Slashdot Mirror
NVIDIA's New Flagship GeForce GTX 580 Tested
MojoKid writes "Even before NVIDIA's GF100 GPU-based GeForce GTX 480 officially arrived, there were a myriad of reports claiming the cards would be hot, loud, and consume a lot of power. Of course, NVIDIA knew that well before the first card ever hit store shelves, so the company got to work on a revision of the GPU and card itself that would attempt to address these concerns. Today the company has launched the GeForce GTX 580 and as its name suggests, it's a next-gen product, but the GF110 GPU powering the card is largely unchanged from the GF100 in terms of its features. However, refinements have been made to the design and manufacturing of the chip, along with its cooling solution and PCB. In short, the GeForce GTX 580 turned out to be the fastest, single-GPU on the market currently. It can put up in-game benchmark scores between 30% and 50% faster than AMD's current flagship single-GPU, the Radeon HD 5870. Take synthetic tests like Unigine into account and the GTX 580 can be up to twice as fast."
http://www.pcper.com/article.php?aid=1034
http://www.hardocp.com/article/2010/11/09/nvidia_geforce_gtx_580_video_card_review
http://www.anandtech.com/show/4008/nvidias-geforce-gtx-580
http://www.legitreviews.com/article/1461/1/
http://www.techreport.com/articles.x/19934
http://www.bit-tech.net/hardware/graphics/2010/11/09/nvidia-geforce-gtx-580-review/1
I am very glad to see the performance crown handed back and forth.
Now if only this was happening in the CPU market...
This GTX580 is a 3 billion transistor chip(not counting the RAM on the same card, just the GPU die itself). Does anybody know what year the number of transistors on the entire planet reached the number on this die?
In an absolute, architectural sense, essentially never. A screamingly fast vector processor isn't going to do much for all your x86 code, and never mind all the little housekeeping chores that the CPU does(most of the modern ones include the system RAM controller(s), do a lot of peripheral wrangling, may be the root of the PCIe bus, and so forth).
In a "designing your next gaming build" sense, they largely already have. Unless you are a money-is-no-object-e-penis-must-get-longer type gamer, you can generally get better bang for your buck by going with a cheaper CPU and spending the savings on a nicer graphics card. It depends on the game, and there are situations where a truly epic(2x or 3x of the top of the line GPU ganged together with SLI or crossfire) graphics system will be CPU bound without the best CPU available; but Joe Gamer is, most of the time, better off with a third tier CPU and a second tier GPU, or a 2nd tier CPU and a 1st tier GPU.
In smaller systems(where board footprint really counts) or in cheap systems(where package costs and board size really count) the integration of CPU and GPU into a single package proceed apace, with AMD rolling low-end ATI tech into certain of their newer parts, and Intel trying to make their GMA stuff suck less. The only real wild card is Nvidia: Unlike Intel or AMD, they have no x86 cores to speak of, on the other hand, their GPU-computing initiatives are arguably the most advanced, in terms of tool and driver maturity. The question is, will they eventually produce an Nvidia equivalent to AMD and Intel's CPU/GPU combo packages(perhaps by buying VIA, who has adequate-but-deeply-unexciting x86 assets; but utter shit GPUs), or will they persist purely as a maker of high end gaming GPUs and GPU-based compute cards?
Unless the heriditary line of the "PC" as we know it is wholly extinguished, there will always be an x86 CPU floating around somewhere in the block diagram(and, in other types of systems, likely an ARM CPU); but it is already the case that, for many applications, the CPU has gotten fast enough to hit diminishing returns for many applications, and the GPU(or just the embedded h.264 decoder) is where the action is.
With a 244 watt TDP, I suspect that they need every inch of the front of the card, and are constrained only by PCIe form-factor concerns from using more of the back, just to keep the thing from burning out without a fan that sounds like a legion of the damned every time you boot the thing. The entire front of the card is a combination of heatsink(and not your extruded aluminum jobby, a phase-change vapor chamber unit) and a shroud to direct air flow.
If you want to see the board, back off a few price/performance tiers, and you'll get a 90% bare PCB with a dinky little slug of aluminum or copper on the main chip.
So, this card is about as fast, and consumes about the same power as a 480, but it's "next gen" anyway ?
That looks like a 480 with the 4 replaced by a 5. Hardly a revolution.
Just watercool the 480, it's how it's supposed to be used.
Does anyone assume that the synthetic benchmarks achieved by either AMD or NVIDIA are representative of anything more than these companies' efforts to tweak their driver sets against the pre-existing criteria for getting a "good score"?
Both companies I believe have been accused over the years of doing just that and pointing the finger at the other as taking part in shennaniganism"
I went to battle M.C. Escher, but drew a blank.
For one, there are a lot of motherboards that don't support it. Even new, reasonably high end boards. I have an Intel P35 board with a Core 2 Quad at home, but it has only 1 16x slot. At work, a Dell Precision T1500 with an i7, again only 1 16x slot. Crossfire/SLI cannot be done in these cases. You have to buy a single, heavier hitting, card if you want performance.
Also you need to do a bit more research if you think multi-card solutions work well all the time. They can, but they also can have some serious problems. Some games work great, others can't use a second card at all. There is something to be said for the simplicity of a single card that does what you need.
In terms of needing the speed? Well depends on what you have and your tastes. You certainly don't need it to play any game, all games are playable on less. However you might need it if you desire extremely high resolutions and high frame rates. If you have a 30" monitor and want to drive it at its native, beyond HD rez (2560x1600) you need some heavy hitting hardware to take care of that, particularly if you'd like the game to run nice and smooth, more around 60fps than around 30. You then need still more if you'd like to crank up anti-aliasing and so on.
Now that clearly isn't for everyone, but that's fine. There is no reason not to have a high end as well as a mid range. You should hate on people who want more performance than you do. In fact, you should thank them. Know why the 5770 is so cheap? Because the 5870 is not. That high end card financed the development of the new tech, it recouped a lot of the R&D costs, making an economical midrange card a reality.
This is why nobody seems to be able to break in and compete with nVidia and ATi in graphics. They target midrange or lower end, because development costs on high end stuff is so much. However nVidia and ATi have extremely solid mid and low range lineups, because they can take the tech on their high end cards, and scale it down.
What may happen, what AMD would like to see happen, is for GPU functions to become a part of the CPU, that GPUs go away because CPUs can do it. However that'll be because CPUs have GPU like logic in addition to their own.
The problem, as Intel found out with Larrabee, is that a cache that works well for CPU tasks does not work well for GPU tasks, and vise-versa. For a GPU the bandwidth is everything, while for a CPU its the latency that matters most.
Our CPU's L1 caches are 32K/64K in size because smaller caches have significantly smaller latencies than larger ones. Its quite obvious that a 64K cache is way too small for a GPU, which could literally process 64K of data in only a few of its clock cycles.
Intel never could solve the problem. Larrabee could either be a GPU with poor CPU-capabilities, or a multi-core CPU with poor GPU-capabilities.
Maybe in the future... not with todays memory types.
"His name was James Damore."
The /. summary ends with:
It can put up in-game benchmark scores between 30% and 50% faster than AMD's current flagship single-GPU, the Radeon HD 5870.
But if you read the original article, the one flaw in the (otherwise good) nVidia card is that is still loses to the 5970 which is -- according to the article -- 'about a year old'. So why is that other article mentioned in the summary talking about the 5870 as if its the flagship? Clearly the 5970 is. Or am I missing something?
Wood Shavings!
- Godai
..never as long as Nvidia refuses to release even a hint of documentation and insists that GNU/Linux users accept their Binary Blob World Order. I don't really care if this new card is faster than the fastest AMD card, atleast I can (ab)use those for something. I still have a Nvidia PCI (not PCIe) card on some shelf which does NOT work with the Binary Blob under GNU/Linux, nor does it work with nouveau joke of a free driver.
9/11: Never forget it was a false-flag operation
HardOCP is famous for their real gameplay ratings. They go and actually play through the game while testing performance. They then find the highest settings that the reviewer finds playable. Now while there is some subjectivity to it they do back it up with FPS numbers, and it is the same reviewer trying everything out. So it gives real, in game, actually playing, results. I find it maps nicely to what actually happens when I get a card and play games.
http://hardocp.com/article/2010/11/09/nvidia_geforce_gtx_580_video_card_review
They are definitely releasing Tegra-branded ARM SoCs that include their own GPU tech as one of the functional blocks. If that is what you mean by "hybrid", then yes.
To the best of my knowledge, though, neither Nvidia, with their ARM SoCs, nor Intel with their on-package GMAs, nor AMD with their upcoming on-die ATI tech are creating what you might call a full "hybrid"(ie. a CPU whose instruction set also includes GPU-esque instructions, like MMX or SSE on steroids). At present, they are all just more heavily integrating, for economic and latency reasons, a discrete "CPU" block and a discrete "GPU" block.
I don't think that they have much choice about "Ion 2" pretty much sucking.
With the prior generation of atoms, the usual pairing was Atom + fairly antiquated Intel chipset with GMA950 and a fairly high TDP. For just a little extra, you could pair the Atom with Nvidia's chipset instead, which had as good or better TDP and much better integrated graphics. Intel wasn't happy; but the end result was good.
With the newer generation, Intel brought most of the chipset functions onboard, and played hardball with licensing, so that "Ion 2" ended up consisting of, in essence, Nvidia's lowest-end discrete GPU added on to the system via the few PCIe lanes available. Unlike Ion, which was a genuine improvement in basically all respects other than OSS linux support, Ion2 meant higher TDP, more board space, and higher BOM.
Intel bears much of the blame for it; but Ion 2 is largely a dog, particularly when compared to the "CULV" options, which will get you a real(albeit low end) Core2 or i3 processor and a similar low end GPU for not much more than the Atom...