AMD Ryzen 7 1800X (14nm Zen)

[scaryjim]

It depends on what sort of optimisation you're referring to. Things like task scheduling are down to the OS kernel, along with e.g. time slicing, prioritisation and interrupts. ....

I wonder if one quick and simple optimisation (if this is even possible in Win 10) might be for the game exe to set its own processor priority - so when it detects a Ryzen with more than 8 logical cores it sets a processor affinity that would only use one CCX ... that should avoid the cache misses that might occur when Win 10 moves threads around.

Huh, in fact, I wonder if any reviewers have considered trying that to see if it has any effect on game performance... should be straightforward to start a game, run a benchmark, manually adjust the affinity, then rerun the benchmark...

[CAT-THE-FIFTH]

I wonder if one quick and simple optimisation (if this is even possible in Win 10) might be for the game exe to set its own processor priority - so when it detects a Ryzen with more than 8 logical cores it sets a processor affinity that would only use one CCX ... that should avoid the cache misses that might occur when Win 10 moves threads around.

Huh, in fact, I wonder if any reviewers have considered trying that to see if it has any effect on game performance... should be straightforward to start a game, run a benchmark, manually adjust the affinity, then rerun the benchmark...

They already tested it - check my post earlier in the thread. 4+0>>>>>>>2+2.

[scaryjim]

They already tested it - check my post earlier in the thread. 4+0>>>>>>>2+2.

Ah, just found it - I missed it in the turn of a page

Time to buy SJ,a big virtual beer?? ...

I'll take a small real beer instead, if you're offering

EDIT: having finally read that review now (had to stop for a takeaway ) I'm a bit disappointed that they didn't test the 4+0 with SMT on/off, would've been an interesting test to see if SMT is still an issue even without the cross-L3 latency. Nonetheless, interesting to see one of the suspected issues tested and demonstrated so easily - if AMD can introduce a CPU driver or MS can release a new Windows scheduler that knows not to move threads off a CCX, that should go a long way towards mitigating some of the gaming performance issues (seriously, that's what, a 10% performance boost just because you're not going across the CCX for L3 cache access...?)

[CAT-THE-FIFTH]

Ah, just found it - I missed it in the turn of a page



I'll take a small real beer instead, if you're offering

EDIT: having finally read that review now (had to stop for a takeaway ) I'm a bit disappointed that they didn't test the 4+0 with SMT on/off, would've been an interesting test to see if SMT is still an issue even without the cross-L3 latency. Nonetheless, interesting to see one of the suspected issues tested and demonstrated so easily - if AMD can introduce a CPU driver or MS can release a new Windows scheduler that knows not to move threads off a CCX, that should go a long way towards mitigating some of the gaming performance issues (seriously, that's what, a 10% performance boost just because you're not going across the CCX for L3 cache access...?)

It looks like Hardware.fr has done some more testing with SMT off.




But here is the big kicker!!



They tested in performance mode and found not all the regressions were actually only down to SMT!

[kalniel]

Core parking was found to slow down intel chips back in the day as well, surprised it still causes problems.

[scaryjim]

... They tested in performance mode and found not all the regressions were actually only down to SMT!

Ugh, too many variables!

So, BF1, Project Cars and Civ 6 all hit the caches hard - they're the titles with the biggest regression when you split the cores 2+2. Gaming's always been quite cache sensitive anyway, so that's not a huge surprise.

Warhammer and Watch Dogs 2 are the only games were SMT causes a significant performance degradation in balanced power mode. All other games show similar or greater performance in Performance mode with SMT on as in balanced mode with SMT off. Of course, we don't have figures for performance mode with SMT off, which might show that SMT is also contributing to the regression...

Yep, too many variables. It looks like there are all sorts of optimisations that could pick up 1% here and there, and those 1% can start to add up quickly...

Core parking was found to slow down intel chips back in the day as well, surprised it still causes problems.

I'm surprised it seems to make up so much of the difference when switching to Performance mode, given AMD's insistence that the problem with Balanced mode is that it slows down the clock and voltage changes - in fact even more so, since they initially blamed core parking then "clarified" the response...

[ik9000]

A bit late to the thread, sorry if this is covered already but am I missing something or does even the high-end Zen only have 16 PCIe3 Lanes? Seems a bit small for a new architecture/chipset... Hopefully I've got this wrong.

[CAT-THE-FIFTH]

A bit late to the thread, sorry if this is covered already but am I missing something or does even the high-end Zen only have 16 PCIe3 Lanes? Seems a bit small for a new architecture/chipset... Hopefully I've got this wrong.

I don't think it is that much of a biggie since you need to buy a £600 Core i7 6850K to get any more on the Intel side.

Also,Ryzen is an SOC so the PCI-E controller is in the CPU itself which actually might be an advantage.]

Ugh, too many variables!

So, BF1, Project Cars and Civ 6 all hit the caches hard - they're the titles with the biggest regression when you split the cores 2+2. Gaming's always been quite cache sensitive anyway, so that's not a huge surprise.

Warhammer and Watch Dogs 2 are the only games were SMT causes a significant performance degradation in balanced power mode. All other games show similar or greater performance in Performance mode with SMT on as in balanced mode with SMT off. Of course, we don't have figures for performance mode with SMT off, which might show that SMT is also contributing to the regression...

Yep, too many variables. It looks like there are all sorts of optimisations that could pick up 1% here and there, and those 1% can start to add up quickly...



I'm surprised it seems to make up so much of the difference when switching to Performance mode, given AMD's insistence that the problem with Balanced mode is that it slows down the clock and voltage changes - in fact even more so, since they initially blamed core parking then "clarified" the response...

I told you it was more like death by a 1000 cuts.

[ik9000]

I don't think it is that much of a biggie since you need to buy a £600 Core i7 6850K to get any more on the Intel side.

No, that is a biggy. A high end CPU ought to offer more. You get 28 Lanes with the i7-5820K at £380.

They're saying they only want to target systems who need limited bandwidth. A multicore CPU, DDR4 but only 1GPU at full speed and no allowance for PCIe cards for NVMe drives, adding USB3.1 or the like.

[kalniel]

A bit late to the thread, sorry if this is covered already but am I missing something or does even the high-end Zen only have 16 PCIe3 Lanes? Seems a bit small for a new architecture/chipset... Hopefully I've got this wrong.

It has 24 lanes, cf 16 for Intel's 7700K.

[CAT-THE-FIFTH]

No, that is a biggy. A high end CPU ought to offer more. You get 28 Lanes with the i7-5820K at £380.

They're saying they only want to target systems who need limited bandwidth. A multicore CPU, DDR4 but only 1GPU at full speed and no allowance for PCIe cards for NVMe drives, adding USB3.1 or the like.

It isn't really since most multi-GPU systems on forums are using consumer socket CPUs like the Core i7 7700K,and 28 lanes on a Core i7 6800K won't give you 16X/16X dual cards. Ryzen has 24.

Hence,most multi-card Intel systems will be running cards at 8X/8X meaning you sill have plenty of lanes left for other stuff with AM4.

Then add the fact,anything over two cards scales poorly anyway and more and more games don't really do well with multiple cards,since they are developed for consoles first.

[scaryjim]

... A multicore CPU, DDR4 but only 1GPU at full speed and no allowance for PCIe cards for NVMe drives, adding USB3.1 or the like.

The SoC supports an x4 NVMe drive and USB3.1 Gen 2 natively on top of the x16 cluster for GPUs. The AM4 chipset supports at least one additional USB 3.1 Gen 2 and 2 USB 3.1 Gen 1 ports, plus various configs of SATA and SATA Express ports, and a cluster of PCIe 2.0 endpoints for peripherals. And when you're not using a chipset the Gen 3 x4 cluster that normally connects to the chipset is available for hanging other peripherals off.

In other words, it's got plenty of allowance for NVMe drives, USB 3.1 ports etc.

AM4 is a mainstream consumer/prosumer platform intended to compete in the same market as s115X. For people who must have more PCIe lanes there will be a workstation/server platform under the Opteron branding...

[CAT-THE-FIFTH]

The SoC supports an x4 NVMe drive and USB3.1 Gen 2 natively on top of the x16 cluster for GPUs. The AM4 chipset supports at least one additional USB 3.1 Gen 2 and 2 USB 3.1 Gen 1 ports, plus various configs of SATA and SATA Express ports, and a cluster of PCIe 2.0 endpoints for peripherals. And when you're not using a chipset the Gen 3 x4 cluster that normally connects to the chipset is available for hanging other peripherals off.

In other words, it's got plenty of allowance for NVMe drives, USB 3.1 ports etc.

AM4 is a mainstream consumer/prosumer platform intended to compete in the same market as s115X. For people who must have more PCIe lanes there will be a workstation/server platform under the Opteron branding...

More importantly Intel and its stupid product segmentation helps AMD - you need to spend £600 on a Core i7 6850K to get above 28 lanes,and unless you really want to push the overclock on a R7 1800X,the B350 motherboards are far cheaper than an X99 one and AFAIK support XFire.

[scaryjim]

... the B350 motherboards are far cheaper than an X99 one and AFAIK support XFire.

Officially only the X370 and X300 support 2 x8 PCIe for GPUs (see the specs tab on http://www.amd.com/en-us/products/chipsets/am4), however the 970 chipset for AM3 didn't officially support Crossfire either but boards often claimed to

I suspect some mobo manufacturers will provide a second physical x16 PCIe slot running at x4 (potentially only PCIe 2.0 x4 at that) and claim to be crossfire compatible...

I find it interesting that X300 supports 2 x8 slots though - that must mean AMD are anticipating it being used in mATX boards as well as mITX (since mITX doesn't have room for 2 slots!). And since we now have confirmation that X300/A300 can use the x4 that usually connects to the chipset, that actually makes a lot of sense...

EDIT: just realised I made a mistake in my last post - the SoC only supports USB 3.1 Gen 1, not Gen 2. Still, 4 USB 3 hanging directly off the CPU isn't bad, is it

[CAT-THE-FIFTH]

Officially only the X370 and X300 support 2 x8 PCIe for GPUs (see the specs tab on http://www.amd.com/en-us/products/chipsets/am4), however the 970 chipset for AM3 didn't officially support Crossfire either but boards often claimed to

I suspect some mobo manufacturers will provide a second physical x16 PCIe slot running at x4 (potentially only PCIe 2.0 x4 at that) and claim to be crossfire compatible...

I find it interesting that X300 supports 2 x8 slots though - that must mean AMD are anticipating it being used in mATX boards as well as mITX (since mITX doesn't have room for 2 slots!). And since we now have confirmation that X300/A300 can use the x4 that usually connects to the chipset, that actually makes a lot of sense...

EDIT: just realised I made a mistake in my last post - the SoC only supports USB 3.1 Gen 1, not Gen 2. Still, 4 USB 3 hanging directly off the CPU isn't bad, is it

OK,unofficially - but if you need 8 cores,an R7 1700 for £320 plus a £100 B350 is barely the price of a Core i7 6800K itself.

Also,apparently people are getting decent overclocks on the R7 1700 using the STOCK COOLER.

It seems the Wraith Spire is not too bad at all.

[scaryjim]

... if you need 8 cores,an R7 1700 for £320 plus a £100 B350 is barely the price of a Core i7 6800K itself.

Also,apparently people are getting decent overclocks on the R7 1700 using the STOCK COOLER. ....

Yeah, as usual the lower power variants are the interesting ones (remember the Phenom II X3 720e and X4 910e? ). Really looking forward to the quad core processors now...