AMD - Zen chitchat

[CAT-THE-FIFTH]

Apparently 14NM+ was meant to be denser,IIRC.

[Corky34]

Bit like me then.

WTF,is happening here:

https://www.pcgamesn.com/amd-raven-ridge-overclocking

So,potentially with BCLK overclocking you could get decent overclocks,but it seems BCLK is locked down.

That seemed really odd when pcgamesn.com reported on it, i thought it had to be a bug but obviously didn't know one way or the other, it turned out it's a repeat of the Summit Ridge bug where Windows thinks time is running slower.

EDIT: Looks like der8auer has tested Raven ridge delidding.

[CAT-THE-FIFTH]

Bit like me then.



That seemed really odd when pcgamesn.com reported on it, i thought it had to be a bug but obviously didn't know one way or the other, it turned out it's a repeat of the Summit Ridge bug where Windows thinks time is running slower.

EDIT: Looks like der8auer has tested Raven ridge delidding.

Ah,a shame.

Also,I noticed that I have not seen any reviews testing CPU performance and RAM speed on Raven Ridge!

Ryzen likes moar RAM bandwidth due to the dual CCX design,but Raven Ridge uses a single CCX.

[Xlucine]

I could be misremembering but doesn't the R5 2400G retain the same TDP as the R5 1400, i would guess the change to 14nm+ had something to do with lowering the power draw, along with losing a CCX obviously, but to throw in a GPU while retaining the same TDP is pretty impressive IMO.

2400G has the same TDP as the 1400, 1500X and 1600

[Ozaron]

Received my 2400g, installed to ASrock A320m less than an hour ago. No luck getting to BIOS, as expected, system powers on and just thinks to itself for a little while before powering off and trying again. Will be applying for boot kit.

[DanceswithUnix]

Anyone seen RR supporting ECC ram yet?

[DanceswithUnix]

Trouble with cheap motherboards and RR at Phoronix, looks like it is not a Linux ready platform at least:

https://www.phoronix.com/scan.php?pa...dge-Mobo-Linux

[DanceswithUnix]

Rumour of 14.5% improvement for single thread geekbench of Ryzen 2600 vs current 1600

https://wccftech.com/amd-ryzen-5-260...formance-leak/

[Xlucine]

Rumour of 14.5% improvement for single thread geekbench of Ryzen 2600 vs current 1600

https://wccftech.com/amd-ryzen-5-260...formance-leak/

That's suspiciously close to the 10% that was ballparked for zen+ previously, but for IPC alone. I wonder how much is down to faster RAM? I'd expect the 2X00 CPUs to at least match the 2.9 GHz rated RAM speed of the APUs

[scaryjim]

Rumour of 14.5% improvement for single thread geekbench of Ryzen 2600 vs current 1600 ...

Interesting .. given the relative clocks that's indicative of a reasonable IPC uptick as well (around 8%, if my maths aren't way out). I'd assume the big jump in multithreaded score (+30% ) is down to Precision Boost 2 being able to run the cores much faster in the MT benchmark.

I suppose the big question is whether the 95W SKUs are going to see similar improvements or bigger ones. If the 12nm process has a flatter v/f_max curve, *and* it can do +200MHz in the 14nm sweet spot, we could see something pretty stunning....

[DanceswithUnix]

I'd assume the big jump in multithreaded score (+30% ) is down to Precision Boost 2 being able to run the cores much faster in the MT benchmark.

Or IMC/cache prefetch improvements.

[CAT-THE-FIFTH]

Looks like AMD has hired a senior person from Cisco:

https://www.theregister.co.uk/2018/0..._epyc_new_gig/

Edit!!

CPU tests on RR with faster RAM:

https://www.phoronix.com/scan.php?pa...Memory-Scaling

It seems to have better CPU performance with faster RAM,so it appears even a single CCX design likes faster RAM(unless OFC it is down to the lower amount of L3 cache).

[DanceswithUnix]

It seems to have better CPU performance with faster RAM,so it appears even a single CCX design likes faster RAM(unless OFC it is down to the lower amount of L3 cache).

I thought it still had the same fabric, just this time tying together a CCX and GPU rather than a CCX pair.

[CAT-THE-FIFTH]

I thought it still had the same fabric, just this time tying together a CCX and GPU rather than a CCX pair.

One of the major arguments for Ryzen needing faster RAM was down to the bandwidth between a pair of CCX modules being limited,and faster RAM would mean less of a cross CCX penalty,when using a second CCX for CPU stuff.

OFC,I would love to see some gaming results with a graphics card,to see if faster RAM does impact CPU performance significantly,but apparently most sites have not bothered doing this. Intel on average gained less from faster RAM.

[scaryjim]

I thought it still had the same fabric, just this time tying together a CCX and GPU rather than a CCX pair.

Part of the assumption about memory clocks boosting Summit Ridge performance was that it sped up communication between the two CCXes, because the fabric clock is tied to the memory clock. On the surface it makes sense: you speed up the access to the other CCX's L3 cache, which increases performance.

Thing is, most people ignored the fact that cross-CCX cache latency was pretty much identical to main memory latency - SR CPUs essentially acted like 2 processors with 8MB L3 cache each and a shared main memory access. Speeding up the memory does reduce the latency of cross-CCX communication, but only to the same extent that it reduces latency back to main memory too (i.e. there's no gain from having the data in the other CCX's L3 cache). So it's not really surprising that you get similar gains from faster memory on RR - you're reducing the latency associated with misses in the L3 cache.

One notable thing about the Ryzen IMC is its very high latency to main memory - having a quick look at the MSI Mortar AM4 board review, Coffee Lake with 2400MHz CL 16 RAM has significantly lower latency than Raven Ridge with 2933MHz CL 14 RAM (as in 13ns - almost 20% - lower)*. That's pretty ridiculous - faster RAM with tighter timings but much lower latencies? That bottleneck has to be in either the IMC or the fabric....


*you get some very interesting results if you look at the theoretical latencies in that review. 14 cycles at 2933MHz gives a theoretical 47ns, v 70ns measured latency. 16 cycles at 2400MHz gives a theoretical 67ns, against a measured latency of 57ns. Somewhere along the line something's gone fishy with the Intel latency testing....

[CAT-THE-FIFTH]

It will be interesting to see if Ryzen+ improves in this regard. Currently being stuck with 16gb of 2400mhz DDR4 does kind of not help my own circumstances when it comes to my next upgrade.