• 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 confirmed for 2017 H1

Status
Not open for further replies.
D.P. said:
Really, there is no other explanation. HBM2 is late.
Click to expand...

Sums it up perfectly really. Raja mentioned in the AMA that they're sourcing from Samsung and Hynix.
The good news is that unlike HBM1, HBM2 is offered from multiple memory vendors – including Samsung and Hynix
Click to expand...

I was actually surprised to see him say Vega is 16GB and using two modules. That's two 8GB modules; something that wasn't even on the Hynix catalogue from 2016, or the latest 2017 ones.

So it would seem that for RX Vega they're going to source from both companies for the 8GB model graphics card at least; and at 480GB/s bandwidth that means the modules are running at around 1.88Gbps. So that's faster than the listed 1.6Gbps at least, but still not the 2.0Gbps they were planning on using; and which Hynix listed for Q3 2016 production.

Even then, still surprised they managed to get 8GB modules at 1.88Gbps.
I just hope them sources from Samsung doesn't mean the consumer Radeon RX Vega cards will be using Samsung's 1.4Gbps 4GB modules.
 
D.P. said:
Really, there is no other explanation. HBM2 is late.
Click to expand...

Or someone else has a secret deal for large quantities of HBM2 ;)

N19h7m4r3 said:
Sums it up perfectly really. Raja mentioned in the AMA that they're sourcing from Samsung and Hynix.


I was actually surprised to see him say Vega is 16GB and using two modules. That's two 8GB modules; something that wasn't even on the Hynix catalogue from 2016.

So it would seem that for RX Vega they're going to source from both companies for the 8GB model graphics card at least; and at 480GB/s bandwidth that means the modules are running at around 1.88Gbps. So that's faster than the listed 1.6Gbps at least, but still not the 2.0Gbps they were planning on using; and which Hynix listed for Q3 2016 production.

Even then, still surprised they managed to get 8GB modules at 1.88Gbps.
I just hope them sources from Samsung doesn't mean the consumer Radeon RX Vega cards will be using Samsung's 1.4Gbps 4GB modules.
Click to expand...

I dunno why - maybe due to supply but even that doesn't entirely make sense, as I said awhile back both AMD and nVidia have spent quite a bit of time exploring 2 stack options - I wondered if it was problems with the interposer or something - maybe its just because they are for various reasons having to make such big cores for the higher end parts - neither Vega or Volta are particularly small for the bigger variants with Volta clocking in at around 800mm2 and looking like Vega is going to be some of the biggest chips AMD have ever made also.
 
Rroff said:
I dunno why - maybe due to supply but even that doesn't entirely make sense, as I said awhile back both AMD and nVidia have spent quite a bit of time exploring 2 stack options - I wondered if it was problems with the interposer or something - maybe its just because they are for various reasons having to make such big cores for the higher end parts.
Click to expand...

Well NVIDIA have been using 3 and 4 stacks for their Tesla P100s and new Quadro GP100's; and it's all the slower Samsung 1.4Gbps modules.

Raja saying they're considering a 16GB Radeon card is nice, and hopefully they manage to get it sorted; especially if they manage 4 stacks of HBM2; that would give amazing bandwidth at 960GB/s on current modules.
 
shankly1985 said:
TBH Kap while I think its excellent what you do on this forum you do the best Benchmark threads.. I have spoken quite a lot of my hate towards just still image or numbers.. For me Raw gameplay will always tell the biggest picture. Its 2017 benchmarking of games has moved on for me!! We now need a new way of benchmarking games!

Here is the 295x vs TitanXp running on DX12 and Ryzen CPU, tells a bigger picture dont it :D Plus its also known that the Tomb Raider bench dont added up to the actual gameplay of the game. So can we ever trust build in benchmarks?
Click to expand...


TPU are also very respected for their reviews

3EukugC.jpg

The difference is TPU are not using an AMD CPU.
 
N19h7m4r3 said:
Well NVIDIA have been using 3 and 4 stacks for their Tesla P100s and new Quadro GP100's; and it's all the slower Samsung 1.4Gbps modules.

Raja saying they're considering a 16GB Radeon card is nice, and hopefully they manage to get it sorted; especially if they manage 4 stacks of HBM2; that would give amazing bandwidth at 960GB/s on current modules.
Click to expand...


Vega wont be able to handle 4 stacks, not without a decent redesign. Iis not a case of simply wiring up 2 extra modules. The entire memory interface would need to get duplicated and the changes passed through to the rest f the chip design. Certainly feasible but requires design work, new spin out, larger chip etc. Then the question is does Vega actually need that much bandwidth, would it make any real difference? I would rather hope AMD made their memory usage much more efficient rather than go for yet another brute force design. And if Vega really needs that kind of bandwidth to shine it would have been designed for 3/4 stacks form the start. Best case scenario is Vega would have had the same bandwidth as the FuryX
 
D.P. said:
Why wasn't vega released last year to have some hope of competing for market share against Pascal?
Click to expand...

It's a shame AMD didn't also create a Vega card to use GDDR5X. Back then they showed a Vega engineering sample running the same speed a 2.1Gh GTX 1080 in Doom.

Had they launched a Vega card with 8GB GDDR5X it would have done well; but alas they put too much into HBM2 and got burnt when their partners failed to deliver on the Q3 2016 production dates.
 
Chuk_Chuk said:
Did AMD ever say it was supposed to be released last year? If they didn't then the answer is because AMD is working to their schedule not Nvidia's.
Click to expand...

Doesn't matter what AMD says. AMD can work to whatever schedule they want but the market doesn't sit around waiting and the shareholders aren't happy loosing sales.
 
N19h7m4r3 said:
Navi, multiple gpu cores, seen as a single large GPU all sharing their HBM2 as High-Bandwidth Cache; and allowing for improved SoC of CPU+GPU.

Just like Ryzen is multiple CCX linked via Infinity Fabric, AMD's goal by Navi is the same for GPUs. Raja spoke about in in 2016 where their end goal is for Multi-GPU cores to be as common as multi-CPU.

Infinity Fabric runs at 30-50GB/s on Ryzen, but goes all the way up to 512GB/s on Naples/Epyc and GPUs.

https://youtu.be/4qJj1ViyyPY?t=4m4s
Click to expand...

Thanks, much appreciated!
 
Status
Not open for further replies.
Back
Top Bottom