Specs of NVIDIA First 40nm GPU GT216 Emerges
AMD’s first 40nm GPU RV740 has been pretty hot, but its rival NVIDIA seems to be quite low-key on this. According to Beyond3d, NVIDIA is preparing its first 40nm product – GT216.
As we know, GT216 will feature 192-bit memory interface and 768MB capacity. It adopts GDDR3, working at 1.2GHz (2.4GHz data rate). Probably there will be a 384MB version of GT216. No words yet on the number of stream processors.
Well, let’s suppose RV740 does have 640 stream processors, according to the proportional relation between number of SPs of RV740/RV770, we can reckon that GT216 might feature 168SPs and 56TMUs. (The original text says GT216 will equip 160SP/32TMU, but considering the structure of GT200, 160SPs sounds impossible. GT200 contains 10 groups of TPC, and each TPC has 24 SPs. For instance, GT280 has 10 groups of TPC, GT260-192 has 8 groups, GT216-216 has 9 and GT216 might have 7).
40nm product is something great to expect. Both RV740 and GT216 are highly cost effective. We will continue to follow the next round of mainstream GPU battle.
December 23rd, 2008 at 6:33 pm
When?
December 23rd, 2008 at 6:53 pm
are all this confirmed?
December 23rd, 2008 at 7:41 pm
Doubt it
December 23rd, 2008 at 7:56 pm
I’m suprised that Nvidia still didn’t adopt GDDR5
December 24th, 2008 at 5:24 pm
What the heck happen to DDR4 ???
December 24th, 2008 at 6:07 pm
LOUDER SON!
December 25th, 2008 at 10:18 am
I’d think outside the box on this one.
To get 160sp/32tmu as the article suggests, you’d have 8 arrays with 20sp and 4 tmus an array. Think of it as a strange cousin of the 8800gtx: Same amount of arrays, up the shaders from 16 to 20 per array, half the tmus from 8 to 4, and split the bus in half (from 384 to 192-bit). Or, perhaps, half the arrays with twice those specs per array.
Other possibilities, for example, could be they may change the sp from 24 to 16 per array.
Also, TMUs could go from 8 to 4 per array.
You’d end up with a 160sp/40tmu part. Sounds familiar donut?
Lets keep it simple though, let’s say it’s 160/32, now that we know it’s possible.
Let’s say it’s clocked at 600c/1500mhz shaders. You would have the same FLOP ability as a gtx260 (192sp) part (720gflops). Of course it would have half the TMUs and 1/2 the bandwidth, but that’s the price of a small die and a lower-end part. Think of it as a 9600gt vs 8800gt (gt216 vs gtx260 192), only without losing flops.
I think such a part could be done at around rv670 size.(150-200mm2), which fits the market. Such a part theoretically matches up against the 4850/rv740 very well in core logic/bandwidth, if you clocked a rv740 @ 775mhz (1TF). Of course that would be if the efficiency of rv740 v gt216 is the same as rv770 vs gt200. GT216 might be slightly less efficient as ATi can compensate by upping the clockspeed vs rv770, where-as the deficits of gt216 would be much greater compared to gt200 in the texture/bw department.
Even if this isn’t the route they go, changing of the array structure is not uncommon for lower-end and/or refresh parts ala the 9600gt (say versus the 8800gt/9800gt), and Nvidia has done it before. They need not, and most likely shall not, constrain themselves to the GT200 structure.
Even so, I think that ATi will have the more efficient setup, and with the ability to use a 128-bit bus and gddr5, it allows them to create a smaller chip than they would have needed to with 256-bit (or wonky 192-bit).
The saving grace, imho, will be 768MB of RAM versus 512mb on the 740, as in this sweet spot RAM does matter. If ATi can squeeze the broke Qimonda for good deals on 1gb gddr5 though…That could surely make a difference.
December 25th, 2008 at 10:45 am
FWIW I think the matchup to wait for is rv770pro (4750/5650?)
It’d be something like this:
RV770pro:
640 shaders @ 625mhz (800gflops)
128-bit
2.4ghz GDDR3 (38.4 gbps)
32 TMUs (20 gt/s)
16 ROPs (10 bps fillrate)
512MB/1GB
<75W
vs
GT216
160 shaders @ 1500mhz (720Gflops)
600mhz core
192-bit
2400mhz 192-bit gddr3 (57.6 gbps)
32 TMUs (19.2 gt/s)
16 ROPs (9.6 bp/s fillrate)
768MB
<75w
If Fudo is right and rv740 is 256-bit though, it’d be more like a slaughter than a competition…We shall see!