AMD Vega Hardware Reviews

I've been impressed with evga's quality in video cards (ball bearing fans ftw) and PSUs. Haven't had one of their MBs though.
There was the small issue of cards catching fire with 1070/1080, though
 
There was the small issue of cards catching fire with 1070/1080, though

But that was happening with Nvidia in general, no? Or was that the previous fire hazzard?
 
Ouch. I've got six EVGA 1070s running in a crypto rig. Any linkage on which items like to burst into flames?
 

So essentially a hot GTX 1080. Now it's down to pricing and mining benchmarks.

From AMD's perspective this release (plus Threadripper) should be enough to prevent the Barclay's $9 share price as it adds product in a tier where AMD had none. My bet is that the mining numbers (H/s coupled with power draw) will not be attractive enough to replicate the RX400/500 boom so AMD will see a drop in GPUs until they can get something better out.
 
I'm pretty sure nVidia can cut 1070 and 1080 prices, it's kind of "old" right now for them. But at they same time, they don't need to, Vega not beeing a threat. I was very attracted by Vega when the tech aspect of it was revealed, I was thinking about upgrading from my Fury X, but what a letdown by the final product.
I hope reviewers will have enough time with the cards, so we understand what went wrong.
 

The liquid version is on par with a custom1080 at 1080p, 15% faster at 1440p and 10% at 4k. Time spy is the worst result with 1080 card ahead.

Would've been good enough if it were the 1600Mhz version or if the Ti wasn't that much faster this time.

As a prospective buyer I'd be leery of the overvolted memory, AMD cards have a habit for black screen memory errors.
 
The liquid version is on par with a custom1080 at 1080p, 15% faster at 1440p and 10% at 4k. Time spy is the worst result with 1080 card ahead.

Would've been good enough if it were the 1600Mhz version or if the Ti wasn't that much faster this time.

As a prospective buyer I'd be leery of the overvolted memory, AMD cards have a habit for black screen memory errors.

MSI GTX 1080 is one of the slower custom designs.

EDIT:

I'm pretty sure nVidia can cut 1070 and 1080 prices, it's kind of "old" right now for them.

Of course and they would still earn a lot of money by it. With 314mm² the GTX 1080 is a much smaller chip in comparison to the RX VEGA 64 with it's 484 mm².
 
Last edited:
Good start, but would be nice to have a white paper explaining where all that SRAM went. There is enough missing for a whole 2nd chip in there. Most cache sizes in the ISA document were inline with previous generations, except the 2MB larger L2.

I tried to find another GPU slide where AMD gave a figure for MB of SRAM per die, without much luck so far.
We may need to refrain from being to worried about that figure in the absence of data from other AMD GPUs.

The wafer shot showed command processors and geometry front ends that appear to have some markedly visible SRAM arrays, and a few extra blocks with at least some SRAM of their own. Doubling L2 capacity means doubling the number of tags, AMD indicated much of its transistor growth was invested in extra clock speed and measures like buffers (SRAM) needed for compensating for the extra speed.

The Vega ISA has significantly expanded the number of memory instructions it can have in-flight, which may have implications as to how many buffers there are for descriptor information and miss handling resources, given that GCN tosses a fair amount of data to the L/S section. That is an area where AMD decided to change the layout for, and may have added more hardware in the process.

There's a band of what AMD indicated in its power management controllers slide to possibly be a major component of the on-die infinity fabric, and each block appears to have at least some SRAM, perhaps with routing information and transaction information.
The memory controllers/HBCC would need to maintain a fair amount of queues and state.
The virtualization section from AMD's slide is likely dominated by storage, and there's a security section that likely is an ARM A5 and secure memory good for hundreds of KB at least. The video block may be good for some fraction of a MB, etc.
If we start thinking each extra block is good for some fraction of a MB per section, or is tens of extra KB over 64 CUs, it would start adding up.
 
I'm pretty sure nVidia can cut 1070 and 1080 prices, it's kind of "old" right now for them. But at they same time, they don't need to, Vega not beeing a threat.

Vega is a threat at the current prices, in a way there are perfectly valid reasons to go with a RX Vega 64/56 instead of a 1080/1070.
But at least the 1070 (not aware of G5X prices) should be able to go down to sub-$300 and still allow nvidia to get a very nice profit.
 
MSI GTX 1080 is one of the slower custom designs.

EDIT:



Of course and they would still earn a lot of money by it. With 314mm² the GTX 1080 is a much smaller chip in comparison to the RX VEGA 64 with it's 484 mm².

It might be, but still the liquid-cooled version is ~20% faster than the 1080FE while 1080Ti FE is at 30%, a performance disparity that I expected with the 1.6Ghz Vega. Maybe the liquid cooled card doesn't hold the 1.75Ghz boost consistently enough?

Otherwise AMD look to be heavily memory bandwidth constrained, a re-release later with better HBM2 should do enough for a rebrand without core clock increases.
 
Back
Top