 7 GHz by 2030, let alone 8.

Bold prediction

 

[Doug S]

Bold prediction

Not really. We had the first 5 GHz CPU about 17 years ago. It isn't that long ago we saw the first 6 GHz. I feel pretty confident in this, but feel free to bookmark it (or better screen capture it if the forum doesn't survive) and make me eat crow if I'm wrong! 😉

 

[fastandfurious6]

We had the first 5 GHz CPU about 17 years ago

exponential difficulty

 

[Saylick]

We won't see 7 GHz by 2030, let alone 8.

In stock form or with non-exotic cooling OC?

 

[igor_kavinski]

(or better screen capture it if the forum doesn't survive) and make me eat crow if I'm wrong! 😉

Done!

Where can you be reached?

 

[StefanR5R]

[wider speculative execution]

[stalls after branch misprediction]
Assuming that a context switch happens in less than 10 clock cycles, that could allow 50% smoother gaming performance on average.

Branching is deciding between two (sometimes more) points in a task to continue at. A context switch is interrupting one task and taking on another.

Re smooth games:
~20 cycles of an e.g. 4 GHz CPU = 5 ns (nanoseconds)¹
frame-to-frame time at e.g. 120 Hz video frame rate: 8 ms (milliseconds, that's million nanosecods)

Also, IIRC 10 ms (milliseconds) is a ballpark under which audio latency needs to stay before musicians perceive a lag while hearing themselves on headphones (i.e. the whole chain of recording, processing, and playing back of live audio, with all the necessary buffering included). I suppose similar measures need to be taken for game audio.

________
¹) The ~20 cycles of penalty of a branch misprediction of which Agner Fog was writing about require that the stalled pipeline can be refilled with instructions and data which do reside in 1st level cache, AFAIU. However, even a last-level cache miss results in "only" a few hundred cycles of a stall, which still is ~five orders of magnitudes smaller than a video frame time. But it follows that [speculative] prefetching of instructions and data is much, much more important for purposes like video game smoothness, compared to speculative execution.

 

[Win2012R2]

Re smooth games:

I guess this won't be fixing Unreal Engine 5 stuttering

 

[soresu]

I guess this won't be fixing Unreal Engine 5 stuttering

Bad optimisation on the part of game devs and low parallelization of the engine are the causes of that.

Epic can fix the latter, and are in the process of doing so in their UE6 roadmap.

But the former is all on game devs, which have become increasingly lax due to the likes of DLSS/FSR "free lunch" performance increases.

 

[soresu]

Not really. We had the first 5 GHz CPU about 17 years ago. It isn't that long ago we saw the first 6 GHz. I feel pretty confident in this, but feel free to bookmark it (or better screen capture it if the forum doesn't survive) and make me eat crow if I'm wrong! 😉

Priorities changed.

They could have kept increasing clock frequency with minimal IPC investment, but the power cost would likely have ballooned significantly faster.

Also adding more cores and more SIMD ALUs per core constrained the rate of clock frequency gains.

Bear in mind that single core CPUs were the norm for many decades longer than that.

 

[Doug S]

In stock form or with non-exotic cooling OC?

IBM POWER 6, 5 GHz shipped in May 2008.

 

[adroc_thurston]

IBM POWER 6, 5 GHz shipped in May 2008.

Hey that's decidedly not mainstream computing.
The first actual real client 5g you could buy was what, 9900KS?

 

[LightningZ71]

Intel core i7 8086K was, IIRC, the first client 5Ghz CPU.

 

[adroc_thurston]

Intel core i7 8086K was, IIRC, the first client 5Ghz CPU.

Meme SKU, just like the 5G HFT parts from eons past.
I mean something you can just go and buy.

 

[lamedude]

AMD FX-9590