Friday, January 2nd 2015
Possible NVIDIA GM200 Specs Surface
Somebody sent our GPU-Z validation database a curious looking entry. Labeled "NVIDIA Quadro M6000" (not to be confused with AMD FirePro M6000), with a device ID of 10DE - 17F0, this card is running on existing Forceware 347.09 drivers, and features a BIOS string that's unlike anything we've seen. Could this be the fabled GM200/GM210 silicon?
The specs certainly look plausible - 3,072 CUDA cores, 50 percent more than those on the GM204; a staggering 96 ROPs, and a 384-bit wide GDDR5 memory interface, holding 12 GB of memory. The memory is clocked at 6.60 GHz (GDDR5-effective), belting out 317 GB/s of bandwidth. The usable bandwidth is higher than that, due to NVIDIA's new lossless texture compression algorithms. The core is running at gigahertz-scraping 988 MHz. The process node and die-size are values we manually program GPU-Z to show, since they're not things the drivers report (to GPU-Z). NVIDIA is planning to hold a presser on the 8th of January, along the sidelines of the 2015 International CES. We're expecting a big announcement (pun intended).
The specs certainly look plausible - 3,072 CUDA cores, 50 percent more than those on the GM204; a staggering 96 ROPs, and a 384-bit wide GDDR5 memory interface, holding 12 GB of memory. The memory is clocked at 6.60 GHz (GDDR5-effective), belting out 317 GB/s of bandwidth. The usable bandwidth is higher than that, due to NVIDIA's new lossless texture compression algorithms. The core is running at gigahertz-scraping 988 MHz. The process node and die-size are values we manually program GPU-Z to show, since they're not things the drivers report (to GPU-Z). NVIDIA is planning to hold a presser on the 8th of January, along the sidelines of the 2015 International CES. We're expecting a big announcement (pun intended).
80 Comments on Possible NVIDIA GM200 Specs Surface
If so, clocks-wise, it augers well for GeForce branded cards...now I guess, we wait around until Nvidia is pressured into releasing it as such.
If I'm reading this right, the ROP and core count are both 50% greater than GM 204, but the 252.9 GTex/sec fillrate implies that the texture address units have increased by 100% from 128 to 256 ( 988 * 256 = 252.928 GTexels/sec) - similar steps to that seen in the Kepler arch, although assuming 128 cores per module (as per GM 107 and GM 204) the number should be 192 ( 3072 cores / 128 per module = 24 SMM * 8 TMU per SMM = 192)
@btarunr
Should run a poll on the thread: "What percentage of posts will howl about pricing?"
A. 90-95%
B. 96-97%
C. 98%+
And yes, I think the flame trolls will be inbound with haste on the cost front.
So (if this is true) figure a core clock roughly 20-25% higher or so (real/'boost'/load), which more-or-less meshes with the earlier rumors.
I have to imagine the default clock (for consumer parts) is fluid dependent upon whatever Fiji/whatever is comparably, but I don't doubt that earlier-reported 1100n/1390b (overclocked?) is within reason for 300w.
I've always figured this arch was engineered with 20nm in mind (which obviously nvidia backed out of at some point, probably not long after they threw that hissy-fit presentation about tsmc's 20nm cost) and the clocks vs older archs reflect that. IOW, 1100 is probably the new '~900mhz' and 1390 the new 11xx (think 1.163-1.175ish volts for both companies earlier products). It meshes with one of nvidia's chief scientists that said 20nm gave a 20-25% boost.
Extrapolate as we may, we still prolly have a ways to go before any hard numbers that could truly give an indication of where it (or the competition) will end up.
The only thing one can safely say is 980 is meant to preempt a faster iteration of a 290x-like product, and a higher segment is typically around 15-20% faster.
From there, each company *should* have two more products on new chips that climb two more approx similar steps, while obviously each lower chip generally overclocks to the level of the stock next level.
If you take that to it's logical conclusion assuming 21 and 24 SMM parts that (over)clock slightly lower (~90%?) than GM204 counterparts (due to extra ram etc), I think it gels...but that's pure speculation.
(and I didn't even mention cost....even though I think 290/290x clearly show a trend for the way forward in AMD's future price structure...which will lead to inevitable 980 price drops and higher-end chips taking it's place.)
Even if this is the pro grade part, the slightly cut down version will still be a monster, but assuming its on the same process node how are they cooling this beast?
Let's be honest, if this is indeed the Titan II then we'll see the 1000 dollar pricetag return. The Titan I was an epic failure. Those who got a 780 basically got a card which was only 10% away from the Titan but at half the price. And let's not even talk about Titan Z vs R295X2.
Nvidia needs to destroy the Titan line. But they won't, because lots of NV fanboys will do almost anything for Nvidia even as the company pisses in their mouths, the fanboys only beg for more.
And btw, if AMD goes down in the GPU space - which is absolutely a possibility - consider the X80 flagships gone and replaced with the 1000 dollar GPU cards instead. But I'm sure you people will defend that, too :D
I'm one of those evil NVidiots. I drive my Nissan Skyline with decals of a winged JSH on the hood/bonnet. I buy NV stock and smoke rolled up AMD shares.
Or I but what suits my needs as long as my budget meets it.
Titan Z is just a showoff piece, not a card that sells. Mind you, it sold even more poorly than nvidia expected, so there is a sentiment of overpricedness there as well.
NVIDIA Planning To Ditch Maxwell GPUs For HPC Purposes Due To Lack of DP Hardware – Will Update Tesla Line With Pascal in 2016, Volta Arriving in 2017
There might not be Titans this time around but high prices is another thing.
And it presumably makes sense with 28nm sticking around longer than expected, the recent launch of the K80 no doubt helps to fill the gap.
Speaking of Titan, that big fat government contract for two new super computers Summit and Sierra should keep them busy.
NVIDIA Volta, IBM POWER9 Land Contracts For New US Government Supercomputers
Nvidia GM200
AMD Fiji
Ask yourself, what source would be in possession of BOTH Nvidia's and AMD's next top cards as well as AMD's second tier offering and Nvidia's GM 200 salvage part. One source having access to four unreleased top tier parts across both vendors :rolleyes:
Not only do they have access to both vendors next offerings, not a single other source has even a single one of those four benchmarked.
Can you source the debunking? How were they debunked by the way? Was it a consensus from people who didn't like the outcome?
On the other hand, a 290X runs on par with GTX980 at 4K so it's not out of the question to expect AMD and NV top be close this time around. The loser will be whoever releases first (IMO). I think the vendor that releases second will tinker with their product to pull a performance edge or use aggressive pricing.
Good for both camps.