POPPUR爱换

标题: GTX 970的谜思 [打印本页]

作者: hd4770 时间: 2015-2-25 11:51
标题: GTX 970的谜思
- It started with a benchmark program finding the memory of last 0.5g is very slow
- Jonah then explained that 7 slices of crossbar at full speed (a total of 196G/s), and accessing the last 0.5G only by the 7th slice of crossbar, thus 28G/s.
- People starting to question the peak 224G/s bandwidth in the spec

Most people so far believe that the peak memory bandwidth is 196G/s, not 224G/s. 这是一个误解. 程序写的好的话，196G/s和28G/s是可以相加的.

作者: eraser666 时间: 2015-2-25 16:04
其实就是驱动程序的锅啦

作者: fengpc 时间: 2015-2-25 21:25
鉴于各大媒体的炒作能力和国内一般人的英文水平，最好是把把那个Senior VP的说话原文或者录音放出来供大家判断才有说服力，况且这种负责高层次管理的人也不一定清楚细节。
另外，人家也没说L2的带宽就等于显存的带宽，L2的带宽是大于DRAM的带宽的，当需要读取的数据在L2里面的时候就不会访问DRAM，这就是L2比较大能提高性能的原因，现在970砍的是L2不是位宽！

作者: 蒙大拿 时间: 2015-2-26 00:46
Gpu也有cache?

作者: G70 时间: 2015-2-26 02:40
提示: 作者被禁止或删除内容自动屏蔽

作者: hd4770 时间: 2015-2-26 09:56
- L2 size and ROP number are not in gtx 970 spec.
So there is no false advertisement in product spec.
The reviewers said they got L2 size and ROP number from NVIDIA. Do they have document to support that? So far no one saw it. It could be misunderstood by the reviewers. Even if they could prove this, any comments from reviewers are not official advertisement from NVIDIA. Are there any prior examples for such a case in lawsuit? Anyway, this would be interesting to everybody how the lawsuit comes out.

作者: gz_easy 时间: 2015-2-26 12:17
3.5GB, 224bit位宽@7GHZ够用了，当初就这样宣布多好

作者: G70 时间: 2015-2-26 12:53
提示: 作者被禁止或删除内容自动屏蔽

作者: fengpc 时间: 2015-2-26 14:33

G70 发表于 2015-2-26 12:53
典型甩锅手法，哈哈。

这下会不会把reviewers 全得罪了。

也不算甩锅，原厂有给一份tech guide，不过那不能算商品销售的spec，而且发现问题之后也马上修正了~~
这次970弄得对手没有还手之力，难得逮到个痛脚还不拼命地抽。。。

作者: eraser666 时间: 2015-2-26 14:58
本帖最后由 eraser666 于 2015-2-26 17:41 编辑

蒙大拿发表于 2015-2-26 00:46
Gpu也有cache?

Maxwell的一个很大改进就是大容量L2好吧
连回复都不用你还要我尊敬你？

作者: 3670636 时间: 2015-2-28 16:43
如果GTX970标注3.5G甚至3G，我相信今天销量一定还是这么好，也不会有这样大的风波，不管怎么样骗人总是不对的，老黄这次真的栽了~

作者: lemonninja 时间: 2015-2-28 23:06
这次主要是透露的资料太少了，要是一开始就介绍新的Cross Bar架构说不定还会引起大家对新架构的兴趣。

作者: hd4770 时间: 2015-2-28 23:42
本帖最后由 hd4770 于 2015-2-28 23:43 编辑

It is normal business conduct to not reveal too much secret. The lesson learned is ordinary users tend to use naive program to play with cards and to satisfy his/her curiosity, which sometimes hit corner cases. The corner case in GTX 970 was really intended to optimize the power vs. the performance. A full memory system might just get 0.1% more perf, while it would consume 3% more power. So the bisection of 3.5+0.5 was really to benefit users (which can be done only in Maxwell with the newly added hardware in memory system, in other words, this is a new feature not without hardware cost). In this sense Jensun claimed this is a feature to enable a better config for a game card is true. If I were a GTX 970 owner, I would probably think it is more like "AHH, is this an opportunity to get a few bucks back?" rather than this is a real problem with the config.

作者: G70 时间: 2015-3-1 02:17
提示: 作者被禁止或删除内容自动屏蔽

作者: hd4770 时间: 2015-3-1 06:01
so far I only saw the naive benchmark shows a memory access speed cliff off above 3.5G. For average game perf, gtx970 is aligned nicely with gtx980 when above 3.5G. It seems you had a gtx 970. Did you find any obvious problem for "970在3.5G以上占用会出问题" before this problem popped up around end of Jan?

作者: G70 时间: 2015-3-1 13:40
提示: 作者被禁止或删除内容自动屏蔽

作者: G70 时间: 2015-3-2 13:51
提示: 作者被禁止或删除内容自动屏蔽

作者: eraser666 时间: 2015-3-4 09:12

G70 发表于 2015-3-2 13:51
因为有人测试过3.5G+占用的时候，970SLI frametime变得极其极其不稳定和偏高。

不需要SLI，单卡的frametime测出来就有跳动

蛮正常的，大部分情况都是980最好，970次之，290X最差。

但是970有跳动

作者: G70 时间: 2015-3-4 22:45
提示: 作者被禁止或删除内容自动屏蔽

作者: Xenomorph 时间: 2015-3-5 15:12

hd4770 发表于 2015-2-28 23:42
It is normal business conduct to not reveal too much secret. The lesson learned is ordinary users te ...

以0.1%性能代价减少3%能耗的某种设计，技术上看肯定是好的，但消费者会认为关键在于是否会有其她不可预知的后果发生。

现状是对桌面显卡而言“超过3.5G显存占用”已经不是corner cases了。但engineer当初在测试时怎么没发现。

相反，notebook平台上由于分辨率受限，高显存使用的情形会相对更少，对电力效率的要求也更高，为什么GTX980M都没有去阉割L2/ROP？

作者: amx004 时间: 2015-3-5 16:41

hd4770 发表于 2015-2-28 23:42
It is normal business conduct to not reveal too much secret. The lesson learned is ordinary users te ...

既然是如此目的
可是媒体评测能耗比一项970不如980,
你该不会说如果不是这么设计, 970能耗比更差吧..

不管怎么说, 给人感觉就是欲盖弥彰, 其实就是阉卡阉到引起公愤嘛, 大家心照就是了

欢迎光临 POPPUR爱换 (http://we.poppur.com/)