Sold products that aren't in consumers hands? That's what's generally called the channel inventory.
It's amazing the things you can find out by reading conference call transcripts:
and
- Status
It's amazing the things you can find out by reading conference call transcripts:
Jawed
Legend
The channel ends once it leaves the AIB's hands (i.e. is purchased by a retailer), as NVidia doesn't sell consumer cards to the public.
Jawed
No, the channel consists of all the parts that have not reached their final destination, the consumer, after they have been sold by the initial producer (i.e. Nvidia). That is, all the parts in the pipeline over which they have no control and limited visibility. A box with a GT250 on the shelf of Fry's is part of channel inventory. This doesn't affect the bottom line in anyway, but too much of it can result in excess inventory in the longer term if demand suddenly drops (due to the 3 month or so lead time in wafer starts.)
Jawed
Legend
OK, the channel inventory NVidia can discuss with analysts ends with what leaves its customers' hands.
Someone's still selling FX5200s:
http://www.google.com/products/cata...&cid=14852880162739718024&sa=title&os=sellers
I wonder if NVidia is still selling those?
Jawed
Someone's still selling FX5200s:
http://www.google.com/products/cata...&cid=14852880162739718024&sa=title&os=sellers
I wonder if NVidia is still selling those?
Jawed
The danger about the channel is about sudden drops in channel velocity. A product like FX5200 already has a velocity of 0, so it doesn't matter.
Nvidia just can't catch a break.
*GT200 the bomb detector, not the GPU*
http://www.cnn.com/2010/WORLD/meast/02/23/iraq.bomb.detectors/index.html?hpt=T2
*GT200 the bomb detector, not the GPU*
Tegra and Ion
Forgot to post this the other day, from Techeye:
The article goes on to speculate about the future of Ion and the prospects of some kind of tablet device.
Would not like to be in above department at the moment, imagine it is chaos. People walking round everywhere with no idea, if they are lucky the new management positions are being decided by arm wrestles as opposed to the usual full blown fist fights the situation normally generates
Forgot to post this the other day, from Techeye:
The article goes on to speculate about the future of Ion and the prospects of some kind of tablet device.
Would not like to be in above department at the moment, imagine it is chaos. People walking round everywhere with no idea, if they are lucky the new management positions are being decided by arm wrestles as opposed to the usual full blown fist fights the situation normally generates
I could be stupid, but I don't see what's so bad or unexpected here (this was confirmed by NV PR to X-bit labs btw). Obviously NV would be much better off if there wasn't any trend towards IGP integration, but they can't prevent it at this point. So despite the rumor mill being rather clueless, MCP89 has been ramping for a while (final silicon) after having taped-out around late 3Q09 afaict. This is a 40nm/Socket 775/128-bit DDR3/48 SP(i.e. GT216)/6Gb SATA chip. Given the death of Socket 775, the lack of a DMI license, and how pointless DMI or AMD chipsets would be anyway given their integration trends, there's not much else this team could have done in x86 after that (and yes, Apple will be using MCP89 in their mainstream products among many other OEMs).
Also the MCP team wasn't so big anymore. Tegra had around 500, so if it's now 650 - does that imply MCP only had 150 people left? Certainly not impossible given some things I heard, although I still think there was some double accounting - for example, Bangalore was recently working on both MCP and Tegra memory controllers iirc. It'd be more interesting to have a manhours number, which obviously we'll never get.
This clearly is not a management shuffle, it's more one engineering team absorbing another. These engineers should be useful for Tegra3 (taping-out around June 2010) and Tegra4. Expertise-wise, MCP and Tegra are getting more similar. Tegra3 should support 64-bit DDR3 and a G80-inspired CUDA GPU, for example. The other two main MCP design tasks are very similar to Tegra's: southbridge design and SoC integration (plus their associated S/W). That doesn't mean it's a perfect fit, and once again the fact that NV is giving up completely on future x86 chipsets isn't good news, but this shouldn't be unexpected and it's certainly not a bad approach given the lack of options at their disposal.
A final intriguing point I'd like to make: if NV ever gets their hands on a x86 core and license for SoC integration, by that time there might not be any major other difference left between a Tegra ARM SoC for netbooks/tablets and a Tegra x86 SoC for netbooks/small notebooks/media centers. And the ARM SoC would be significantly faster CPU-wise, too.
Also the MCP team wasn't so big anymore. Tegra had around 500, so if it's now 650 - does that imply MCP only had 150 people left? Certainly not impossible given some things I heard, although I still think there was some double accounting - for example, Bangalore was recently working on both MCP and Tegra memory controllers iirc. It'd be more interesting to have a manhours number, which obviously we'll never get.
This clearly is not a management shuffle, it's more one engineering team absorbing another. These engineers should be useful for Tegra3 (taping-out around June 2010) and Tegra4. Expertise-wise, MCP and Tegra are getting more similar. Tegra3 should support 64-bit DDR3 and a G80-inspired CUDA GPU, for example. The other two main MCP design tasks are very similar to Tegra's: southbridge design and SoC integration (plus their associated S/W). That doesn't mean it's a perfect fit, and once again the fact that NV is giving up completely on future x86 chipsets isn't good news, but this shouldn't be unexpected and it's certainly not a bad approach given the lack of options at their disposal.
A final intriguing point I'd like to make: if NV ever gets their hands on a x86 core and license for SoC integration, by that time there might not be any major other difference left between a Tegra ARM SoC for netbooks/tablets and a Tegra x86 SoC for netbooks/small notebooks/media centers. And the ARM SoC would be significantly faster CPU-wise, too.
Perhaps, but in absence of an x86 license, don't you think nv will be left out in the cold in the <$100 gpu market, as fusion chomps away at this market. Which would make an x86 license particularly valuable to them even if the only difference between a Tegra and an 86gra is the cpu, and 86gra is lame in comparison.
Starting in 2H11 when Socket 775 really starts being phased out, yes, certainly. There's still the 40nm dual-core VIA Nano though, but it would obviously be slightly absurd to expect as many units for that.
'86gra'... I like it!
But sure, they'd prefer competing in both places at once. It just won't happen anytime soon.There's still the question of what their ex-Transmeta engineers are working on. I'm not sure if I've said this before, but my *guess* is they are designing a specialized processor to transcode x86 code to ARM on the fly, and they'd want to integrate that along with ARM Eagle cores (>Core 2 Duo perf/clock for native code afaik) on a future Tegra SoC. But as I said, that's just a guess, and there are problems with it (e.g. it's harder to transcode to an ISA that wasn't designed with that in mind) but it'd certainly spice up things a lot, while probably costing less and risking less than a fully custom x86 core.
So you're sure that they have a backup strategy in place to fight back against fusion. Wouldn't a VIA86 based SoC be a cheaper/faster answer?Starting in 2H11 when Socket 775 really starts being phased out, yes, certainly. There's still the 40nm dual-core VIA Nano though, but it would obviously be slightly absurd to expect as many units for that.
'86gra'... I like it!
Eagle, faster than conroe, sounds great. When are devices having it expected to ship?
I'm not sure how Tegra3 or a two-chip platform with VIA is a 'backup plan', it's just a separate thing that would happen either way, so I'll assume you are referring to my speculation about what the ex-Transmeta engineers were working on. And no, I have no idea if they have any backup plan of that sort; it's genuinely pure speculation. It's also very much possible that they've got a small team dedicated to that but that they'll never ramp up/finish the project. I am personally *very* skeptical that Jen-Hsun would invest in creating their own CPU even if based on transcoding though, which is why I come up with that idea practically by elimination.
ARM will deliver final RTL to partners around 1H11 iirc (or am I confusing that with another of their new cores? subtract or add one quarter max either way), whereas they delivered RTL for the Cortex-A9 sometime in 1H08. So basically add three years from when Cortex-A9 devices will be available, which is around mid-2010, and realistically that means mid-2013 for the first Eagle-based devices in retail. There's still plenty of room for Cortex-A9 to improve before then as triple gate oxide and/or High-K plus improved design methodologies/experience could result in much nicer clock speeds.
Interestingly, unless TSMC becomes more aggressive for 22nm, Eagle might still be on 28nm, ugh! It's strange really; they certainly aren't going to use anything else than double-immersion initially on 22nm (they're still on single-immersion for 28nm!) so I don't see why they'd need to delay 22nm to 3Q12/1Q13 for HP/LP respectively, unless they're planning to do something funky at the gate level that they haven't yet announced. Intel is still on track for 22nm in 4Q11, so they're basically losing back all the time they gained in the last several generations in a single step for a very incremental node - it's crazy. Here's hoping they are just bluffing...
By a backup plan I meant some kind of strategy to remain in the low end desktop/laptop market, as fusion is all set to nuke this market. With a large profit/volume market dead, I am not sure whether nv's discrete gpu business is going to have enough legs to stand upon.
It's also very much possible that they've got a small team dedicated to that but that they'll never ramp up/finish the project. I am personally *very* skeptical that Jen-Hsun would invest in creating their own CPU even if based on transcoding though, which is why I come up with that idea practically by elimination.
Mid 2013
that's 7 years after conroe, Gee, things really move slowly in the embedded world.aaronspink
Veteran
2043. So just around the corner.
And what is the basis for this estimate of yours?
aaronspink
Veteran
about the same basis as the original comment, none. Its always nice to claim something, but rarely sufficient. Considering the issues ARM has had in bringing out A9 vs previous designs, the confidence certainly isn't there.
I don't have the time or the sources to know about that kind of thing in any detail, so I'll just take you at your word that A9 had many more early issues than previous ARM designs. On the bright side... at least the end result seems very satisfactory, unlike the A8 which didn't really meet its goals.
Because I am basing my speculative schedule on the time it took for the A9 from RTL delivery to device availability though, I am not sure how important this is unless you're suggesting that Eagle might plausibly take even substantially more. Or are you just (and very rightfully) pointing out that RTL delivery schedules are probably too optimistic and they'll very likely slip? Or maybe just that performance targets won't be met... All of those are possible, and you're right to be a bit skeptical, although personally I'd rather not be so very cynical!
Starting in 2H11 when Socket 775 really starts being phased out, yes, certainly. There's still the 40nm dual-core VIA Nano though, but it would obviously be slightly absurd to expect as many units for that.
'86gra'... I like it!
There's still the question of what their ex-Transmeta engineers are working on. I'm not sure if I've said this before, but my *guess* is they are designing a specialized processor to transcode x86 code to ARM on the fly, and they'd want to integrate that along with ARM Eagle cores (>Core 2 Duo perf/clock for native code afaik) on a future Tegra SoC. But as I said, that's just a guess, and there are problems with it (e.g. it's harder to transcode to an ISA that wasn't designed with that in mind) but it'd certainly spice up things a lot, while probably costing less and risking less than a fully custom x86 core.
I'd be surprised if ARM was going to produce something with greater perf/GHz than Core2 in a reasonable frequency range any time soon. The A9 isn't even close, it's 2-issue OOE, with 1 LSU. Although surprisingly the controller supports upto 8MB L2!
I believe ARM can go wider, but they'll have a hard time keeping power down while creating a high performance memory hierarchy.
Also, what software and workload are you talking about? ARM still quotes performance numbers using Dhrystone MIPS, which is basically useless. For some reason the embedded world didn't get the memo in the late 1980's that synthetic toy benchmarks are not helpful...
It's quite possible that ARM can produce a high performance microarchitecture, they'll have had 3 shrinks since 65nm, so that gives an awful lot of room for improvement.
DK
- Status
Similar threads
