AMD Vega 10, Vega 11, Vega 12 and Vega 20 Rumors and Discussion

$1200 for a $400 GPU, a 400W PSU and a thunderbolt controller?

Those guys are mental...
A prosumer workstation card with Apple brand. Not surprised. Unclear what the product support looks like, but that could be a fair price for a pro product. Just look at WX and Quadro cards.

Still a bit cringeworthy.
 
A prosumer workstation card with Apple brand. Not surprised. Unclear what the product support looks like, but that could be a fair price for a pro product. Just look at WX and Quadro cards.

Still a bit cringeworthy.
It's not a Pro card, though. It's a Vega 56, not a WX8200...
 
Memory controllers aren't the issue, it's the caching and culling/sorting/prediction algorithms. Essentially a cache miss issue where memory latency is a factor in addition to having the miss. Hard to know precisely, but it ties into the geometry bottleneck. Somewhere there is a critical path that was likely missed and accounts for the higher voltages and power usage. Only takes one weak link to slow everything down. In theory that could be somewhat easily fixed in newer Vega variants, but no indication of that so far. Not unless Vega20 or the new mobile version is uncharacteristically fast.

Computerbase did test RX480 and it says the opposite, the memory bandwith is issue with Polaris.

https://forums.anandtech.com/threads/how-memory-frequency-starved-is-polaris.2479648/

https://forums.anandtech.com/threads/how-memory-frequency-starved-is-polaris.2479648/#post-38348803
 
Computerbase did test RX480 and it says the opposite, the memory bandwith is issue with Polaris.
If the entire workload fit into the cache practically zero memory bandwidth would be required. The issue would be the geometry pipeline stalling and making memory the bottleneck due to cache misses. Then consider the effect of any bandwidth saving techniques.

The newer memory types aren't anything special. Just typical process improvements applied to the prior version. The question is why one card requires more bandwidth than another.
 
If the entire workload fit into the cache practically zero memory bandwidth would be required. The issue would be the geometry pipeline stalling and making memory the bottleneck due to cache misses. Then consider the effect of any bandwidth saving techniques.

The newer memory types aren't anything special. Just typical process improvements applied to the prior version. The question is why one card requires more bandwidth than another.

if that were true, no high-performance GPU needs a bus larger than 256bit and Nvidia could save some money on GTX 1080/1080Ti memory as well as AMD on Vega 64/56 ...
 
if that were true, no high-performance GPU needs a bus larger than 256bit and Nvidia could save some money on GTX 1080/1080Ti memory as well as AMD on Vega 64/56 ...
They simply wouldn't if the caches were that large. Caches and register files have a huge amount of bandwidth relative to actial memory. It's the culling, reuse, and compression schemes, or lack of effective implementation, that can make a huge difference. Check out Nvidias tiled rasterizer for example.
 
They simply wouldn't if the caches were that large. Caches and register files have a huge amount of bandwidth relative to actial memory. It's the culling, reuse, and compression schemes, or lack of effective implementation, that can make a huge difference. Check out Nvidias tiled rasterizer for example.

too much "if" ....
 
From Anandtech's live blog of AMD's Next Horizon event:

12:21PM EST - First 7nm GPU in the industry

12:21PM EST - Additional specs will be coming for MI60

MI60?

If the MI25 was a single Vega with 12.3 TFLOPs FP32 / 24.6 TFLOPs FP16, going by the same logic the MI60 should be ~30 TFLOPs FP32 / 60 TFLOPs FP16.

Maybe it's a 2x Vega 20 card with each chip clocked at 1.8GHz?
And with 4x HBM2 stacks, we could be looking at cards with a whopping 64 GB VRAM.
 
anandtech quotes:
7.4 TGLOPS FP64

14.7 TFLOPS FP32

118 TOPS INT4

if double fp16, it would be close to 30TF fp16. Which would require 2 of them for 60TF, which doesn't really make sense to me.
 
1.8ghz at 300w ? We're not that far under Vega 10, I was expecting better, even if it's not "just" a die shrink...
 
2hBpr7G.png
 
1.8ghz at 300w ? We're not that far under Vega 10, I was expecting better, even if it's not "just" a die shrink...

Pushing for the upper range of performance runs into decidedly non-linear penalties. The promised gains in circuit performance iso-power were stated as being ~1.25x, which if comparable to earlier slides from AMD is a downgrade from ~1.35x. Something more fundamental to the architecture would need to change to do more than incrementally shift where the chip hits a wall.
 
Pushing for the upper range of performance runs into decidedly non-linear penalties. The promised gains in circuit performance iso-power were stated as being ~1.25x, which if comparable to earlier slides from AMD is a downgrade from ~1.35x. Something more fundamental to the architecture would need to change to do more than incrementally shift where the chip hits a wall.
Exactly my thoughts, like I said, they would need to significantly change the arc in order to scale the frequency more than Vega 10. Also it seems power efficiency is rather bad at the core of the Vega architecture, 300W @1.8GHz on 7nm is not really that big of an improvement after all.
 
Much more attractive is Vega 10 performance at 150 watts. With some Navi improvements is what next gen consoles will get. On the other hand is sad to think they need 7nm to get with their current architecture the perf/watt of Pascal with 16nm.
 
There was some speculation concerning the Apple A12's clock speeds on 7nm, which may have hinted at somewhat less power or clock scaling than originally expected. Wire delay and resistance was something everyone was wary of, and it's possible that Vega's architectural targets feel those effects more keenly--given that Vega's changes at 14nm were notably underpinned by trying compensate for wire delay or reach higher clock speeds.

Additionally, there's a wider and faster memory subsystem and a decent amount of additional IO. There's a power hit when those features were added to Nvidia's enterprise lines as well.
 
300W, 7nm, +300MHz - guess they shove this laughable GPU out of their doors solely because Hawaii got really rusty.

Seeing the predictions "north of 2GHz"... oh my, god. Vega simply has to be clocked to ~1.2GHz to be effective. How would a 7nm Fiji do? The same I guess.
 
AMD is comparing V100 PCI-E (250W) vs MI60 (300W). Both cards trade blows in a number of benchmarks. Though they avoided the comparison with the 10% faster V100 NVLink (300W).

DrWUNzkWkAMH3eG.jpg


DrWUNzmX0AAgxXX.jpg
 
Back
Top