If AMD can pull off seamless multi-gpu with Navi then that will put them in a seriously good position in both gaming and non-gaming markets. The current problem with SLI and Crossfire is that games must support the configurations but with Navi any game would work with all the gpu's present in the system.
-
Competitor rules
Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.
Vega refresh in 2018, current Vega is broken!
- Thread starter lowdog
- Start date
If AMD can pull off seamless multi-gpu with Navi then that will put them in a seriously good position in both gaming and non-gaming markets. The current problem with SLI and Crossfire is that games must support the configurations but with Navi any game would work with all the gpu's present in the system.
Both nVidia and AMD are looking at the same kind of approach - multi-chip-modules in the long term - this isn't traditional multi GPU or even multi GPU really and probably won't come with first generation Navi which will likely be architectural changes as a start down that road. This is basically dissecting a traditional monolithic die into different modules that allow you to flexibly implement and scale the specification of the GPU.
Traditionally you couldn't have the latency and bandwidth that allowed the command processors in one GPU access shaders or other sub-systems that were off die directly which is a big limitation for multi GPU - recent advanced in substrate technology means that a seperate "command" package could interface with multiple dumb/headless processing packages without penalty.
No... that's in fact our own job description
Pseudo logic.
The Bandwidth of HBM is needed because GPU's are getting too powerful for GDDR, that doesn't mean they are not too powerful for HBM.
You stick GDDR5X on the same GPU it will be slower.
Permabanned
- Joined
- 12 Sep 2013
- Posts
- 9,221
- Location
- Knowhere
If it works as well in a gpu as it has with Threadripper they should be able to jump ahead, at least until Nvidia get their own version going. This fabric has the potential to give us 100 to 150 % performance increases going from Vega to Navi.
Permabanned
- Joined
- 12 Sep 2013
- Posts
- 9,221
- Location
- Knowhere
I don't think it would be. If your gpu is bandwidth starved yes it'll make a difference but if it has enough anyway not so much. So far we haven't seen HBM making a big difference with Fiji or Vega. As an example, comparing the Hawaii/Grenada chips with the 512 bit bus the Fiji chips weren't as far ahead of Grenada as we were all hoping. Certainly not by an amount that was enough to split in half and say that's from the newer architecture and that's from the HBM advantage. In fact there were occasions where they were on an even pegging.
That does not stack up.
Compare the Titan Xp to the Titan V in DX11 benchmarks and core for core the Xp is probably edging it.
The Titan Xp uses GDDR5X and the Titan V HBM2.
HBM is poor for gaming which needs high framerates not massive data transfer.
To get the most out of the Titan V you need to run it @2160p with max settings which plays to the strengths of HBM2, if you run it at 1080p you are going to be very bottlenecked. My Titan Xp cards are faster at the Orange room VR bench for example and my older Titan P cards are a little bit faster again.
HBM2 is very nice for gaming @2160p but a bottleneck at 1080p.
Caporegime
- Joined
- 18 Sep 2009
- Posts
- 30,505
- Location
- Dormanstown.
The 1080ti says "hi"
not bothered with this place for ages and the first thing I see is that you are still spouting this crud.
And back to the topic at hand, the DSBR and Prim shaders are working in hardware. They have it working in some vulkan apps, in openGL as shown by the gains over Fiji radeon pro duo and on Metal2.
The main problem is that it is taking a considerable driver rewrite to get it working properly and with DX11.
Associate
- Joined
- 31 Oct 2012
- Posts
- 2,243
- Location
- Edinburgh
Which of latency or throughput do you think is worse? As these are all that exist from the GPUs perspective.
Yep AMD are still working on the drivers that were supposed to work wonders for the Fury X.
I think these days more people are realising that HBM1 on the Fiji cards was a mistake.
All I understand from testing is every HBM equipped card I have used all suffer from the same problem, they are not at the best at 1080p where the fps are high but are great at 2160p where the fps are lower.
I was talking about drivers for DSBR and Prim shaders.
Using mantle, vulkan and dx12 already showed ages ago that HBM is not the issue with GCN. As stated many times in the past.
Soldato
- Joined
- 25 Sep 2009
- Posts
- 10,215
- Location
- Billericay, UK
Not really it's just a different approach (but yes Nvidia are at a generation ahead of AND at moment on metrics like performance per mm2 or TDP). The Titan V is something like 800mn square, yields on that chip are probably very low so why not break it into 4x 200mm square chips and stitch them together? Seems like a sensible approach on paper imo.
HBM1 is the problem on the Fury X and thankfully more people are agreeing with this these days.
If AMD had put 8gb on the cards they would still chug a bit at 1080p but they would have been fantastic at 2160p.
There is nothing worse than running out of memory at high resolution on a high end card regardless of what API they are using.
The HBM2 on the Titan V overclocks really well which is a good thing as even a +220mhz increase is not enough to get all of the performance out of the card at 1080p.
HBM2 is nice providing you are running the right software at high resolution but it is not good and was never intended to be at 1080p on a compute card like the Titan V.
I am pretty sure that all the proper Volta gaming cards if there are going to be any will pack GDDR5(X) or GDDR6 which I think will not throttle the cards at 1080p.
Permabanned
- Joined
- 12 Sep 2013
- Posts
- 9,221
- Location
- Knowhere
I haven't seen anything. But I'd also like to know if anyone's heard real news with a reliable source. If you're referring to the claims this thread was started on I'd bet money on it being a load of rubbish. If anything I'm hoping AMD are trying to get Navi released sooner.
Last edited:
Prim shaders are active, as the way Vega works now in hardware requires it. But the thing that is not active is NGG fast path. This is still being worked on as the driver is being rebuilt from the ground up. Part of the work on this is shown by the hiring of a new lead shader compiler programmer.
Once the above is done, the DSBR will also be able to operate with the new Work distributor, allowing the ROPs to remain constantly filled with tiles from any screen space coordinates instead of the native quadrants they cover.
Both should provide a reasonable if not considerable performance bump when fully active. At the moment the DSBR and Prim shaders are only working in native mode, which is why Vega is no better than Fiji with higher clocks. The DSBR when active is reducing required bandwidth and power, but working sub optimally as it cannot make use of the new work distribution engine without NGG also being active.
There is no problem with the hardware itself, the driver side of things is just taking considerably more work.
Once the above is done, the DSBR will also be able to operate with the new Work distributor, allowing the ROPs to remain constantly filled with tiles from any screen space coordinates instead of the native quadrants they cover.
Both should provide a reasonable if not considerable performance bump when fully active. At the moment the DSBR and Prim shaders are only working in native mode, which is why Vega is no better than Fiji with higher clocks. The DSBR when active is reducing required bandwidth and power, but working sub optimally as it cannot make use of the new work distribution engine without NGG also being active.
There is no problem with the hardware itself, the driver side of things is just taking considerably more work.
Permabanned
- Joined
- 23 Apr 2014
- Posts
- 23,552
- Location
- Hertfordshire
They cant even meet Vega demand, never mind a refresh already.