It still remember the 9800SE version, I softmodded mine to a total of 16 pipelines. It was awesome. Then it was replaced with a Geforce 6800.

It think it had 8 pipelines instead of 12. So you got a near 50 percent boost when you softmodded the card. Dem nais.

I had an X1800XT, bought it when it first came out, it cost about 1300 ringgit. Dem expensif, but very song play games. But then like yours, mine overheated, artifacted and then mati katak.

HBM is interesting. But knowing how they dealt with the challenges of DRAM stacking is even more interesting.

But here are my predictions;

- Memory overclocking will be quite shit. Simply the issue of multi layering DRAM makes routing and vias a huge issue and signal integrity and timing will be tight. But this is a trade of for shorter delays.

- Thermal management. Sharing the same substrate is an issue as thermal density increases, also proximity to the core which has higher leakage current than the mem poses thermal management issues.

- Price. Yields will be crap, I do not expect prices to do down until a process improvement takes place and that will take longer than the usual cycle.

Waiting for a beyond3d or Arstechnica article, but not holding my breath haha!

Not necessarily. The performance would be pretty good out of the box, just that OCing would be an issue. I personally don't overclock most of the time as I don't see a point to it.

Above all the multiple die single package system would save a lot of board space and increase density, which means interesting prospects for SFF systems.

I never have and probably never will take into context OC performance when it comes to any review. A card that has a lot of OC potential means that production quality control isn't tight (thus the high spread) and thus reliability is more variable. Because if manufacturers could reliably clock their chips higher and bin the higher performance parts, they'd do it because it means higher profits.

When the original GTX reference cards came out and people found that they had significant headway, Nvidia started allowing most AIB partners stock OC the cards as yields and quality tightened up. Now all cards are overclocked from the box when compared to the original speed.

I feel no need to OC, if I wanted a faster Porsche, I'd just buy the next faster model. Much like if I feel my M235i is too slow, I'd order an M4

It's not black and white, as depending on the inherent efficiency of the rendering interface and any additional abstraction steps that the driver may or may not have to make, a choice of Mantle or DX can affect things.

Graphics is a lot more complex than just a simple SIP pipeline. For example a driver might have issues with texture calls and transfers being the most cycle heavy, thus the use of a CTM (Close to Metal) like Mantle where you can actively manage hierarchy can reduce CPU usage as a whole even on the driver side of things. This is dependency in action.

True. But at the end of the day, if by reducing abstraction and increasing efficiency with Mantle say reduces certain flow on effects that the driver might not be efficient with, then yes, there are effects.

Quoting Mike Houston (now a chief engineer at NVidia) when commenting on API performance overheads.

His comment on the small batch problem generally stems from SIMD Vector widths I believe, which an API mostly abstracts, but with a CTM API can optimize.

TL;DR, it's all related at the end of the day. You can change something high up the chain like a compiler, but the effects will be felt everywhere.

Dunno why ppl until dai dai wanna argue driver overhead on CPU, just go buy 8 core Xeon, kaotim

Hmmmm, might be a good purchase lol

I always use the FMUL and MADD feature as a good analogy on how optimization can have a flow on effect from drivers to final processing.

For example, you have an instruction that states you want the answer to 4x4, processor A can complete a FMUL in 4 cycles while it can finish a MADD in 1 cycle. Instead of using a single MUL, you can just use 4 ADD and get the same answer in half the cycles (4+4+4+4 = 16 wat)

But what if using that many add cycles saturated the mem bus? or overtaxed the queue? That's why I believe the discussion about driver/api overhead is slightly moot without an understanding of the underlying principles of computer engineering.

I do not believe that all is lost, given Carizzo has interested me greatly I see that it has great potential for optimization and class beating integrated graphics.

What I think they're doing is "buying time", much like how Intel bought time post Merom to get the processor enhanced and well built enough for the desktop environment by bumping up the raw speed on the Prescott Pentium 4 processors.

AMD know where they want to go with GCN, they just need more time. That's why it's an R9 390 and not an R10

Quick and dirty answer, no idea. It depends on a lot of things, BUT more RAM is always nicer.

The most likely answer is that AMD have calculated that high performance nature of HBM and the higher bandwidth offered by processor mediated DMA and PCI-E 3.0 should be enough to cover for things.

Technical explanation.

Hmmm everything I see is a card that is very much a gamble on DX12's CTM (Close to Metal) philosophy, which I feel is a natural progression of graphics card architecture, much like Mike Houston predicted a few years back when he was still at AMD.

But what is going to be interesting is their ability to pull it off given that DX12 games are still in the pipeline and optimization always takes time. ATI was first to market with built in tessellation with the R5xx series but it didn't do jack shit for performance simply because devs needed another gen to catch up via optimization.

My bet though is because most game devs are quite in tune with optimization for GCN due to AMD's omnipresence in the console market that AMD is betting that any performance gains will come fast and scale quickly, coupled with the increased memory efficiency gained from GCN 1.2. It's a bold gamble, not sure it'll work though.

You guis tell me if you get any review samples!

ATI has always been pretty innovative with a lot of their architecture enhancements, they were first with AVIVO where the GC took up video processing and started the trend, but for 1 year plus only 1 piece of software supported it, they were first with non-linear pipeline architecture and a lot of other things both companies use as the norm these days. NVidia has always bided its time when comes to these things, preferring market maturity. That's why Maxwell debuted in the GTX 750 TI for more than a year before being scaled up.

It's like how AMD foresaw the need for even more Compute performance and wasted a significant amount of die space while NVidia cut it down in Maxwell. Compute performance is the way of the future if you ask me, but AMD is jumping the gun a bit early for mass market.

AMD has always been crap with samples and if you gave them a shit review, you'd be magically bumped to the bottom of the sample list next time around. They sure knew how to make enemies back then. But that was like 8 years ago. Maybe their marketing dept is a little smarter now.

Looks liek Tantalum Polymer capacitors near the power capacitors (eg; Panasonic/Sanyo SMD POSCAP). Might be wrong.

Also the increased compactness of the board and a superior VRM might negate the need for bulk capacitance.

That remains to be seen.

It was true when I did it for money. Nowadays computers and gaming are one of my side hobbies, I like discussing it more than I like reviewing it, I prefer spending my money on cars and guns.

Naw brah, 3DFX brought that in with the Voodoo 2, so when they were bought over by NVidia, NV had access to all their expertise. But yeah, ATi's solution was effin retarded.

I oso not up to date, once in a while when something new comes out, I come out of my tech cave and find out that I'm generations behind in hardware. My last PC lasted me from 2008 to 2014 before that sial DFI mobo meletop

It's not a controlled item and thus they can set any price they want ler.

I know the law, but the law also requires that the defined "price controller" in this case a ministry officer make full review of the "Goods" under review, it can only apply to 1 item at a time and can be subject to review because computer items are not under a "gazette"

By the time this process is done, the R9 400 series will be out already haha!

You're just better off waiting 3 months for the price to normalize than the 1 year + it'll take for a review.

When it comes to buying graphic cards, I only have 1 rule,

1) If it can run the games you want at the speed you want it at, it's all good.

Worse comes to worst, can just jual