NVIDIA Maxwell Speculation Thread

tmu:138 ?
138/8=17.25
138/13=10.615...

???
Its a Kepler based calculation: 1664 / 192 * 16 = 138. The newest GPU-Z should show some more correct data.
Maybe still wrong if they went to back to 16 TMUs per SMM.



That sounds like a high estimate. GM107 is 25% larger than GK107 where the difference is 2.5x the number of SM's and the large L2.

If we assume 2x the number of SM's in GM204 vs GK104 and a similar increase in L2 it should be closer to 370mm^2. Of course there could be more to Maxwell than we've seen so far.
They got another more direct shot: http://cdn3.wccftech.com/wp-content/uploads/2014/09/NVIDIA-GM204.jpg article
... package is supposed to be 40 x 40 mm, like GK104.

With some marketing calculation of die-size - not counting border space of the die, you might come to ~380mm².

But still a massive die if it is only 16 SMM and 2MiB L2. Maybe there are some secrets like in Tonga... :LOL:
 
The L2 usually scales with the number of memory controllers so we should expect 4MB.
But the L2 per MC is also scalable: GK208 64-Bit / GK104 256-Bit had booth 512KiB L2.

Compared to GM107 2MiB sounds a bit low for GM204, but SiSoft read it from somewhere:
http://forum.beyond3d.com/showpost.php?p=1872390&postcount=1976 ... and I do not see that it somewhere calculated, when architecture is not detected and there was now 2MiB 256-Bit GPU.

2MiB would be still a jump from GK104 and GK110 (1.5MiB). GM107 was probably optimized to perform fast with DDR3, like GK208.
 
But the L2 per MC is also scalable: GK208 64-Bit / GK104 256-Bit had booth 512KiB L2.



Compared to GM107 2MiB sounds a bit low for GM204, but SiSoft read it from somewhere:

http://forum.beyond3d.com/showpost.php?p=1872390&postcount=1976 ... and I do not see that it somewhere calculated, when architecture is not detected and there was now 2MiB 256-Bit GPU.



2MiB would be still a jump from GK104 and GK110 (1.5MiB). GM107 was probably optimized to perform fast with DDR3, like GK208.


You can read L2 size from CUDA or OpenCL device properties.

http://developer.download.nvidia.co...DEVICE_g5aa4f47938af8276f08074d09b7d520c.html

http://www.khronos.org/registry/cl/sdk/1.0/docs/man/xhtml/clGetDeviceInfo.html

If it really is only 2MB then that improves the odds of a smaller die.
 
According to those Sisoft benchmarks Fermi is twice as efficient per flop than Kepler/Maxwell. That's pretty sad. Hopefully it's just due to a regression in nVidia's OpenCL driver.

Given that Maxwell has reverted to single-issue flops and has increased per thread registers and bandwidth and lower latency hiding requirements I would expect it to at least match Fermi's efficiency.
 
No doubt it will be more expensive but I don't think they want to learn new node lessons with GM2x0.

Ok then in order to bounce back to your original sentence I quoted, it depends what kind of node lessons they'd need. Considering they need Pascal out the door truly ASAP they can play around with testchips without even bringing them to tape out.
 
Price:
Zotac 970 pictured: https://www.facebook.com/pchub/photos/p.894462047250388/894462047250388/?type=1

PCHub Price range:
GTX 980 : 31-34k
GTX 970 : 19-22k

https://www.facebook.com/pchub/phot...0.1410452114./894392987257294/?type=1&theater

This could be $399 for GTX 970, but GTX 980 seems to have a nice 50% premium - $599?
*Imaginary* NV marketing speek: "up to 20% faster, 30% less power consumption, 4GB for 4K and $100 cheaper than GTX 780 Ti". :LOL:


TDP:
http://www.techpowerup.com/gpudb/2621/geforce-gtx-980.html
http://www.techpowerup.com/gpudb/2620/geforce-gtx-970.html
156W and 175W, maybe TPU got some BIOS dumps with these data?
 
Last edited by a moderator:
Price:
Zotac 970 pictured: https://www.facebook.com/pchub/photos/p.894462047250388/894462047250388/?type=1



https://www.facebook.com/pchub/phot...0.1410452114./894392987257294/?type=1&theater

This could be $399 for GTX 970, but GTX 980 seems to have a nice 50% premium - $599?
NV marketing speek: "up to 20% faster, 30% less power consumption, 4GB for 4K and $100 cheaper than GTX 780 Ti". :LOL:


TDP:
http://www.techpowerup.com/gpudb/2621/geforce-gtx-980.html
http://www.techpowerup.com/gpudb/2620/geforce-gtx-970.html
156W and 175W, maybe TPU got some BIOS dumps with these data?

Excuse me, where did you read this?
 
http://wccftech.com/geforce-gtx-970-980-pricing-leaked-399-499-zotac-gtx-970-pictured/

Reported specs for the GTX 970 are 1664 CUDA cores with 4GB of GDDR5 memory clocked at 7Ghz effective operating across a 256bit buss. Rumored TDP and price are 150W and $399 respectively. The GTX 970 is supposed to go head to head with the R9 290 and the GTX 780 in terms of performance but offer better power efficiency and an updated set of features. The GTX 980 is supposed to outperform and replace the GTX 780 TI as the fastest single GPU card from Nvidia and also introduce enhanced efficiency and an updated feature set.

The e-tailer also mentions pricing, stating that the GTX 970 would cost 19-22K Philippine Pesos which amount to 430 to 500 US Dollars and the GTX 980 would cost 31-34K which amounts to 700 to 770 US Dollars. These are quite higher than the previously rumored $499 and $399 for the GTX 980 and GTX 970 respectively. The leaked MSRPs by the e-tailer might end up being pre-order prices which are usually quite inflated so I’d take them with a grain of salt.
 
http://wccftech.com/geforce-gtx-970-980-pricing-leaked-399-499-zotac-gtx-970-pictured/
The e-tailer also mentions pricing, stating that the GTX 970 would cost 19-22K Philippine Pesos which amount to 430 to 500 US Dollars and the GTX 980 would cost 31-34K which amounts to 700 to 770 US Dollars. These are quite higher than the previously rumored $499 and $399 for the GTX 980 and GTX 970 respectively. The leaked MSRPs by the e-tailer might end up being pre-order prices which are usually quite inflated so I’d take them with a grain of salt.
Read more: http://wccftech.com/geforce-gtx-970-980-pricing-leaked-399-499-zotac-gtx-970-pictured/#ixzz3D3F7vWrp
Strangely enough, the Philippines were the first "source" of the GTX 670 pricing....at the same 20K Pesos ($470). Actual pricing: $400.

Wouldn't surprise me if Nvidia encouraged the circulation of higher prices prior to launch as they did with the GTX 680 also. Adds a little more to the PR war chest for no actual effort.
 
Back
Top