I think they misslead what presentation benchmark is and real life benchmarks are .. Many "IF" and "maybe" in this article
AMD Vega 10, Vega 11, Vega 12 and Vega 20 Rumors and Discussion
I think they misslead what presentation benchmark is and real life benchmarks are .. Many "IF" and "maybe" in this article
Any chance of english summary?
Aerows
Regular
If AMD has been more tense about leaks these days, I'd say they are doing fairly OK.
I'd personally like some consumer based products to evaluate it, not to mention prices.
As, "caffeinated" incarnated, I want some real news from AMD. We don't have a card yet, but the mantra is "Vega is coming soon!"
Now what exactly will this Vega that is coming provide?
I'd personally like some consumer based products to evaluate it, not to mention prices.
As, "caffeinated" incarnated, I want some real news from AMD. We don't have a card yet, but the mantra is "Vega is coming soon!"
Now what exactly will this Vega that is coming provide?
ImSpartacus
Regular
I haven't seen get posted here yet about potentially the next Vega.
I was previously of the opinion that Vega 11 would be "half Vega 10" (i.e. 32 CUs, 1 stack HBM2, etc) and replace Polaris 10 in the mobile/semi-custom space (e.g. Apple-bait). However, I'm wondering if it makes more sense for it to use GDDR6 based on info that I came across.
Of course AMD would eventually transition from GDDR5 to GDDR6, but a LinkedIn job description makes it appear sooner than I would've expected.
This senior AMD fella has had "New DRAM Controller : GDDR6" in his job description since September 2015. Comparatively, this guy was working on HBM1 in July 2014 and that shipped barely 12 months later in Summer 2015.
https://www.linkedin.com/in/daehyun-jun-bb135a22/
My guess is that AMD could be ready to implement GDDR6 effectively as soon as suppliers hit their stride in early 2018. To my knowledge, the only GPU that could arrive in the 2018-ish time frame is Vega 11.
If Vega 11 XT sports 12 Gbps GDDR6 (as per SK Hynix's Q2 2017 lineup) on a 256-bit bus (pure speculation), then that gives it a 384 GB/s to play with. That's enough to fuel something like 48-ish CUs. That's staunchly in 1017 territory. That would fill in the last gap in their lineup on the desktop gaming side while still fitting on MXM-B.
I was previously of the opinion that Vega 11 would be "half Vega 10" (i.e. 32 CUs, 1 stack HBM2, etc) and replace Polaris 10 in the mobile/semi-custom space (e.g. Apple-bait). However, I'm wondering if it makes more sense for it to use GDDR6 based on info that I came across.
Of course AMD would eventually transition from GDDR5 to GDDR6, but a LinkedIn job description makes it appear sooner than I would've expected.
This senior AMD fella has had "New DRAM Controller : GDDR6" in his job description since September 2015. Comparatively, this guy was working on HBM1 in July 2014 and that shipped barely 12 months later in Summer 2015.
https://www.linkedin.com/in/daehyun-jun-bb135a22/
My guess is that AMD could be ready to implement GDDR6 effectively as soon as suppliers hit their stride in early 2018. To my knowledge, the only GPU that could arrive in the 2018-ish time frame is Vega 11.
If Vega 11 XT sports 12 Gbps GDDR6 (as per SK Hynix's Q2 2017 lineup) on a 256-bit bus (pure speculation), then that gives it a 384 GB/s to play with. That's enough to fuel something like 48-ish CUs. That's staunchly in 1017 territory. That would fill in the last gap in their lineup on the desktop gaming side while still fitting on MXM-B.
If these slides were legitimate (and still accurate), there will be Vega 11 products available this year:
https://videocardz.com/65521/amd-vega-10-and-vega-20-slides-revealed
https://videocardz.com/65521/amd-vega-10-and-vega-20-slides-revealed
D
Deleted member 13524
Guest
As kalelovil stated, Vega 11 is supposed to be coming this year. AMD is naturally developing GDDR6 controllers, but it could only be used in 2019 for Navi.
ImSpartacus
Regular
I'm familiar with those slides, hence my previous belief that Vega 11 would be half of Vega 10. To my knowledge, those slides are the only Vega 11 info out there.
But since half of 2017 is effectively over and it's been radio silence on that chip, my gut is that Vega 11 is delayed.
Combine that with the likely shortages of fast HBM2 and I can see a path where it makes sense for AMD to utilize SK Hynix's initial GDDR6 supply in early 2018.
D
Deleted member 13524
Guest
It would, but GDDR6 may be coming too late for Vega 11.
Why are shortages of HBM2 likely?
Besides, typical design-to-production cycle for AMD's chips is around 3 years. I doubt they would drastically change the memory controller mid way.
D
Deleted member 13524
Guest
AMD results could be somehow somewhere interpreted to be misleading by someone.
But hey, at least they start at zero.
Well, the 7nm plans probably changed, as in the roadmap slides we see 14nm+ for Vega rather than 7nm.
But I don't imagine they abandoned Vega 11.
gamervivek
Regular
Maybe AMD found they could get good enough results with the 14nm+ vs. 7nm and brought ahead the Vega 20 and that Volta is coming soon enough.
There hasn't been any talk of Vega 11 from AMD themselves and whether it's even real. Also the chip that goes in Scorpio can be a decent upgrade on Polaris 10 as well if its frequency scales like PS4 pro's vs RX 480 and probably 48ROPs(?), so if Vega 11 is the smaller chip there isn't much room to wiggle into. Though the scorpio chip would be on TSMC's process.
There hasn't been any talk of Vega 11 from AMD themselves and whether it's even real. Also the chip that goes in Scorpio can be a decent upgrade on Polaris 10 as well if its frequency scales like PS4 pro's vs RX 480 and probably 48ROPs(?), so if Vega 11 is the smaller chip there isn't much room to wiggle into. Though the scorpio chip would be on TSMC's process.
You're comparing apples with oranges, consoles use APUs, which are designed from the ground up to be just that, they don't have "Polaris 10" or "Pitcairn" or whichever GPUs have been suggested.
AMD has confirmed Vega 11's existence through their OpenCL-drivers
Anarchist4000
Veteran
There's barely been any talk of "Vega10" either since initial roadmaps. AMD has been referring to all SKUs as Vega.
It could be nothing is delayed(HBM2 rumors), but waiting on the official HDMI2.1 spec to be released 2Q17. So any time now. That would have the 4k120, variable refresh, and HDR they keep indirectly showing that would be useful for gamers and probably upcoming VR headsets.
Their main point of argument is IMHO, that AMD supposedly added the series of individual results in milliseconds instead of providing a geometric mean. Then there's different results for Pascal-cards at Baidu and AMD allegedly did not pick the fastest one. Lastly and mainly in comparison to Intel, that the libraries used for KNL are not accelerating every workload, which was pretty obvious when Intel showed their KNL-results, that there are optimizations missing (link from the heise article: https://software.intel.com/en-us/ar...nce-for-deep-learning-and-getting-better-fast).
That said and judging from the officially available results at github and (guesstimating from) the screenshot from AMDs FAD (courtesy of heise.de), I fail to see where there is a dramatic difference.
Yes, the 288 ms for the M40 seem to be wrong, since they should read 226 ms (it's 288 including column "n" [Forward (msec)], but this is included for M40 only). And yes, the DeepBench results seem to be added "Total Time" (column T in Baidu's github results) only and yes, that part of the benchmark seems to use only FP32 according to the github description.
But WRT the comparison Vega vs. Tesla P100 (PCIe), I do not see anything dramatically wrong here, if you do not insist on geometric means.
Yes, the 288 ms for the M40 seem to be wrong, since they should read 226 ms (it's 288 including column "n" [Forward (msec)], but this is included for M40 only). And yes, the DeepBench results seem to be added "Total Time" (column T in Baidu's github results) only and yes, that part of the benchmark seems to use only FP32 according to the github description.
But WRT the comparison Vega vs. Tesla P100 (PCIe), I do not see anything dramatically wrong here, if you do not insist on geometric means.
Last edited:
D
Deleted member 2197
Guest
I think the point is the way Baidu developed the benchmark is for the result to be based on geometric means. I guess it really doesn't matter since anyone who is interested in using products in this space will know how to interpret the benchmark.
Silent_Buddha
Legend
Would like to hear the wording that she uses. Did she say that it'll be launching "in the next 2 months" or "2 months after" FE?
Would have been nice for that poster to have at least attempted to quote what she actually said as apparently other people heard something different. There aren't too many combinations of words that could be misinterpreted in such a way, IMO, unless you aren't a native English speaker.
Regards,
SB
Would have been nice for that poster to have at least attempted to quote what she actually said as apparently other people heard something different. There aren't too many combinations of words that could be misinterpreted in such a way, IMO, unless you aren't a native English speaker.
Regards,
SB
The quote seems to be
The next couple of months is from now to end of July. We knew FE is mid-late June and were hoping RX Vega in July sometime. Nothing about that quote changes that.
Ah the internet.....
Queue inflaming WCCFTech "article" then pharma posting that here as well.
The next couple of months is from now to end of July. We knew FE is mid-late June and were hoping RX Vega in July sometime. Nothing about that quote changes that.
Ah the internet.....
Queue inflaming WCCFTech "article" then pharma posting that here as well.