If I've learned anything over the years, generalizations with current games don't necessarily mean jack about the future. Vulkan and D3D 12 are immature and volatile at the moment.

We don't even know if future games will be well designed for current hardware. It seems like these low level APIs need TLC to run properly on any given hardware. So maybe get ready to upgrade even more often if you like to play the latest eyecandy.

swaaye wrote:

It seems like these low level APIs need TLC to run properly on any given hardware.

Yup... well, they have been warned.
As they say: be careful what you wish for.

Scali wrote:

These are lies spread by clueless AMD fanboys, who read something about pre-emption in NVidia's whitepaper, and don't quite unde […]
Show full quote

laxdragon wrote:

edit: It has come to my attention that the 3Dmark bench did not really support full async, more of a preemptive mode.

These are lies spread by clueless AMD fanboys, who read something about pre-emption in NVidia's whitepaper, and don't quite understand the whole story.
Firstly, the DX12 API cannot control how async workloads are executed. So you can't specify whether you want 'full async', 'pre-emptive mode', or 'sequential' or whatever.
Only the driver can decide this, the application can't even tell what the driver does. FutureMark explains it very well in their press release: http://www.futuremark.com/pressreleases/a-clo … 3dmark-time-spy
Secondly, fine-grained pre-emption is ONE feature that Pascal has introduced to improve latencies on async workloads: http://www.anandtech.com/show/10325/the-nvidi … ition-review/10
This is more relevant to combinations of heavy/long and high-priority/time-critical workloads, and doesn't really apply to Time Spy.
There is however ANOTHER feature for async compute, and that is a hardware-based dynamic load balancing: http://www.anandtech.com/show/10325/the-nvidi … dition-review/9
This is very similar to AMD's implementation with ACE's, which is why Time Spy gets decent gains from the same code running on either AMD or NVidia Pascal cards.

Maxwell v2 does not have this feature, and its async compute is more difficult to optimize for. That is why the driver makes all the code run sequentially by default, which is usually faster than running async code that is not properly tuned for the architecture. Effectively that means turning async on or off in Time Spy does nothing, because even though Time Spy uses multiple command queues when async is on, the driver will put them into a single queue internally, which comes down to the same thing as running Time Spy with async off, where the application puts everything into a single queue.

Another article dealing with Time Spy: http://www.extremetech.com/gaming/232122-lies … 2-time-spy-test

In case anyone was still in doubt, AMD has given Time Spy its official seal-of-approval: http://radeon.com/radeon-wins-3dmark-dx12
Seems AMD is perfectly happy with how 3DMark is show-casing async compute. Then again, who doubted this? AMD was involved in the development of Time Spy (as were nVidia and Intel).

Scali wrote:

Not really. The gap between AMD and nVidia is (currently) a bit smaller in DX12/Vulkan. But AMD still loses.

if you are talking 480 vs 1060 that is not correct

the 1060 is only superior on the Nvidia sponsored Rise of the Tomb Raider, on the AMD sponsored and more neutral games the 480 tends to be faster.

the Vulkan CPU bound testing is interesting and it needs further development (the current patch/drivers fails to even apply vsync properly!)

AMD made a good job with low CPU overhead and high drawcalls with Mantle (plenty of tests showing) and DX12, their problem with high overhead on DX11 doesn't seem to apply to those, also those Vulkan results don't seem to apply to DX12 games.

but I have to say, seeing all the benchmarks being done with just 6700Ks OC and not a cheaper CPU is disappointing, I remember when Anandtech over 15 years ago would test the new VGAs with 2 or 3 different CPUs of significant performance differences.
I think we need that for cards like the 1060 and lower at least, since most consumers don't run 4.5GHz i7s I would think.... and for DX11 alone the advantage Nvidia can maintain with slower CPUs is significant, Vulkan is unexpected, but as I said, it's to immature it seems.

I think Time Spy is solid at trying to be neutral, but reality will see some Rise of the Tomb Raiders and some Hitmans, favoring each vendor due to deals they made 😐

SPBHM wrote:

but I have to say, seeing all the benchmarks being done with just 6700Ks OC and not a cheaper CPU is disappointing, I remember when Anandtech over 15 years ago would test the new VGAs with 2 or 3 different CPUs of significant performance differences.
I think we need that for cards like the 1060 and lower at least, since most consumers don't run 4.5GHz i7s I would think.... and for DX11 alone the advantage Nvidia can maintain with slower CPUs is significant, Vulkan is unexpected, but as I said, it's to immature it seems:

Yeah, it's a little frustrating when review sites almost exclusively test with hardware I could never afford, at resolutions I could never achieve. The results aren't quite as meaningful to those with modest hardware budgets. It comes off as elitist, and if you're not rich enough to play, too bad.

clueless1 wrote:

Yeah, it's a little frustrating when review sites almost exclusively test with hardware I could never afford, at resolutions I could never achieve. The results aren't quite as meaningful to those with modest hardware budgets. It comes off as elitist, and if you're not rich enough to play, too bad.

I guess the scenery is also changing, with the APUs and all.
If you look at nVidia, the most low-end card of the last generation was the GTX960. They didn't have a 940 or 950. Possibly because an integrated GPU would be too close in performance for people to bother getting such a graphics card. So if you have a smaller budget, then you probably don't even buy a separate graphics card at all anymore, these days.
I wonder if the GTX1060 will be the slowest, or if a GTX1050 will arrive, and then, how much more low-end will that be? The GTX1060 is still a pretty powerful card, so this time there's more room for cards at the bottom (until iGPUs catch up at least).

Scali wrote:

I guess the scenery is also changing, with the APUs and all. If you look at nVidia, the most low-end card of the last generation […]
Show full quote

clueless1 wrote:

Yeah, it's a little frustrating when review sites almost exclusively test with hardware I could never afford, at resolutions I could never achieve. The results aren't quite as meaningful to those with modest hardware budgets. It comes off as elitist, and if you're not rich enough to play, too bad.

I guess the scenery is also changing, with the APUs and all.
If you look at nVidia, the most low-end card of the last generation was the GTX960. They didn't have a 940 or 950. Possibly because an integrated GPU would be too close in performance for people to bother getting such a graphics card. So if you have a smaller budget, then you probably don't even buy a separate graphics card at all anymore, these days.
I wonder if the GTX1060 will be the slowest, or if a GTX1050 will arrive, and then, how much more low-end will that be? The GTX1060 is still a pretty powerful card, so this time there's more room for cards at the bottom (until iGPUs catch up at least).

I think GTX 750Ti has around 2-3 times more gaming power than Intel's Iris Pro GPU's. There would be lots of space for a GTX 1050.

I understand why they have to use i7 processors to eliminate bottlenecks in the results, but it would be nice if when they tested a mid-range or entry-level graphics card, they'd test it as well in a 2nd testbed, one that matches its performance profile. Then you'd get a whole new dynamic, like perhaps the 1060 beating the 480 more significantly, along with the knowledge that the 480 may be less efficient/more cpu-bound, etc. Just sayin.

clueless1 wrote:

I understand why they have to use i7 processors to eliminate bottlenecks in the results, but it would be nice if when they tested a mid-range or entry-level graphics card, they'd test it as well in a 2nd testbed, one that matches its performance profile. Then you'd get a whole new dynamic, like perhaps the 1060 beating the 480 more significantly, along with the knowledge that the 480 may be less efficient/more cpu-bound, etc. Just sayin.

Yup, the above DOOM Vulkan results don't bode well for RX480... These new APIs were originally marketed as "lower CPU overhead", so you can use a low-end CPU and still get great framerates.
But the differences are so significant that you'd be better off with an nVidia card running OpenGL than with an AMD card running Vulkan, if you have a very slow CPU.
On nVidia the CPU really doesn't matter much, only 4 fps difference between the slowest CPU and a super high-end one.

It's very important that reviewers become aware of this, and do more testing with CPU scaling.

SPBHM wrote:

since most consumers don't run 4.5GHz i7s

The funny thing is that X5667 (3.06GHz base quad w/ lots of OC headroom) can be found for $35 on ebay US (I was searching for hexa X5670/5675) ... Yes, good OC'ing X58 boards may be still expensive but it pays to know this stuff.

Saw on a reddit user review I came across, even an i7 930 is still enough to keep up with a GTX 1070.

I have a similar setup. I figure my 8C/16T @ 3.9Ghz system is going to be plenty for even multiple 480's.

archsan wrote:

Unfortunately not all manufacturers list that in their specs. Is there a jonnyguru equivalent for gfx cards? […]
Show full quote

nforce4max wrote:

Just look for custom models using a different pcb design with a good phase count then everything will be fine plus the overclocking will be a bit better.

Unfortunately not all manufacturers list that in their specs. Is there a jonnyguru equivalent for gfx cards?

I bought reference GTX 470s days after launch and don't want to repeat it again... Sold one early and still have one that survives til today though, but that's more because it's been very rarely used.

Anyway this pretty much explains all those "FE" cards being pretty much ignored on the market. I almost shed a tear for those who preordered and bought them enthusiastically earlier. Reminds me of my 2010 self.

edit: OK, I found this and this and some more like this. I hope for non-OC purposes, 8-phase (e.g. Palit Super Jetstream) is good enough? I could really use some primer on this.

All you got to do is look at the bare card and if you got an idea what to look for you can count the phases by just the number of coils and go from there. I would just avoid the trouble and just wait till more custom models come out while keeping an eye out for any complaints by early buyers. If there starts being complaints by owners about failing or dead cards with in the first six months then chances are the rest of the generation is likely to suffer the same faults.

The more phases the better as the load per phase is lower and it allows for redundancy should a phase fail the card can continue as normal with the remaining phases.

The 8+2 cards are good, just make sure that there are no hot spots and maintain them they should be good for years 10,000 to 20,000 hours easily. The crippled cards are likely to fail in the first or second year of service depending on workload and climate.

Aye aye cap'n!!! 😁

Suddenly I have new appreciation for Galax(y) cards and some of the other specialist makes.

As for you guys looking for custom 1060 this is what I found so far: http://wccftech.com/nvidia-geforce-gtx-1060-c … model-round-up/

nVidia continues on its Pascal mission, here comes the big one: http://www.geforce.com/hardware/10series/titan-x-pascal

Ah PC gaming, not only do we battle console gamers, but we squabble amongst ourselves.

Scali wrote:

I guess the scenery is also changing, with the APUs and all. If you look at nVidia, the most low-end card of the last generation […]
Show full quote

clueless1 wrote:

Yeah, it's a little frustrating when review sites almost exclusively test with hardware I could never afford, at resolutions I could never achieve. The results aren't quite as meaningful to those with modest hardware budgets. It comes off as elitist, and if you're not rich enough to play, too bad.

I guess the scenery is also changing, with the APUs and all.
If you look at nVidia, the most low-end card of the last generation was the GTX960. They didn't have a 940 or 950. Possibly because an integrated GPU would be too close in performance for people to bother getting such a graphics card. So if you have a smaller budget, then you probably don't even buy a separate graphics card at all anymore, these days.
I wonder if the GTX1060 will be the slowest, or if a GTX1050 will arrive, and then, how much more low-end will that be? The GTX1060 is still a pretty powerful card, so this time there's more room for cards at the bottom (until iGPUs catch up at least).

Uhh, there WAS a 950. https://techreport.com/review/29061/nvidia-ge … s-card-reviewed

mr_bigmouth_502 wrote:

Uhh, there WAS a 950. https://techreport.com/review/29061/nvidia-ge … s-card-reviewed

Ah right. Point remains though, there used to be many more low-end cards than there are today.
You had a 705, 710, 720, 730, 740 and 745.

That is true. Slowest card is the 950, but that card still costs just under A$ 200 over here. The cheaper cards are all 730 cards or similar. No 940 or 930. I'm not quite sure why, are onboard graphics that good these days?

Scali wrote:

mr_bigmouth_502 wrote:

Uhh, there WAS a 950. https://techreport.com/review/29061/nvidia-ge … s-card-reviewed

Ah right. Point remains though, there used to be many more low-end cards than there are today.
You had a 705, 710, 720, 730, 740 and 745.

Weren't a lot of those like OEM specials?

mr_bigmouth_502 wrote:

Scali wrote:

mr_bigmouth_502 wrote:

Uhh, there WAS a 950. https://techreport.com/review/29061/nvidia-ge … s-card-reviewed

Ah right. Point remains though, there used to be many more low-end cards than there are today.
You had a 705, 710, 720, 730, 740 and 745.

Weren't a lot of those like OEM specials?

Some of them were, others were sold as retail. I see 710, 720, 730 and 740 listed in pricewatch.