Sure. Hence the 'eh' part of my comment.
-
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.
AMD Vega Architechture
- Thread starter jjgreenwood
- Start date
I do too, generally. I'd expect something between maybe 330-380mm^.
But I raise the possibility of 300mm^ depending on how much more efficiency Vega actually brings to the table. They may get away with a similar competitive chip, but at a slightly smaller size IF Vega is another decent step up from Polaris. I'm personally not expecting it to make that much of a difference, but who knows.
Caporegime
- Joined
- 18 Oct 2002
- Posts
- 33,188
Something I said about Fury before it was launched, most architectures are memory limited, almost every chip is designed to work within the confines of it's memory access both bandwidth and latency. Fury was going to get a huge boost in bandwidth but the actual architecture, effectively a small update from GCN 1.0, wasn't designed to utilise that much bandwidth.
Vega is more likely to be an architecture designed to utilise higher memory bandwidth. If you have more memory bandwidth you can change how data is accessed and used, you might order processing data differently, beef up the front end drastically as it can request much more data then rearrange the shaders to deal with that much more information.
Internal bandwidth efficiency tests for Hawaii are exceptional, it gives a theoretical vs utilisable bandwidth, Hawaii is near 100%, Fury was at 60-70% or something. Again the internal bus for moving around data was potentially simply not designed to move that much data. Think back to R600, it had a 1024bit internal ringbus and it turned out to be WAY more than it needed and ahead of it's time, with 3870 I believe it was knocked down to 512bit. But these things are part of your architecture design, it's perfectly possible Fury had the reverse and simply didn't have the ability to utilise that much bandwidth, as in GCN 1.0 chips effectively had the equivalent of a 512bit ringbus but needed a 1024bit ringbus and as such became a limit.
An architecture designed around the specific ability of HBM2 is even more likely because it looks like every chip in the Vega range will use HBM2. Having an architecture need to work and scale with both HBM2 and gddr5 and way different bandwidth/shader ratio is not going to get a great balance.
Fury is good, it used the power advantage of HBM and it made the card way better than it would have been and like GP100, it was a necessary step for a company to learn what they need to then produce higher volume parts, but architecturally it wasn't optimised at all for massive bandwidth.
I think Vega is going to be a very large step and the performance/W charts AMD showed for Vega and considering the stated performance/w gain for RX470 vs 270x, it's going to be very very big.
Vega is more likely to be an architecture designed to utilise higher memory bandwidth. If you have more memory bandwidth you can change how data is accessed and used, you might order processing data differently, beef up the front end drastically as it can request much more data then rearrange the shaders to deal with that much more information.
Internal bandwidth efficiency tests for Hawaii are exceptional, it gives a theoretical vs utilisable bandwidth, Hawaii is near 100%, Fury was at 60-70% or something. Again the internal bus for moving around data was potentially simply not designed to move that much data. Think back to R600, it had a 1024bit internal ringbus and it turned out to be WAY more than it needed and ahead of it's time, with 3870 I believe it was knocked down to 512bit. But these things are part of your architecture design, it's perfectly possible Fury had the reverse and simply didn't have the ability to utilise that much bandwidth, as in GCN 1.0 chips effectively had the equivalent of a 512bit ringbus but needed a 1024bit ringbus and as such became a limit.
An architecture designed around the specific ability of HBM2 is even more likely because it looks like every chip in the Vega range will use HBM2. Having an architecture need to work and scale with both HBM2 and gddr5 and way different bandwidth/shader ratio is not going to get a great balance.
Fury is good, it used the power advantage of HBM and it made the card way better than it would have been and like GP100, it was a necessary step for a company to learn what they need to then produce higher volume parts, but architecturally it wasn't optimised at all for massive bandwidth.
I think Vega is going to be a very large step and the performance/W charts AMD showed for Vega and considering the stated performance/w gain for RX470 vs 270x, it's going to be very very big.
Last edited:
Did a quick search but couldn't find a thread on this. I'n not overly knowledgeable about the technical details of graphics cards but thought it may be fun to ask the question of those that know more. Apparently Vega is due in October I was wondering if it's possible to estimate the probable performance of the card given what we know of Polaris.
This has been doing the rounds suggesting the specs of the cards:
And this has been posted as suggesting the performance of Polaris:
Given the performance delta of last gen's FuryX vs the older 290X:
Is it possible to make a relatively safe assumption that performance is likely to have a similar delta for the high end Vega part due later in the year.
Presumably AMD will want this part to hit 1080 levels of performance at least, and on the surface of things it appears they will need quite a significant improvement in clock speeds & architecture to get them there.
i just noticed P10 and P11 are wrong in that chart.
2306 Shaders and 896 Shaders.
https://www.techpowerup.com/gpudb/2849/radeon-rx-460
https://www.techpowerup.com/gpudb/2848/radeon-rx-480
Soldato
- Joined
- 19 Feb 2011
- Posts
- 5,849
I think that would make more sense than the Polaris, probably a good bet it would either use GDDR5X rather than HBM2 as well, unless Microsoft dont care about the price of the thing.
MS very much cares about pricing - why it went with AMD apu; as price and performance nothing could match it. but time frame fits possibly with using Vega arch; just depends on how much power it uses otherwise would be smarter to go with Polaris......specially if the numbers are true on it....
Pardon me if I'm mistaken, but isn't there already an official Vega thread? Or will this become the new Vega thread? With 200+ pages of posts that have nothing to do with the topic at hand lol...
Anyhow, the mid 2017 releases look reasonable for Vega, but that's an awful lot of time to let Nvidia soak up the higher-end sales. Looks like it'll be roughly a year till I buy a new GPU for my rig... I guess money saved to spend on other stuff then?
Anyhow, the mid 2017 releases look reasonable for Vega, but that's an awful lot of time to let Nvidia soak up the higher-end sales. Looks like it'll be roughly a year till I buy a new GPU for my rig... I guess money saved to spend on other stuff then?
Big Vega is early to mid - I'm guessing possibly more early than we believe - rumors that little Vega has been pulled forward to Oct - possibly we could see 3 versions of Vega......little vega with gddr - little vega with hbm2 and big vega with hbm2......
Tbh, I don't think that's at all unexpected. Isn't the 1080 314mm^2?
But yes I'd expect Vega to be little bigger than that to not be too close to Polaris 10.However, if it has p/w benefits anywhere near that chart they showed l, they could size it closer to polaris and still have a significant performance delta.
Last edited:
Associate
- Joined
- 1 Oct 2014
- Posts
- 100
LoL ,,,, there's a new Hype Train in town. All Aboarddddd.