分類彙整: GPU

在GDC08登場的各個引擎

http://www.watch.impress.co.jp/game/docs/20080228/3de.htm
西川善司の3Dゲームファンのための「GDC 2008最新ゲームエンジン」講座(前編)
和製CAE対応「オクターブエンジン」、ソニー製なのにXbox 360対応「PHYRE ENGINE」、Valve「Team Fortress 2」
http://www.watch.impress.co.jp/game/docs/20080229/3df.htm
西川善司の3Dゲームファンのための「GDC 2008最新ゲームエンジン事情」講座(後編)
「CRY ENGINE」独自拡張「DUNIA ENGINE」、新生「UNREAL ENGINE3」

所以會場那個「Octave Engine HONGO」 demo其實用到了總共七個GPU:
GPGPU:Quadro FX 5600 x 4 + Tesla C870 x2
Render:GeForce 8800GTS 512MB x1
這樣可以real time 跑 300個partical,然後一張單獨的GPU來render。
實際上Prometech宣稱的性能是四GPU可以在30fps下模擬100萬個partical,目標是CAE市場。

另外,Octave Engine Casual超有趣XD

—–
所以PSSG其實只是繪圖的部分,加上其他的工具合起來才是Phyre Engine…._A_
http://www.watch.impress.co.jp/game/docs/20080228/3de057.htm

不過由於除了PSSG可以跨LibGCM、OpenGL和DirectX,剩餘的所有的實作都是CPU相關,實質上就是PS3靠SPE撐的部分。
雖然有辦法移到其他平台,但是實際上在XBOX360和PC上的話,表現是會有一點落差;當然其他平台就得用繪圖部分的資源來補足啦….

比較有趣的事情是後面要提供的Terrain Engine,可以有效率地生成大規模的場景;而這個引擎有streaming texture的能力,不需要勞煩美工做texture tile。
http://www.watch.impress.co.jp/game/docs/20080228/3de.htm

也就是說,其實這已經有點MegaTexture等級的感覺啦!XD

9800GTX本體照

http://www.vr-zone.com/articles/GeForce_9800_GTX_Card_Photos_%26_Specs_Unveiled/5599.html
GeForce 9800 GTX Card Photos & Specs Unveiled

http://en.expreview.com/?p=277
9800GTX GPU-z screen shot

PCB為P392,12層、與88GTX等長(約27cm)、兩個6pin電源、TPC約168w。
果然還是要衝一下賣相嗎XD
而9800GT應該就和8800GT/GTS的PCB不會相差太多才是,雖然有兩組MIO-port。

—-
話說回想先前曾經認為G92的存在也許類似6800GS…._A_

至少掛上了9800GTX的名字之後應該是不至於啦。
而且G94的結構最後看起來也是和G92共通性很高,不像NV42和G73其實落差還不小。

—-
http://hkepc.com/?id=806&page=4&fs=idn#view
沒注意到G94支援HDMI 1.3a

9600GT正式發表

http://www.4gamer.net/games/047/G004716/20080220011/
NVIDIA,ミドルレンジ向け新型GPU「GeForce 9600 GT」を発表

http://pc.watch.impress.co.jp/docs/2008/0222/nvidia.htm
NVIDIA、ミドルレンジ向けで256bitメモリの「GeForce 9600 GT」

http://pc.watch.impress.co.jp/docs/2008/0222/tawada131.htm
多和田新也のニューアイテム診断室
NVIDIAの新ミドルレンジGPU「GeForce 9600 GT」

G94的時候特別拿出來講,不過基本上G92的ROP和G94一樣,都是相對於G80有改善,基本結構的話G9x都是共通的。
所以和8800GT相比雖然命名上有點爭議,不過隨著9800GTX推出,這個爭議應該可以消除。

G94的特色是”整顆正好是G84的兩倍”,包含shader、ROP(fillrate)、記憶體頻寬與容量,所以性能也達到了差不多兩倍的表現,大部分狀況只低於8800GT約15%左右,是C/P值非常高的設定。
如果當初把G72(7600GT)做成256bit的話,與7900GTX之間的關係就會很接近現在9600GT與9800GTX之間的關係。

時脈為650/1625/1800,64SP、512MB GDDR3,TDP為95w,散熱器與8800GT相同,值得注意的是它是NVIDIA第一個內建DisplayPort的產品,所以毫無疑問未來會有DisplayPort直接onboard的產品出現。

此外,由於預期9600GT會有超頻板的出現,AMD宣布3870降價到189美金來對應。

9800GTX預計發表時程

GeForce 9800 GTX (D9E-20) to launch in Late March – VR-Zone

所以D9E-20其實是G92版8800GTS 512MB把時脈提高後下重新包裝的產品。XD
這樣看來8800GTS其實C/P值也沒有想像的低,因為有更低的產品可能要推出了….預計是9800GX2推出後一小段時間就會登場,以滿足市場對高速單晶片的需求。
而對替代8800GTX的定位來說,用9800GTX對成本結構的改善也有好處,不僅是PureVideo HD2的存在,讓G92搭配3way SLI應該也是可以期待。

——
The follow up to the 8800GTX is….. – The INQUIRER
INQ這兩天寫9800GTX會有768MB,這點應該是煙幕彈。
不然192bit還得了…. G92又不太容易有384bit。

[EDIT] 1GB確認。

同時會有9800GT,正好就是8800GT + 3way SLI。

NVIDIA 2007Q4財報

http://www.nvidia.com/object/io_1202940080671.html
NVIDIA Reports Record Results for Fourth Quarter and Fiscal Year 2008

http://forum.beyond3d.com/showthread.php?p=1128580#post1128580
Default NVIDIA’s Record Q407: ~$1.2B Revenue/~$300M Net Income

http://pc.watch.impress.co.jp/docs/2008/0214/nvidia.htm
NVIDIA、2008年度売上高は過去最高となる41億ドル

8800GT在2個月內出貨兩百萬顆(當初R300是半年出貨一百萬顆左右,G80是五個月出貨100萬個),雖然有些成本上的問題。
AGEIA購併完成,營業額創新紀錄。

內容錄音:
http://phx.corporate-ir.net/phoenix.zhtml?c=116466&p=irol-audioArchives
http://biz.yahoo.com/cc/2/89332.html

全文:
http://www.aiplus.idv.tw/phpBB2/viewtopic.php?f=9&t=8366

CEO
– Another record quarter and record year.
– [Income Figures]
– [New Products/Initiatives]
– [CUDA as a first step towards Heterogenous Computing and also interesting for consumer workloads]
– [New Products/Initiatives including Hybrid SLI]
– [Growth percentages, MCP was the lowest at +7%]
– [Mental Images & Ageia acquisitions]
– Very satisfied with our strategic position heading into Fiscal 2009.

CFO
– Differences between GAAP and non-GAAP is stock compensation and Mental Images charges.
– Notebook and MCP were up slightly at 1 and 2% sequentially.
– Desktop was up 69% year-over-year.
– Consumer business [mostly handheld] declined.

– Some cost issues with the 8800 GT during the quarter which we hope to resolve in the next two quarters.
– The revenue for Mental Images during the quarter was less than one million dollars.
– Employees up 902 year-over-year, two thirds of that being R&D.
– Cash was down $43M quarter-over-quarter. $178M stock repurchase.
– Cash flow a little bit over $250M in the quarter.
– Capital expenditures were $70M in Q4.

Outlook
– We expect revenue in Q1 to be better than seasonal.
– We believe revenue will be slightly down.
– We will get back on track to improving gross margins, will be flat to slightly up.
– Operating expenses will increase 8-10% due to less vacation and $7M extra charges.
– Tax rate will be higher in 2009 since the R&D Tax Credit hasn’t been renewed.

CEO
[…]
“The CPU has become good enough for most users.”
[Enthusiasm about low-end CPU + high-end GPU configurations]
[…]


Q&A

Q: Marv, Q1 slightly stronger than seasonal – how do you perceive seasonal and that strength? Strength in desktop?
A: Seasonal is down 5% historically. PC in general is 5-10%. So I mean it’ll be down less than that.
Regarding desktop, performance GPUs are doing very well (8800GT).
Q: Notebook arena volatility? + Growth in the 20% range?
A: Notebook in Q4 was up 1%, it certainly didn’t decline. I would expect it to follow seasonality in Q1.
Regarding 2008, I don’t think I have any update compared to what I said earlier.
Focus on growth in desktop GPUs, growth in MCPs, growth in Workstation.

Q: April quarter, strength in the desktop GPU market. Do you see growth in the overall business?
A: It’s all over the segments, increasing adoption of the GPU. Bias towards slightly higher-end GPUs.
People understand that rebalancing the system to be more GPU-centric will result in muigher performance.
One PC OEM after another, they are increasing the budget allocating to the GPU.
Q: Competitive landscape in MCP? Notebook market? Dual-chip X2 from AMD?
A: We have the only Motherboard GeForce solution for both AMD and Intel.
Anybody who doesn’t want to sacrifice a rich graphic experience and not sacrifice compatibility,
we believe our GPUs on the motherboard are ideal. This is our first year for that on the Intel platform
and we’d expect that to do quite well. We’re enthusiastic about MCP.
[marketing about X2 vs 8800GTX] “We will shortly announced Tri-SLI”. [errr?]

Q: Cost issues regarding 8800 GT?
A: Yeah. Hmmm. 8800 GT was ramped probably faster than any other high-end GPU in our history.
In just 2 months of production, we shipped 2 millions of them. Yet the die size is larger than any other GPU.
We had some manufacturing challenges, but now yields should be much better going forward.
Marv: From my standpoint, it was about getting the product out. We focused on that even though cost was high.
Q: Ageia?
A: […] Our strategy is to port the physics engine on top of CUDA.
Our expectation is that this is going to encourage people to buy even better GPUs. […]

Q: Timeline for when we should start to see the impact in the actual product offering?
A: We’re working on porting the physics engine to CUDA right now.
Every CUDA-enabled GPU will be able to do it. […]

Q: Segment breakout? When could Ageia be accretive? 8800 GT costs? Need to port to 65nm?
A: […] Consumer was down 18% quarter-to-quarter. […]
“8800 GT costs is just about doing the process engineering […]”
Ageia doesn’t have high operating expenses […]

Q: CUDA? Requires a CPU business?
A: We believe in Heterogeneous Computing, the CPU is already there [from Intel/AMD], it’s fabulous […]
[GPU Computing marketing] [Physics marketing]
Q: DX10 Chipset. What are your expectations there?
A: In order to get Vista Certification by the second half of the year, you require DX10.
So I think the answer is we’ll get there as fast as possible.

Q: X2-type of products, will NVIDIA introduce products like that? Will that become the standard at the high-end?
A: We would do an X2-like product only if it delivered performance that’s possible only
But there’s no question a single GPU approach is better. But either way, it has to be the absolute highest performance.
If a X2 has the highest performance, it will be accepted. Otherwise, you’ll have big problems.
Q: In the near-term?
A: [huge laughter] We can’t talk about future products.
Q: Cash?
A: […] Intend to buy more stock back in Q1.

Q: Growth targets for the year still around 20%; did you discuss share gains etc.? AMD in the notebook area?
A: Maybe we’re just unsuccessful in communicating this, but the entire GPU market is booming.
We might be gaining share but it’s not significant. […] We’re just selling more GPUs and slightly higher-end.
Q: For all platforms? Desktop/Laptop/Workstation?
A: Absolutely.
Q: US vs International demand?
A: Not dramatically different demand in the different regions.
Our expectation is that the growth of GPUs will overcome the seasonalities.
Q: Pricing environment?
A: Pricing is good. More money spent on GPUs so that helps Average Selling Prices for the entire GPU business.

Q: ASPs?
A: Up nicely, since we shipped a lot of 8800 GTs.
Q: Operating expense growth in Q1; what about after Q1?
A: Working hard to keep it flat, but we’ll see how successful we are.

Q: Supply etc.?
A: We are the world’s fabless semiconductor company.
The products that we build are very complex and the die sizes of our largest GPUs are significantly bigger than CPUs.
So these processors consume a lot of wafers, it’s imporant we work very very closely with our partners.
TSMC has done a wonderful job for us, and my expectation is we’ll get the fabless model to new heights.
So we obviously take supply very seriously.
Q: Lead time?
A: It hasn’t really changed.

Q: Guidance for better than seasonal for Q1. Unique for NVIDIA?
A: I think the GPU market is just growing. It ought to help anybody who builds a GPU.
The way I see it is that visual computing is just becoming ever more important.
Q: Competitive landscape for the fiscal year. Unit share or revenue share perspective for main business lines?
A: Can’t break that down, sorry. [market is growing] [segmentation of the market and increase of ASPs]
Q: Gross margins. TSMC has talked about trying to renegociate pricing. Did this affect the margins?
A: Gross margin impact of 8800 GT was completely my fault. It has nothing to do with TSMC.
We’ll have to take all of that responsibility ourselves. But the good news is that since it’s entirely our fault,
we can fix it entirely ourselves. Can still expand margins.

Q: Competitive landscape in Notebooks. You vs AMD and IGPs in the next cycle?
A: In the notebook segment, AMD has wonderful offerings. Don’t forget they still have a terrific GPU design organization.
Between NVIDIA and ATI, or AMD now, it is the unambiguously the two best GPU design teams the world has ever seen.
I give them a lot of respect. And in the lower-end segment, we can’t differentiate against them. So we don’t expect
to win much of the business. In the upper-end, we can differentiate and we invest much more in R&D.
Q: Inventory increase?
A: New products in the pipeline.

Q: Overall PC environment opinion?
A: We’re fine. All we see is fine. We watch every single week around the world, and it looks good for us.
I don’t really know what else to say.
Q: OEM vs Whitebox difference?
A: Not really. I think it’s possible that the SKUs we are in are selling well and people are just prefering
platforms that we are designed into, but I don’t really know.
[visual computing & reallocation of resources within the PC]
I think it’s logical that should be a permanent thing.

NVIDIA的APX2500

http://pc.watch.impress.co.jp/docs/2008/0212/nvidia.htm
NVIDIA、720p動画対応の
Windows Mobile向けプロセッサ「APX 2500」

http://www.watch.impress.co.jp/av/docs/20080212/nvidia.htm
NVIDIA、スマートフォン向けの720p再生対応プロセッサ
-Windows Mobile端末で「スマートフォン 2.0」

http://www.nvidia.com/object/apx_2500.html
官方製品情報

http://www.beyond3d.com/content/articles/101
B3D的分析

ARM11 core + HD AVP(720p H.264 decode) + ULP GeForce(注意,有CSAA support)….這對Windows Mobile廠商來說應該會是個相當有吸引力的東西。
讓Windows Mobile設備可以解720p的訊源,一方面是為了HDMI輸出,一方面是為了減少”掌上機專用訊源”的必需性。畢竟基本上轉檔還是一件麻煩事情….

支援CSAA的ULP GeForce….真的是G8x嗎?還是改ROP的G7x?或者是完全無關的東西?

[EDIT]
從B3D提供的資料來看,應該是GoForce5500的強化版。

—-
其實如果不是PC市場,光是CPU + GPU整合的話,去買個ARM-core應該就很充足了….
由於NVIDIA的主業是GPU,如果打算要讓GPU可以獨立動作(現在的Tesla需要Host PC),其實不見得一定要x86;而且現在根本沒有機會去搞個和Intel/AMD相比之下有競爭力的x86 core,成本也太高了(更別提沒有晶圓廠)。

反過來說,就算是Tesla,感覺上也不需要弄到可以獨立動作才行:
目前Tesla把Host PC的記憶體和GPU上的on-board memory當成階層記憶體結構來使用,一邊是高頻寬一邊是高容量,全部攬到GPU自己身上的話不見得會比較好。

老文:David Kirk 在東京大學的Tesla/GPGPU相關演講

http://journal.mycom.co.jp/articles/2007/09/13/nvidia/index.html
“GeForceの父” David Kirk博士、東大で並列コンピューティングについて講演

這篇的重點有:
1. 每個Shader Multi-processor(SM)有768個thread、8個SP、每個TPC有兩個SM、所以總共有12288個data-thread可以on-the-fly執行。
2. 每32個data-thread組成一個warp,只要以這個單位來切換,就可以達到無延遲。
3. 每個SM有自己的share memory。(G80是16KB)

其次是CUDA programming的技巧
4. programmer可以設定grid,每個grid則由1~512個block來組成。
5. 每個block內部的資料可以以1~3次元的方式排列,這個block稱為Cooperative Thread Array(CTA)。

到這邊已經很清楚了:這樣一段一段的把資料結構切出來,就是希望programmer替GPU做好data的localize。
這樣的話每個SM對每個CTA進行處理,SM的share memory也只以CTA內部的單位來處理,理論上只要CTA有設好 = localize有做好,就不會有太多cache miss。
然後等到要進行記憶體存取的時候,就會發生latency,這時則把整個CTA(data set)給卸載換掉,來確保GPU維持滿載。

所以,實際在寫CUDA程式的人,除了上述的要點之外,不需要再深入顧忌GPU實際上的結構;
或者說,上面的要點實際上就是G8x這個GPU最終的可擴充範圍。
在ISA固定(保留擴充性前提)的狀況下,CTA的資料結構固定住之後,接下來NVIDIA只要擴充SM & TPC的規模即可。
可以增加shader processor數量與增設指令、增加share memory容量、增加TMU數量與功能(G9x已經做了)….在這些範疇內都還算是G8x的結構延伸。

於是在維持CUDA相容性的狀況下,其實還是有很多東西可以增補,調整。
為了培育CUDA的軟體資源,NVIDIA短期內的GPU結構變化,大概是離不開「TPC內部的擴充」、和「GPU總TPC數量的擴充」這兩項了。

當然這樣講的話,好像GPU就只剩shader變化一樣,只是shader之外的變化其實真的就不容易看出來,因為大家也不知道shader以外的繪圖技術進展會朝向什麼方向。
比方說NVIDIA資助發展的Logarithmic Shadow Maps,讓shadow map的效率進一步提升,而只需要Triangle Setup Engine支援對數值域即可,這不是研究的最尖端本來就很難知道。

http://www.xbitlabs.com/articles/video/display/radeon-hd3870-hd3850.html
Xbit針對RV670與G92/G80的benchmark。

可以看出G92相對於G80的底層還是有很多改善….比方說分支性能就有改善。

Crysis patch 1.1心得

關於Crysis patch 1.1聽到一些反應:

1. 性能改善,特別是SLI性能。
88GS 768MB SLI因此獲得非常大的幫助,1920×1200下的Very High提升60%以上,衝贏8800GTX….
而且總價低於8800GTX(在大陸的報價似乎是1500RMB,差不多是一張88GTS-512的錢)

2. CSAA支援
性能與4x MSAA幾乎相同(+/- 0.5fps),但是品質與8x MSAA不相上下….有非常大的吸引力。

http://myevilprocessor.blogspot.com/2008/01/geforce-9600-gt-benchmark.html
GeForce 9600 GT Benchmark by 原力部落格
在06底下9600GT和88GT-256/512還蠻接近的:

  • Core : 650MHz
  • Shader : 1625MHz
  • Memory:1800MHz
  • Stream Processors : 64

和88GS 768MB一樣,記憶體容量增加讓高解析度底下的性能跌幅減緩,只是Driver似乎還有問題,所以2560×1600還是會暴跌,這可能要Driver完善之後才能解決。

[EDIT]
http://nueda.main.jp/blog/archives/003216.html
GeForce 9600 GT(G94)のレビュー

G94的die size有240mm(16 x 15 mm^2)?!
有G92的73%、只比G84小一點點,但是這是65nm啊XD

CES2008:次世代光碟戰爭今年結束?

http://www.watch.impress.co.jp/av/docs/20080107/ces05.htm
ソニーのストリンガーCEOがBlu-ray勝利をアピール
-「すべてをブルーに覆った」。有機ELテレビを米国発売

同時公布了VAIO VGX-TP20/25系列、BDU-X10S。
前者是以前TP1的BD-ROM裝備版,CPU換成Core2Duo T8100,HDMI也改為支援1080p的版本;後者的話是預計200usd以下的新BD-ROM。
現在就看DT1要不要跟著改版上去….

米Funai、北米に300ドルを切るBlu-rayプレーヤーを投入
-Profile1.1対応。HDサラウンドのストリーム出力も

第一台真正的”低價BD-Player”終於要推出了。

http://www.watch.impress.co.jp/av/docs/20080107/sce.htm
SCE、1月下旬のアップデートでPSPに「Skype」機能追加
-新型PSP専用。マイクも発売し、固定/携帯電話と通話も

PSP-2000專用,沿用以前的USB mic,提供通話功能,似乎沒有IM功能。

http://plusd.itmedia.co.jp/lifestyle/articles/0801/08/news010.html
「Cellテレビ」、登場は意外に早い?

Toshiba展出內建CELL的電視試作機。
由於這次展示的是90nm 的CELL,所以其實發熱量算是相當大。
SpursEngine的性能自然是受限許多,比較沒辦法做這次的48個SD stream同時decode/合成。

http://pc.watch.impress.co.jp/docs/2008/0108/ces03.htm
NVIDIA、Hybrid SLI技術を正式発表
~PureVideo HDはVC-1アクセラレーションを実装

第一個功能是「Hybrid Power」,這是在NVIDIA的內建顯示晶片組上可以利用的功能:在使用NVIDIA內建顯示晶片組與繪圖晶片的前提下,才能使用Hybrid SLI。
在省電狀態下,GPU會完全關閉,只使用內建晶片組來運作;在高負荷繪圖工作下,GPU會輸出到內建晶片組的繪圖晶片上(應該是render到主記憶體),然後再由內建繪圖輸出。
所以不必改變輸出介面,在這個狀況下GPU完全是”加速”用,而且既有的SLI模式可以結合這個功能使用;缺點是由於它需要GPU的PCIe介面可以接受SMBUS的電源管理訊號,所以沒有這種設計的現行產品應該都無法支援這個功能。

這有個隱藏的功能,就是GPU不需要在本地端:
你可以用 infiniband接一大堆GPU給一台PC用,某種意味上是虛擬化。
軟體隨便你怎麼做,但是如何不改以前的遊戲hack過去的話可能要動到Driver model,這樣應該會費不少周章。
但是那種外接的1U GPU rack就可以拿來跑遊戲了w

第二個功能是「GeForce Boost」,基本上就是傳統的SLI,只是讓不同種類的GPU可以在一起運作,所以現在分離GPU可以來幫助內建顯示,得到效能boost。

此外,同時發表的「GeForce 8200」,就是先前已經傳聞一段時間的MCP7A,具備PureVideo HD2,能夠做VC-1加速;
值得注意的是,PCWatch的說法是VC-1加速是「Driver實作」,也就是硬體沒有改變。
同時發表的nForce700a似乎也是具備這個繪圖核心的樣子,支援Hybrid SLI,並宣稱與GeForce 8200「性能不變」,只是沒有提到有沒有PureVideo HD2。