guess fermi\'s yield by selling price and availability

hd4770

facts so far:
1) #470 >> #480
2) 470 & 480 have limit availability 4 weeks after launch
3) 470 MRSP = $349
4) NV has 70+% TSMC 40nm capacity
5) GT21x has 40+% margin
6) 58xx selling at a premium
7) About 100 fermi chips per wafer
-------------------------------------------------
If 470 & 480 has 60+% of GT21x margin, i.e. 24+% margin, the overall margin (due to tesla 65+% margin) could be 35+%, in this case, it is worth reducing GT21x volume (meanwhile gaining highend market share), thus we should have seen large volume of fermi by now, contradicting to 2). Conclusion 1: Fermi's margin is less than 60% of GT21x margin.

Given 6, Fermi's margin should not be very low. Thus, a reasonable guess of 20%.

Now assume 25% channel margin (including service cost), 349*25% = $87. Per chip nv is selling 470 at 349 - 87 = $262. Applying 20% margin, per card cost = $210. Assume $85 board & memory cost, gtx470 per chip cost = 210 - 85 = $125.

Assume 80% wafer cost ($5k) is by 470, #470/wafer = 5k*0.8/0.125 = 32.

Given 1), #470 : #480 = 4 : 1 looks reasonable, thus 8 480 per wafer. If the tesla chip yield is similar to 480, another 8 per wafer. Therefore, 8 + 8 + 32 = 48 chips per wafer.

Conclusion 2: Given 7), the real fermi yield is 48%.

Although 48% is only a result after some calculations, it is based on all the above facts, thus won't deviate from the real number too much. Moreover, gtx460 could add a few more percentage to fermi's overall yield.  

Any predict of 20-30% overall fermi yield is a lie. BTW, many newbies feel that AMD/ATI should lower their 58xx/57xx prices because they believe AMD has a lot of room for a price war. But the fact is AMD graphic profit Q1 2010 = 47 millons, and sold 4 million DX11 units, say, 1 million 58xx, 3 million 57xx/56xx/55xx. If AMD only lower 25$ per 58xx and 10$ per 57xx/56xx/55xx, AMD's graphic division would end up with a  - 8 million profit.

lightbb

顶沙发翻译

苯苯小哥

NV费米良品率计算方法？成本算法？

panjanstoneborg

从fermi的售价和供应推测其良率

到目前为止
1) 470数量大大于480
2) 发布4周数量仍有限
3) 470卖349刀
4) nv占据70%以上的TSMC 40nm产量
5) GT21x系列利润率在40%以上
6) 5800系列卖的稍贵
7) 约100个dies / wafer 对于fermi来说

如果 470和480的利润率达到gt21x的60%以上，并计入tesla系列的利润率，fermi的利润率可以达到35%以上。这可以允许nv通过减产gt21x来增加fermi的数量。这与2) 不符。
结论1： 因而fermi的利润率不足35%。但是考虑到6) fermi的利润率应该不会很低，20%是一个合理的猜测。

那么假定渠道利润率是25%，可以推算出470卖价在262刀，按利润率25%来算，470就是210刀的成本。如果假定其他部件成本是85刀，则470用的GF100成本是125刀。

如果470占用80%的wafer成本(5k刀)，那么一个wafer上470的个数是5k刀*80%/125刀=32个。

根据1) 如果认为470和480的数目为4比1，再计算上tesla用的芯片，就是总共48个/wafer。
结论2： 那么良率就是48%。基于以上事实，fermi的良率不会和这个推算相差很大。考虑到将来460，还会增加。

所以说fermi良率在20-30%是不对的。

此外，ati 10q1的利润是47M，总共销售了1M的cypess和3M的其他，那么假定ati把5800系列的定价都降低25刀，其他系列的定价都降低10刀，利润就只剩8M了。
（ati的这个推算太粗糙了）