National LME49600 Reference Design Project

[RocketScientist]

There have been several threads here talking about the National LME49600. The biggest is this one:

The Wire Ultra-High Performance Headphone Amplifier

I'm checking if there's any interest from others in at least PC boards for a LME49600 design. I suspect most got their itch well scratched with the above project, but I wanted to at least ask.

For those not familiar with it, the LME49600 is an ultra low distortion output buffer that National pairs with one of several of their ultra low distortion audio op amps for a complete servo controlled headphone amp. The design can be found in the LME49600 datasheet here and in their application note AN-1768:

National LME49600 AN-1768 Application Note

The design I have in mind would be virtually identical to the one in the app note above with perhaps a few useful additions. It would be significantly less expensive than the OPC "Wire" version as instead of 6 expensive op amps you only need 2 dual op amps (there's no issue sharing the servo with the gain stage for each channel in a single package as National did in their reference design). And the dual parts are barely more expensive than the single versions. So that eliminates almost $20 of op amps.

I also plan to use through hole parts for everything but the LME49600 (and it's fairly big and relatively easy to solder). The dual op amps used in the National design are available in an 8 pin DIP and could even be socketed for those who like to experiment with different op amps.

I intend to fully measure the design with a Prism dScope III, conduct ABX blind listening tests, and use AudioDiffmaker with my Benchmark ADC1 as well as analog (i.e. "Hafler Distortion Test") audio differencing to evaluate the design under real world conditions.

Part of my interest in the LME49600 is to serve as the "ultra low distortion" entry in a comparison between genres of headphone amps. I hope to conduct measurements, blind ABX listening, and audio differencing on the National LME49600 reference design, a popular lower feedback discrete Class-A design (i.e. Gilmore/Kumisa/Balkishan), and a lowly single IC "cmoy" design using a decent op amp like the OPA551.

If the National reference LME49600 amp is of interest for others wanting PC boards, please let me know? If there's enough other interest it will change how much effort I put it into the PCB and documentation as well as my purchase quantity for the PCBs.

If there's a lot of interest I could also consider a group buy of parts and/or even having the boards fully made at the assembly contractor I use. If the volumes are really high, then a full SMT version can be done on their mask/pick/place/reflow lines and lower the price.

[IanAS]

If you want to do a budget version then why not use the Buf634P as that's about half the price of the identical(?) 49600.

Otherwise, my own findings using my ears, are that the plastic chips are grossly inferior to the TO99 version of the 49720 which is very inferior to two single TO99 49710HA.

So unless someone really does not care about quality and needs bargain basement, why sink the ship for a few $ of tar?

[agdr]

I'm in! I would also likely be interested in a group buy if one occurs.

I've been messing around with the LME49600 for several months, including that headamp circuit in the data sheet. I did see that National ap note on the eval board, which contains the details of the parts specifications.

Unfortunately I saw the "wire" amp thread too late to order, but opc put together a fine amp there. My uses are mainly for unbalanced input through.

My random thoughts on some of the comments and discussion so far (in the "wire" thread about this)...

I've read several postings of BUF634 vs. LME49600 and the LME seems to consistently come up as the winner. I have a BUF634 headphone amp, the previous version of the HA-INFO NG98. Sounds good, but I think the LME sounds better, just imho.

See my post in the "wire" thread about paralleling two LME49600s per channel. I've read at least one posting where that was tried and (subjectively) sounded better. The BUF634 apparently has 10R on the emitter of each output transistor. The LME49600 shows a similar resistor but I've yet to read a value. Paralleling may get that impedance down looking into the output port. At least having the option for the second chip per channel would be nice. For what it is worth, that Jim's Audio LME49710+LME49600 board on eBay (no DC servo), which seems to suffer from layout problems (noise) by some posts I've seen, has two LME49600s in parallel per channel. Would be interesting to hear it with servo and with a really good layout. Some thing to experiment with...

I have read those posts before by the (ex) National guy saying the metal can versions sounded better. I think he wrote that he was about to look into it when the staffing "bomb" went off there at National. From that little bit of info in the post I would agree that it may be good to socket the LME49720 so that a metal can version could be stuffed in. Those HAs are about $20 last I looked. Pricey, but hey still in the ballpark if they make better sound, and it would be optional anyway. But I would definitely agree that the circuit board should be able to use just the simple, cheap, plastic DIP LME49720 as an option regardless. A lot of the "sounds better" stuff seems pretty listener-specific, even if it is backed up with low distortion, low noise, etc tests. So if someone likes the sound of the low cost part, more power to them in my book.

Having the option to use 2 metal can LME49710s instead of the 20 would also be interesting, but I can't think of a good layout way to make that work offhand, while still preserving the ability to use one plastic dip 20. I'm never a fan of super packed-in circuit boards - I'm kind of heretical there and like some space between my components even if it makes the board larger - so maybe one solution is to put in holes for for the DIP socket and the can in parallel, then populate one or the other.

I've seen a few posts where people building the headamp circuit in the LME49600 datasheet wound up with 10-14mV offset on the output despite the servo circuit. Mine was about 8mV. The "wire" amp didn't seem to have that problem from what I read in the thread. I haven't tried the theory yet, but maybe much more accurate (as in 0.1%) resistors in a few crucial places could solve that problem, and/or offset trim for the input op amps. I've actually read at least one post where someone thought the VR1s were offset trim.

The datasheet circuit shows the various chip's decoupling capacitors going from rail(s) to ground. My best understanding is that is not best practice anymore, rail to rail is preferred as close to the chip (?). I'm sure someone will correct me if this isn't the case.

There actually are heatsinks for the TO-263 with solder rails that solder down next to the chip

Digi-Key - DA-T263-101E-ND (Manufacturer - DA-T263-101E)

I have a few here but haven't tried them yet. They may reduce that foil surface area needed for the LME49600 chips. The formula is given in the datasheet, of course. If using two paralleled LME49600s per channel the power would be divided and the heatsinks might really shrink things down.

Good luck with the project!

[RocketScientist]

@agdr thanks for the great tips! I'll do some more research and comment a bit later. Off the top of my head, I suspect the paralleling idea makes more sense for some headphones over others but if you're designing a "one-size-fits-all" amp, it might be worth it.

[agdr]

Good point - I have low impedance (44R) Shures. For a high(er) impedance headphones the buffer paralleling may not add anything. But again... you could make it optional on the board. If you had pads for two buffer chips per channel the builder could decide to populate one or two. I'm a fan of population-optional board designs to allow one board to cover many different near-similar configurations.

My comment about "sounds better" being listener specific also applies to me. My comment above about the LME49600 sounding better than the BUF634 is pretty meaningless since I've heard them in different amps with different front ends and different layouts. I notice the pinout is the same for both chips in the 5 pin surface mount versions and most of the parameters are similar for either. How about a layout that would allow either buffer chip to be used at the builders discretion? IanAS has a good point about the price difference. If the BUF634 sounds good to any given builder, then great!

I have also read about the stacking option IanAS mentioned with the BUF634s in the DIP version - been meaning to try it myself in my NG98 - and that is really just paralleling those chips, of course.

[RocketScientist]

Originally Posted by agdr

I'm in! I would also likely be interested in a group buy if one occurs.

Great!

Originally Posted by agdr

Unfortunately I saw the "wire" amp thread too late to order, but opc put together a fine amp there. My uses are mainly for unbalanced input through.

I agree he did a good job. Had I made it in on this buy soon enough, I probably would have just bought one of his. My Benchmark DAC1 and Squeezebox Transporter both have balanced outputs. But I think balanced sources are less common for the majority looking for a reasonably priced headphone amp.

I've read several postings of BUF634 vs. LME49600 and the LME seems to consistently come up as the winner. I have a BUF634 headphone amp, the previous version of the HA-INFO NG98. Sounds good, but I think the LME sounds better, just imho.

Good to know. They're pin compatible so anyone would be free to experiment if they want, or save a few bucks and use the BUF634F. I might even be able to work the TO220-5 through hole footprint into the layout but it would require a heatsink.

See my post in the "wire" thread about paralleling two LME49600s per channel. I've read at least one posting where that was tried and (subjectively) sounded better. The BUF634 apparently has 10R on the emitter of each output transistor. The LME49600 shows a similar resistor but I've yet to read a value.

Yeah, the vendors often are intentionally (and understandably) very vague with their "equivalent circuits". But I do wonder to what degree random pairs will play nice and reasonably share the load. That would be something that's easy enough to explore experimentally with some very low value series resistors to enable reading differential currents on a scope.

I generally like the idea of adding the pads for parallel output devices for those with especially low sensitivity or low impedance cans who like their music loud.

How well they share current might be subject to random matching, temperature, etc. It's also possible National may use different fab lines which may have slightly different processes (especially over time). So, at the least, it would be best to use devices with the same date/lot code. The good news is the devices won't self destruct even if they don't share the current at all.

I'm also a bit concerned about the loading to the gain stage with paralleled buffers. You get roughly twice the capacitance and half the impedance. It's a direct connection to the gain stage with no series resistance and paralleling will change the open loop transfer function. But the effects should be easy enough to measure.

I'm a huge fan of "measuring as I go" and have all sorts of custom tests and scripts predefined in the dScope software. It's not unlike writing software. I usually get a much better end result by testing regularly along the way.

For what it is worth, that Jim's Audio LME49710+LME49600 board on eBay (no DC servo), which seems to suffer from layout problems (noise) by some posts I've seen, has two LME49600s in parallel per channel.

Yeah, I found that design and his "modular diamond buffer" board earlier on eBay. Neither is the National reference implementation (as you said no servo), and both seem kind of "random". So I'm not that surprised people have had problems. I've yet to measure any audiophile DIY eBay PCB/kit from Asia that didn't have at least one serious flaw. I don't understand why more of them don't just copy the reference designs and at least try to follow the reference PCB layout?

The KECES amps out of Asia using the LME49600/49710 pair sell for around $400 and have received some fairly good reviews. But when I read you can upgrade to a "pure silver" IEC AC power socket for "even better dynamics" for only $90 more it's hard to take "KECES" seriously and trust their design (and business) objectives.

I will confess I've thought about buying one of the cheaper eBay headphones amps in a nice KECES-like custom enclosure, throw away the main PCB, and just use the chassis, transformer, hardware and other bits.

I have read those posts before by the (ex) National guy saying the metal can versions sounded better. I think he wrote that he was about to look into it when the staffing "bomb" went off there at National. From that little bit of info in the post I would agree that it may be good to socket the LME49720 so that a metal can version could be stuffed in.

That would work. I also might be able to squeeze some extra pads in with the DIP layout for those who want to solder in the TO-99 version without mangling the lead wires as much.

A lot of the "sounds better" stuff seems pretty listener-specific, even if it is backed up with low distortion, low noise, etc tests. So if someone likes the sound of the low cost part, more power to them in my book.

Yeah. Personally I think much of it is purely psychological. But, if it doesn't cost much extra for more piece of mind, and it brings more listening pleasure, why not?

Having the option to use 2 metal can LME49710s instead of the 20 would also be interesting, but I can't think of a good layout way to make that work offhand, while still preserving the ability to use one plastic dip 20.

I would be reluctant to share the gain stages between the two channels in a single package (despite that fact it's done all the time in even expensive high-end audiophile gear). But sharing the servo and gain stage within a channel is much less an issue. There's just not much going on with the servo for it to cause any problems that I can see. And National's reference design helps prove that out.

There are various links around the web of golden ears who heard the national LME49600 reference board and gushed praise all over it. The National guys were seriously obsessive with those designs. And if there was a significant performance gain from using single amp devices, they probably would have done so. It's a $270 board so I don't think cost was much of an issue when they can source the op amps internally.

I'm never a fan of super packed-in circuit boards - I'm kind of heretical there and like some space between my components even if it makes the board larger - so maybe one solution is to put in holes for for the DIP socket and the can in parallel, then populate one or the other.

I agree--especially if there's the possibility of modifications. It's nice to leave room for bigger caps, etc. And it also helps with crosstalk and sometimes other performance areas (although large layouts can also create stability problems).

I've seen a few posts where people building the headamp circuit in the LME49600 datasheet wound up with 10-14mV offset on the output despite the servo circuit. Mine was about 8mV. The "wire" amp didn't seem to have that problem from what I read in the thread. I haven't tried the theory yet, but maybe much more accurate (as in 0.1%) resistors in a few crucial places could solve that problem, and/or offset trim for the input op amps.

Hmmm... that might need some investigation. < 15 mV doesn't bother me that much but I know there are some cans that are very sensitive. What does bother me is a fully direct coupled design with no dc blocking caps anywhere and no protection--especially when the source could be most anything. I have to wonder if some of the offsets people have seen are from their source, RFI (especially with the input floating), poor grounding schemes, etc?

As was discussed in the Wire thread, DC protection might not be a bad idea and that's one of the options I was considering at least laying out if the extra real estate wasn't a problem. The protection scheme could take many forms depending on cost and real estate. It can be done without any series relay contacts in the signal path. Another option might be an LED that just warns of excessive offset.

I've actually read at least one post where someone thought the VR1s were offset trim.

It often happens when sharing DIY projects. That's one reason sites like Headwize and AMB are so great is they form a better repository of information--more of a wiki--than some thread with 80 pages and 800 posts in it. Few want to wade through that much text.

The datasheet circuit shows the various chip's decoupling capacitors going from rail(s) to ground. My best understanding is that is not best practice anymore, rail to rail is preferred as close to the chip (?). I'm sure someone will correct me if this isn't the case.

I think that partly depends on the PCB, your grounding scheme, and the device itself. I'm not sure there's any one right answer. To me it goes back to "measure as I go". The right kinds of measurements often show the subtle differences from various decoupling options, grounding options, etc. That usually makes it easier to figure out what's best for a particular design rather than following some generic rule of thumb. Of course if they all yield identical measurements, then you fall back on audiophile legend....

There actually are heatsinks for the TO-263 with solder rails that solder down next to the chip

Calculating real world power dissipation in a Class-AB output stage is always a bit of a dice roll. In this case it's even worse as different power supply voltages might be used, the loads will vary widely, etc. But even when all that's known, different music still is a big variable with radically different peak to average ratios. I trust the engineers at National did their homework but they do hint that extreme conditions may limit the current available due to high temps. It's easy enough to torture test the finished design and monitor the tab temp and outputs for sine of thermal induced current limiting. But I agree heatsinks, and/or parallel devices might help shrink the design.

Thanks again for the great ideas! I would encourage others to join in. I've done searches here on LME49600, etc. but I'm sure there's more I've missed, haven't thought of or perhaps wrongly dismissed.

[Atilla]

I've got a few of those in my parts drawer and it's always good to put them to use.

I've tried one design, with a simple board from Ebay, that works quite nice but it's got a bit poor grounding scheme. Uses metal-can LME49710 version of the opamps.
EDIT: Right, the one you're talking about up there. I get about 3-5mV DC offset with that one.

With two of those buffers in parallel it can drive small speakers without issues however, I'm not sure there's any benefit of paralleling those. They really mean business.

[IanAS]

Originally Posted by agdr

...use 2 metal can LME49710s instead of the 20 ... I can't think of a good .. way to make that work

[IanAS]

Originally Posted by agdr

chip's decoupling capacitors going from rail(s) to ground. My best understanding is that is not best practice anymore, rail to rail is preferred as close to the chip (?).

I've done both here on this sound card. 100uF Polymer caps from ± rails near the chip(s) to ground and a 0.33uF//0.022uF from rail to rail.

The 0.33 increased the transient response, the 0.022 took away a coarseness. The 0.33 also raised the perceived loudness from the lower mid range downwards, improving (with that system) the tonal balance to a warmer more 'musical' balance.

One 0.33uF was a lot better sounding than three 0.1uF of the same cap type. Also tried 1, 2, 4 and 5 of the 0.1uf and also the 0.33 with 2 x 0.1uF. Up to 0.5uF sounded wrong, dull somehow, soft, blurred. Less than three of the 0.1uF's and there was not as much increase in warmth or transient. A single 0.33uF was cleaner, faster, better transient edges.

[Atilla]

Oh lord, what have you done to the poor Auzentech ;(