Nvidia 9300/9400 chipsets exposed

Update
While miffed that Charlie has gone and got the scoop before it was ready to announce it - as usual - Nvidia also asked us to point out that the 9300 and 9400 chipsets will launch on October 15th, somewhat later than it originally hoped, evidently.

NVIDIA WILL RELEASE the 9300 and 9400 chipsets tomorrow after a long delay, and they put it out accompanied by the now standard sleazy and half-true presentation. Read on for all the details.

As usual, we will go through it bit by bit, with some not included numbers here and there. It is a yawner, and Nvidia seriously cherry picks the data, bends it past the breaking point in places, and in general, acts like we expect them to. Nothing we haven't seen before, but if any NV people are reading this, check this link.

Anyway, on with the show. Slide 1 is entitled "Introducing GeForce 9-Series mGPUs", subtitled "Game Changing Performance September 2008". mGPUs are their new term for integrated graphics, they are trying to call them 'motherboard GPUs', one of the dumber names I have seen in a long time. I guess they are trying to make people think that integrated parts don't suck by changing the name. If you hear anyone using the term, please, mock them for me.

Slide 2 is when the NDAs go up, Sept 30, 9am EST, 2pm GMT, and 9PM in Taipei, Taiwan. Yay, now you know the times. They then go on to say in Slide 3, titled GeForce 9-Series mGPUs for Intel that they are "5x Faster!!", with two exclamation marks for slower reviewers. What they neglect to say is what it is 5x faster than, a pretty good indication that they are making numbers up. I would point out that it is several hundred times faster than the old VLB Tseng Labs VGA card I have on my shelf, and probably throws out millions more polygons than the Permedia PCI card I have as well.

It then slimes over to a box with GPU+, with CUDA and PhysX logos on it. Wow, who would have guessed they are going to cheat on benchmarks with integrated GPUs now as well? The next box has 70+% Boost with a Hybrid SLI logo below that. Looks like they (maybe) finally got it working a year after ATI. Good job, I guess they didn't lose all their engineers to Intel of late. I wonder if it really works this time? Given their 'hybrid power' track record, I doubt it.

The last box is "Blu-Ray that works", and has a Purevideo Logo as well as a Blu-Ray logo. Wow, cool one there guys, you finally caught up to ATI's year and a half old integrated parts, pat yourselves on the back.

Slide 4 is where the sleaze really begins, and it is all about 3DMark Vantage, a benchmark that is so broken and gamed at this point it is useless. The slide "A New Class of mGPU Performance" takes uneven and skewed benchmarks to a whole new level. It uses 3DMark Vantage, which uses an Nvidia-owned API, and shows 'superiority' with it. It is not an apples to apples comparison because they are cheating by owning PhysX and using it when the Intel parts can not.

They don't show any game numbers of course, because that would expose the lies. Instead, Nvidia shows a G45, ATI 780G and a 9400, all using the P-level Vantage tests. Remember, this is the one that shows off Nvidia cheating the most, and they don't show the test meant for integrated graphics, the low end 'entry' test.

That said, the scores are P104 for the G45, P262 for the 780G, and P575 for the 9400. They don't have the cahones to break up the numbers so you can see how badly they cheat on the physics, nor do they show the system configurations. Imagine that.

If you run the entry level tests independently on a 9300, you get E3600 and ~1800 on 3DMark06. In contrast, an ATI 790GX scores E2700 and 2000, all number rounded off to protect the tester. I have personally tested the 790GX, and gotten much higher 3DMark06 scores.

So, how is NV cheating? Several ways. First, on they are comparing vastly different CPUs across architectures, the current Phenoms are woefully underpowered compared to a top end Intel Core Number Numeral Quad, and that can bump 3DMark06 scores by hundreds of points. Second, NV has replaced the PhysX DLLs with custom versions, so the workloads are quite different. You can see that in my numbers, the 3DMark06, NV DLL-free, has the 9300 about 200 points lower than a 790GX, about 10 per cent. The Vantage numbers are 33 per cent higher. Hmm, what does that mean? See the honesty link above.

Slide 5 is refreshingly honest. It is titled "30 FPS playability across top games", and it shows greater than 30 FPS with a 9300 and the games on minimum settings. Some of the 20+ are at 1024*768, others are at 800*600, but it is a plausible claim. Crysis at 8*6 is hardly what I would call a satisfying gaming experience, but it is technically doable.

"14 of Top 30 PC Games Fail on Intel Graphics Even on lowest resolution, lowest settings" claims slide 6. This is using the broken OS, 32-bit with old drivers, and using NPD sales reports from June. Need I say that they should test with the latest drivers when making that claim? Pshaw, that would be ethical, and Nvidia seems to frown upon using such words, much less acting on them.

That said, of the games pictured in the slide, several are flat out inacurate, Age of Conan and Crysis DX9 work perfectly, and two others, Bioshock and Half-Life 2 have a problem with alt-tabbing out. To call that a game failing is a pushing the bounds of good taste. Independent testing does not back up the Nvidia claims here, not by a lot.

Slide 7 picks on Assasins Creed, an Nvidia sponsored game mind you. It shows it crashing on a computer from the 'competition' in a blue box. If they had used the latest drivers, this would not have happened, it has been fixed for a long time. Can you say dishonest? Yes, you can, the green goblins might not understand though.

Age of Conan is the topic of slide 8, and they show that you can play on higher settings with a 9400 than a G45. Yup you can. You can also play with higher settings on a 4870, at more frames. The 9400 is a better gaming GPU than a G45, is that really news? Slides 9 and 10 do the same to Second Life and Spore, basically you can play them at higher settings on a 9400. Shock.

Slide 11 gets back to naked sleaze once again. They take two $445 systems, identical except for the CPU and the mobo. The difference is that the 'Intel' system has a G31 GPU and a Core 2 7200 CPU. on the 'Nvidia' machine, they have a 9300 board and a Pentium DC E2180. Costs on the parts are the same in total, $50, $120, $100 and $70.

Wow, talk about unethical. They picked the slowest possible three generation old Intel IGP and compared it to a non-released cutting edge NV part. That is like putting a Premier League football team up against a wheelchair basketball team and calling things fair.

Imagine for a moment if you had used the same CPU and gone to Newegg and bought a $50 ATI 3650? Who do you think would win that one? Handily. Crushingly even, scoring around E6750 and 4100. Nvidia can't seem to attempt fair comparisons any more, I wonder why? You also have to question the $100 for an IGP mobo, not only is it really high, but NV has a habit of promising the press rather optimistic numbers before release. Keep an eye on what they really end up selling for in two weeks.

Slide 12 is equally laughable, it says that the CPU is two cores and the NV IGP is a '16 core CUDA processor'. Yup, sure. And all the software out there runs on x86, not cuda. Actually, no software other than an NV paid-for demo or two runs on cuda, so the slide of 2 x86 cores plus '16 cores' makes little sense.

13 compares one of those paid-for demos, Badaboom, against iTunes using two CPUs. The more expensive $158 Intel E8400 is slower than the $80 E2180, presumably plus a 9400, at encoding a 1280*720 clip to 640*360. They don't point out that the 9300 costs $100 a few slides before, and a 9400 presumably costs a lot more, blowing any value proposition out of the water. Then there is the non-trivial problem that you can't do a high-quality encode on a cuda-equipped program for some reason, but they don't even mention that in passing.

To make matters funnier, what happened to the 200x speedups in doing anything? The graphs show a less than 2x speedup, much closer to 80 per cent or so, vs a very slow CPU. Yawn. Slide 14 does more of this and shows that there are five, count them... five, paid-for demos that use GPU acceleration. They are the aforementioned Badaboom, folding@home, MotionDSP (NV owns a chunk of it), TMPGenc and TotalMedia Theater. Be still my beating heart, five apps vs how many for x86? I think the difference is more than five orders of magnitude.

Slide 16 proudly proclaims to the world that they cheat on 3DMark vantage with PhysX. We knew that already though. The NV PR staff should run for goverment, they have the right mindset.

17 shows that there is a "70 per cent boost" with hybrid SLI if you add an 8400 or 8500 GPU, not backed up by any numbers. We will see if those numbers are real, or it actually works at all, NV's track record on such matters is spotty on a good day.

Slide 18 shows 11 games with a 9300 + 8400GS beating a vanilla 8400GS by such margins. I wonder why they can show 'proof' in slides like this, but never against ATI or Intel? Odd.

19 shows the same 70 per cent number with pictures, and screenshots do illustrate frame rates quite well.

Slide 20 is entitled "Blu-Ray that Works", and has four boxes. They are Lowest CPU Utilisation, Industry Leading Video Quality, Across All Blu-Ray titles, and Unmatched 7.1 audio fidelity. None are accompanied by any data at all, and although it may be flogging a dead horse, NV isn't exactly known for honesty here.

The video quality claim uses HQV, a subjective scoring system. It is basically saying that 10/10 people on the Nvidia payroll found their playback to be perfect, take their word for it. The 7.1 audio is quite the ballsy claim considering that both ATI and Intel have had this for a long time, and NV is laughably late to the game. It is a win because 10/10 people on our payroll say it is!

21 shows Windows CPU utilisation graphs running Cyberlink PowerDVD, not really against anything, just curiously showing CPU utilisation with three different Intel CPUs. Not sure what they were trying to prove, but they did make the slide. The next slide shows that they finally have caught up with ATI and Intel to put all the decode in hardware, VC-1 is finally done! Months after Intel. Years after ATI. Victory!

The next slide, 23, shows that they can do 7.1 PCM over HDMI, again months and years after Intel and ATI, respectively. The Americans are nowhere near Baghdad, we will crush their tanks...

That is followed by a slide showing that the G45 has problems with a very specific HD mode, 1080p/24. Nothing new there, G45 is broken on that mode, but they don't mention ATI who has better quality, is cheaper, and came out more than a year ago. 25 points this out with pictures in case any of the tame reviewers can't read big words.

Slide 26 is a block diagram of the 9300, and it is pretty standard. It has a 1333FSB, DDR2 and DDR3 controllers, up to 800MHz and 1333MHz respectively, a 16x PCIe port, four PCIe 1x slots, and six SATA. I wonder if the SATA ports come with a "Don't corrupt data" option in the BIOS this time around? There are also five vanilla PCI slots.

For video, there is DL-DVI, HDMI, and Displayport as well as VGA. Dirt standard stuff here. It has Azalia audio, 12 USB ports, and a GigE port with " Firstpacket Technology". Given NV's track record on things like this, see TCP/IP acceleration, firewalling, channel bonding etc etc, I wouldn't count on it actually functioning, much less pay extra for it.

Slide 27 is the raw specs of the 9400 and the 9300. They both have '16 cores'. and the 9400 runs at 580/1400 while the 9300 is at 450/1200. Both will do 2560*1600 rez, have Purevideo, 7.1 sound, and all the ports listed above. In fact, they list the things in the block diagram with only RAID 0, 1, 0+1 and 5 as really new. They do not list the "do not corrupt" BIOS option here either, so use at your own risk.

Slide 28 shows a picture of the back of the mobo, and it indeed has all the ports claimed. What is the world coming to? The boards on slide 29 list four launch 9400 from ASUS, EVGA, Gigabyte and J&W. They are, respectively, the P5N7A-EM-HDMI, 113-YW-E173-A1, GA-7AUM-S2H and the JWN7AM01. There are also seven 9300 boards listed from ASUS, Biostar, ECS, EVGA, Foxconn, J&W and MSI.

Slide 30 is a point for point copy of slide 3, understandable considering how content-free the whole deck is, they probably get paid by the page. Slide 31 copies slide 2, exactly, I guess someone needs to make up for those tanking stock options. Lastly slide 32 says "Thank you". No, thank you guys for stiffing us on another briefing, we are honestly thankful.

Overall, the 9300/9400 moves the bar a little. It looks to be about the same speed as an ATI 780G/790GX with scores pumped up by the sheer power of the Intel CPUs over the equivalent AMD models. Nothing special, Nvidia is now caught up with the same tech as the other guys, just a year later. Nothing new there.

The listed prices are notably higher than the 780G line, something that won't help matters at all. The real question is what is the TDP, a spec usually in the slide decks, but notably absent this time, is. The chipset is 6+ months late, during CeBit, OEMs were saying that the main holdup is heat. One has to wonder if Nvidia just gave up on efficiency to get speed, or actually did the engineering that it is notably unable to do on other parts? If it just jacked up the power, will the chips live?

Don't expect much from the launch Tuesday, it will beat the Intel G45, but that is not a shock. The only real test is whether or not it can beat a 780G given the same CPU resources and wattage. µ