GPU Rumors

[John P. Myers]

2377

Though rumors, these are realistic expectations. Remember they're going from 28nm to 16nm GPUs which is a monstrous leap. Final specs could change, but this should be ballpark anyway. The GDDR5 in the X80 and X80Ti may turn out to be GDDR5X. Remains to be seen.

The X80, X80Ti and X80 Titan are not official product names. Those are not yet known.

[DrPop]

Man, those look nice. First BIG upgrade in GPU tech I've seen in a while...will make my 7970 seem long in the tooth. ;) Wonder what the price point will be, though....

[nanoprobe]

Man, those look nice. First BIG upgrade in GPU tech I've seen in a while... Wonder what the price point will be, though....

Your first born. =))

[Dave gpu]

2377

Though rumors, these are realistic expectations. Remember they're going from 28nm to 16nm GPUs which is a monstrous leap. Final specs could change, but this should be ballpark anyway. The GDDR5 in the X80 and X80Ti may turn out to be GDDR5X. Remains to be seen.

The X80, X80Ti and X80 Titan are not official product names. Those are not yet known.


This was on Collatz Conjecture Message boards


Profile Crunch3r
Volunteer moderator
Project developer
Project tester





Message 22125 - Posted: 21 Mar 2016, 11:29:21 UTC - in response to Message 22117.


How Accurate Are These Specs?

Now that we’ve gotten the specs out of the way there are a few red flags that I should direct your attention towards. The first being that the source of these specs is claiming that the GTX 1080 Ti will feature 8GB of GDDR5 and a 512bit memory interface. All the while the Pascal based Titan, which is based on the very same GP100 GPU, will feature an entirely different memory interface and an entirely different memory standard.

Any person with basic knowledge of chip design will tell you that this will NOT happen. Each GPU will only feature one memory standard, either GDDR5/X or HBM/2, not both. You will not see a GPU with two completely different memory interfaces supporting two completely different memory technologies. It will only support one or the other.

The “Alleged” Memory Specs For The GTX 1080 Don’t Make A Whole Lot Of Sense Either

There’s another red flag here and that this source is claiming that Nvidia will be using GDDR5, rather than GDDR5X for its GP104 powered GTX 1080. Not only that but that it’s also allegedly configured with a 384bit memory interface and 6GB of memory.

For starts it’s unlikely that we’ll see Nvidia use GDDR5 rather than GDDR5X – which offers double the bandwidth – in its high end graphics cards. Secondly, benchlife, a source which we hold in a much higher regard for its impeccable accuracy over the years claims that the GTX 1080 will be configured with 8GB of memory rather than six. Which, credibility aside, is a much more believable figure to begin with.

Considering that GDDR5X will not only offer twice the bandwidth of GDDR5 but also twice the capacity. Eight gigabytes of GDDR5X would require a 256bit memory interface and would deliver 448 GB/s of memory bandwidth. That amount of bandwidth is enough to keep a GPU with 33% more performance than the GTX TItan X happily fed. Even if benchlife had never leaked these memory specs, 8GB of memory is 100% what we would’ve bet on. It simply makes perfect sense for this type of chip and its time to market.


Both of these redflags will prompt us to label the “confidential” table above as very questionable at best.

[John P. Myers]

Those are good points, however not entirely true. First let me say that i did state this was a rumor and Crunch3r could be half or entirely right. Or entirely wrong.

His first point about not using more than 1 memory standard on the same GPU is false. For example, the GK107, GK208, GM107, GF116, etc. were made to use either DDR3 or GDDR5, and could be purchased in either configuration, with different bus widths as well. It's also entirely possible HBM2's supply won't be abundant enough and all new GPUs end up with GDDR5X instead (AMD was going to use HBM2 as well).

I already mentioned GDDR5X would possibly be used instead of GDDR5, as he mentioned as well. The amount of VRAM on the 1080 is something i somewhat agree with him about. 8GB just makes alot more sense, but at the same time i don't have any info verifying it one way or the other. Also it must be remembered that the GP104 is of the same class as the GTX 770 and GTX 680.

Edit: Another thing to point out are his quotes on the projected bandwidth of GDDR5X of 448GB/s. This is true after the process matures, but is not true at all today. Today, at best, you'd get 384GB/s @ 256bit bus width.

[Dave gpu]

Those are good points, however not entirely true. First let me say that i did state this was a rumor and Crunch3r could be half or entirely right. Or entirely wrong.

His first point about not using more than 1 memory standard on the same GPU is false. For example, the GK107, GK208, GM107, GF116, etc. were made to use either DDR3 or GDDR5, and could be purchased in either configuration, with different bus widths as well. It's also entirely possible HBM2's supply won't be abundant enough and all new GPUs end up with GDDR5X instead (AMD was going to use HBM2 as well).

I already mentioned GDDR5X would possibly be used instead of GDDR5, as he mentioned as well. The amount of VRAM on the 1080 is something i somewhat agree with him about. 8GB just makes alot more sense, but at the same time i don't have any info verifying it one way or the other. Also it must be remembered that the GP104 is of the same class as the GTX 770 and GTX 680.

Edit: Another thing to point out are his quotes on the projected bandwidth of GDDR5X of 448GB/s. This is true after the process matures, but is not true at all today. Today, at best, you'd get 384GB/s @ 256bit bus width.

Thanks for your info, looks like this is going to be a work in progress.

[dcushing]

Your first born. =))

Nvidia 1080, available May 2016, is going to run about $259, from what I can find on newegg and other sources.

[John P. Myers]

More info today and this is all 100% confirmed from Nvidia themselves.

In the Tesla GPUs, the GP100 chip has the following specs:

3584 cores (full ungimped chip has 3840 - these specs are for the gimped chip))
21.216 TFLOPS FP16 (FP16 in the past has been exactly equal to FP32)
10.608 TFLOPS FP32
5.304 TFLOPS FP64 (In the past, Nvidia's FP64 rate (when ungimped) has been equal to 1/3 the FP32 rate. Now improved to 1/2)
1328 MHz core clock/1480MHz boost
16GB HBM2 @ 720GB/s bandwidth

Note these are specs for the Tesla version. Consumer GPUs usually have a higher FP32 rating than the Teslas due to higher base clocks so we'll see. This is the same chip that would be used in the 1080Ti and/or Pascal Titan.

[scole of TSBT]

How do the server/data center models stay cool with passive cooling?

[FourOh]

I think a GTX 1080 would more likely start in the $600-650 range, at least until AMD launches their 14nm Polaris lineup. Both are exciting to me... more GFLOPS with lower power usage :D

[dcushing]

More info today and this is all 100% confirmed from Nvidia themselves.

In the Tesla GPUs, the GP100 chip has the following specs:

3584 cores (full ungimped chip has 3840 - these specs are for the gimped chip))
21.216 TFLOPS FP16 (FP16 in the past has been exactly equal to FP32)
10.608 TFLOPS FP32
5.304 TFLOPS FP64 (In the past, Nvidia's FP64 rate (when ungimped) has been equal to 1/3 the FP32 rate. Now improved to 1/2)
1328 MHz core clock/1480MHz boost
16GB HBM2 @ 720GB/s bandwidth

Note these are specs for the Tesla version. Consumer GPUs usually have a higher FP32 rating than the Teslas due to higher base clocks so we'll see. This is the same chip that would be used in the 1080Ti and/or Pascal Titan.

Thats what I read also... Still looking like around $259 for a nvidia 1080.

[John P. Myers]

With some new data today, it appears the GTX 1080 is ~7% slower clock for clock than a 980 Ti. So much for architectural improvements :/ The 1080 also appears to have fewer cores than a 980Ti. However since the 1080 runs at much faster clocks, it will be faster out of the box than a 980Ti. Looks like about 15% faster. The 1080 and 1070 both use the GP104 chip. The 1080 will have GDDR5X and the 1070 will have GDDR5. The real GPU based on the GP100 chip (whatever the Pascal Titan will be called) will be the real flagship. It will use HBM2 but won't be out til Q1 2017 or so. Same goes for AMD's flagship. In the meantime we get these filler cards :p

[John P. Myers]

1080:
2560 Cores
1607MHz Base Clock
1733MHz Boost Clock
180W TDP

8.22784 TFLOPS FP32 at base clock speed - about 30% faster than Titan X

[Mumps]

DP or SP on those TFLOPs? :)

[dcushing]

Watched the live stream on ustream. All hype about the 1080

gtx 1080 extreme will be $599 retail

gtx 1070 a little cheaper by 100

No thanks!

Invested 2 billion in the technology, bullsh*t

[John P. Myers]

DP or SP on those TFLOPs? :)
SP of course :p

And i should probably say these aren't rumors anymore. All confirmed.

[scole of TSBT]

CardMemorySP GLOPSDP GLOPSWattsMSRP
GTX 7702GB3213134230
GTX 7803GB3977166250
GTX 780ti3GB5046210250
Titan6GB45001500250
Titan Black6GB51211707250
Titan Z12GB81222707375

GTX 7502GB10443260
GTX 750ti2GB13064060
GTX 9602GB230872120
GTX 9704GB3494109145
GTX 9804GB4612144165
Titan X12GB6144192250

GTX 10708GB6500203 (1:32 ?)150$379
GTX 10808GB8227257 (1:32 ?)180$599
Tesla P10016GB106005300300(if you have to ask, you can't afford it)



http://www.tsbt.co.uk/forum/download/file.php?id=57

[Al]

Too bad I'm retiring in 515 days 17 hours and 35 minutes (not that I'm counting) because I'd be able to afford a few if I wasn't.

[zombie67]

I would love to see the same comparison with DP.

[scole of TSBT]

I put an updated table in with guess at FP64 being 1:32 like the Maxwell cards. The GTX 1070 and 1080 use a GP104 chip and will not have the same 1:2 FP32:FP64 ratio as the Tesla, which will use a GP100 chip.

[zombie67]

Thanks!

[John P. Myers]

CardMemorySP GLOPSDP GLOPSWattsMSRP
GTX 7702GB3213134230
GTX 7803GB3977166250
GTX 780ti3GB5046210250
Titan6GB45001500250
Titan Black6GB51211707250
Titan Z12GB81222707375

GTX 7502GB10443260
GTX 750ti2GB13064060
GTX 9602GB230872120
GTX 9704GB3494109145
GTX 9804GB4612144165
Titan X12GB6144192250

GTX 10708GB6200 (?)194 (1:32 ?)150$449 ($379 and up custom at later date)
GTX 10808GB8228257 (1:32 ?)180$699 ($599 and up custom at later date)
Tesla P10016GB106005300300(if you have to ask, you can't afford it)



Editing table. The GFLOPS listed for the 1070 are approximate and based off boost clock so i'm changing the 1070 GFLOPS to reflect the base clock like the others. Changed release prices of 1070 and 1080 since Nvidia is charging $100 more for their reference cards (aka Flounderer's Edition) this time around. Lower priced GPUs won't be available at first.

Additional info: GP100 likely will not be used on any consumer GPU and will instead use GP102 for continued crippling of FP64. Some first release GPUs maybe be additionally crippled by having overclocked capped at 2114MHz (2000MHz base). Do not purchase these GPUs (please :) ). There will be GPUs released soon that will run faster than this out of the box, on air, and can be OC'd up to 2400MHz. On water up to 2500MHz.

Of course, when the 1080Ti comes out, it will be as fast as an overclocked 1080 for about the same price. Also let's not forget that BOINC projects sometimes have issues running new architectures at first anyway.

[John P. Myers]

CardMemorySP GLOPSDP GLOPSWattsMSRP
GTX 7702GB3213134230
GTX 7803GB3977166250
GTX 780ti3GB5046210250
Titan6GB45001500250
Titan Black6GB51211707250
Titan Z12GB81222707375

GTX 7502GB10443260
GTX 750ti2GB13064060
GTX 9602GB230872120
GTX 9704GB3494109145
GTX 9804GB4612144165
Titan X12GB6144192250

GTX 10606GB3855120120$299 ($249 and up custom at later date)
GTX 10708GB5979187150$449 ($379 and up custom at later date)
GTX 10808GB8228257180$699 ($599 and up custom)
GTX 1080Ti12GB9385293250 $1000
Pascal Titan X12GB10157317250$1200




@Scole Please steal your chart back :p Just click "reply with quote" then copy paste it from my text

[scole of TSBT]

@Scole Please steal your chart back :pno, no, no...you own that table now and you're keeping it updated far better than I would.

[John P. Myers]

Pascal Titan expected to launch August 2nd

[John P. Myers]

AMD RX 500 series expected to launch in May.

Up to 4,096 cores

Up to 8GB of 2048-bit HBM2

Up to 12.5 TFLOPS FP32, 781 GFLOPS FP64

~225W TDP

[MindCrime]

I'd buy an nvidia card next if they didn't neuter DP as much as they possibly could. I'm not an AMD fanboy so much but Nvidia's business model feels like Comcast and I'll do whatever I can to not give them money.

[John P. Myers]

GTX 1080Ti just announced. 11GB VRAM @ 352 bits. 3584 cores. 1480MHz base clock. 250W TDP $700

[scole of TSBT]

No posts about the AMD Vega GPU?


BaseBoostBaseBoostTBP (W)
Mem.GFLOPS SPGFLOPS SPGFLOPS DPGFLOPS DPAvg $GFLOPS/$GFLOPS/W
RX Vega 568GB828610544518659210???13.839.5
RX Vega 648GB1021512665638792295$700 14.634.6
RX Vega 64 Liquid8GB1151813738720859345$1,00013.733.4