Author's precis: Mike Moffat will certainly cover this in his own blog, but he's still got a few decades of catching up with the present to do...so, until then, you have my perspective on the Bifrost Multibit.
2015, Chapter 17:
The Multibit Revolution Gets Cheap
Well, maybe not
absolutely cheap, because $599 is still a decent chunk of money, even when counted in 2015’s inflation-decimated dollars.
But $599 is what we were charging for the Cobalt 307 back in 1993, and that was considered a breakthrough in terms of value-for-dollar. When you consider that $599 in 2015 is $364 in 1993, the Bifrost Multibit is pretty amazing.
Don’t believe me? Consider:
- Bifrost Multibit uses a 100% multibit DAC, where the Cobalt 307 used a hybrid multibit+delta-digma DAC—and it’s a multibit DAC we would have swapped a testicle for in 1993 (weeeellll, maybe not literally, but you know what I mean)
- Bifrost Multibit uses our proprietary time- and frequency-domain optimized, closed-form filter (AKA the “supercomboburrito” filter), where the Cobalt 307 used an off-the-shelf digital filter
- Bifrost Multibit has 3 inputs, including asynchronous USB 2.0, where Cobalt had only 2—optical and coax
And, the big one:
- Bifrost Multibit is built as an upgradable platform, so future changes in USB input or DAC technology don’t mean instant obsolescence, where the Cobalt 307 was never intended to be upgraded—it was a “disposable” product
So, for about 40% less cost, the Bifrost Multibit delivers a hell of a lot more value—and a hell of a lot less obsolescence—than a DAC Mike and I designed 22 years ago.
So, even if it isn’t “cheap” in absolute terms, it’s certainly moving in the right direction!
Okay. Pause for breath.
I kinda jumped in here, getting all bean-countery on you right up front. I probably should have made it clear that this chapter is about Bifrost Multibit. And yes, it’ll have all the back-story and design details you expect.
But leading with the financial aspect is interesting, isn’t it? It shows that we
can push costs down—and deliver more—in high-end audio.
And that’s something both Mike and I are very proud of.
Multibit on the Desktop
From long before Yggy, Mike had been planning for multibit upgrades all the way up and down the upgradable DAC line. In fact, he’d already made special provisions for the current requirements and the low-voltage power supplies of the Analog Devices SHARC DSP that we were using.
And, as soon as we introduced Yggdrasil, people began wondering how far down the line it would go. Some guessed that Gungnir would be a relatively easy target, but less of you were convinced it would hit the Bifrost.
And for Gungnir, transforming it into a multibit DAC was relatively straightforward, as I’ve covered before. In fact, it was so straightforward that Gungnir Multibit is very similar to Yggdrasil, in terms of number of DACs, type of DACs, and discrete output buffers.
But Bifrost was another matter. Although the electronic provisions were in place for the DSP engine, there was still a lot of debate about which DAC we’d end up using. After all, Bifrost has less than ¼ the total analog board space as the Gungnir, so the packaging would have to be extremely efficient.
In fact, early development had Mike and Dave wondering if we’d have to go in a different direction than the Analog Devices DSP:
“What about the filter?” Mike asked, referring to our supercomboburrito filter (that runs on the DSP that needs quite a bit of board space for it and its accouterments).
“Weeellll, there are some really good sample rate converters we could use,” Dave said.
“Sample rate converters?” I asked. “For what?”
“Instead of the filter,” Mike said.
“Instead of the burrito?” I asked, mouth hanging open. “And
sample rate converters?”
“Not
asynchronous sample rate converters,” Mike said, waving a hand. “
Synchronous. Use it to bring up the input sample rate 4X—synchronously—so you get, say, 176.4 out for 44.1 in, then it gets interesting. If they sound good.”
“There are some good ones out there,” Dave said.
“A sample rate converter?” I repeated, wondering inwardly,
How am I gonna explain this to everyone…hey, we got your shiny new multibit Bifrost here, but it uses a sample rate converter, and yeah, I know we talked some Schiit about sample rate converters, but those were asynchronous and these are synchronous…
“A
good sample rate converter,” Mike corrected.
Still, I shook my head. Copy-wise, it would take some fancy footwork. But I remained silent. Because the Bifrost Multibit was still vapor. There were no prototype boards, no detailed schematics, no decision on a D/A. There was an AD5761 16-bit DAC, similar to the AD5791 and AD5781, but that family is processing-intense, needing headers written in before each sample…which would use up even more board space.
There was also the geriatric AD1851 16-bit audio DAC from Analog Devices, which was paradoxically still in production, and not super expensive. I mentioned that one to Mike and Dave, but they just muttered and left it at that.
And it wasn’t long before Mike found the DAC he really wanted to use.
The Bizarre DAC
“I’ve got it,” Mike said, one day when we were both at Schiit. “The AD5547.”
I frowned. “What the hell is that?”
“It’s a DAC. Look it up.”
I did. “16 bit?” I asked, doubtfully.
“Look harder,” Mike said.
I did. And I realized that the specs on this DAC were really amazing. So amazing, I should call out the relevant points with numbers.
1.It features an integral nonlinearity (INL) plot that's better than +/-0.5LSB. This is a spec they never provide for audio DACs, because (a) it would be terrifying in the case of audio multibit DACs and (b) it is not possible to measure delta-sigma DACs in this way, since the output depends on the preceding and following samples.
2.The THD performance is actually scary good for a 16-bit DAC—far better than 16 bits. 16 bit level THD, from a theoretically perfect 16-bit DAC, is -96dB. Most 16 bit DACs from the Jurassic Age of Digital didn't hit this number. The AD5547 is -104dB, much better than 16 bit.
3.Its noise level is down at the 22 bit level. Think about that for a bit. Why would you need 22-bit equivalent noise for 16-bit audio? The answer is, you don’t. And in the old days, you couldn’t. Many, many DACs from the early years of the PCM era couldn’t do 16-bit noise levels.
And, it was a dual DAC. As in, we’d only need one chip per stereo Bifrost. This meant less board space, and an easier implementation…except for one thing: the AD5547 is a parallel input DAC.
“Huh?” you may be asking. “What’s that, and who cares?”
Well, most DACs take their data in serial form—data goes in on one single pin. You send it that stream of data, plus a few clocks, and it’s happy. Most audio DACs are built so they easily interface with the output of typical USB and SPDIF receiver chips, so it’s kinda like building blocks. Hook up a few pins and go. Even I can do it.
But the AD5547 takes the data in parallel, on 16 separate pins, plus other clocks to ensure it converts samples at the right time, and both at once. Bottom line, the clock logic and data input to this DAC are fundamentally different than audio DACs—so there’s no easy paint-by-numbers implementation. Mike and Dave would be figuring out how to interface a sample rate converter (or, I still hoped, the full burrito filter) to another bizarre DAC—one completely different than the AD5791 and AD5781.
There are many days I’m happy to be an analog designer.
Hard Design Decisions
From there, Mike and Dave seemed to drop the development of a Multibit Bifrost. It made sense—we were getting deep into Yggdrasil production and ramping up Gungnir Multibit, so they had plenty to do on those fronts without any other distractions.
I didn’t realize how close we were to having a Bifrost Multibit until Mike asked, “You know, if Bifrost Multibit had an op-amp analog stage, how upset would you be?”
“Why?” I asked. “Are we really that space-limited? I know a real discrete I/V is a nightmare, but if we do parts on both sides of the board—“
“You need two gain stages per channel,” Mike said. “It’s a multiplying DAC.”
“Oh,” I said, letting that sink in.
“It comes down to the burrito or a discrete I/V,” Mike said. “And even then, I don’t know—“
“Burrito,” I said.
“You sure?”
I nodded. “No question. The burrito is the DAC.”
Mike rubbed his chin and considered for a moment. “How’re you gonna spin this, after all the words about discrete?” Mike asked.
I laughed. “It’s a lot easier to explain this than sample rate converters.”
“So you wouldn’t be upset?”
“No.”
“Good,” Mike grinned, and handed me the finished, working Bifrost Multibit board.
Right Decisions, Stealth Modes, and the Trouble With First Times
Yep. That’s right. Mike and Dave had just gone ahead and done it. And, as far as I’m concerned, they made exactly the right decision to retain the filter and go to op-amps for the I/V stage.
Why? Because discrete I/Vs are (a) complex, (b) have complex power supplies, or (c) both of the above. Any simple discrete I/V was out of the question because we didn’t have the four independent, non-ground-referenced floating supplies they would require. So that meant it would have to be complex.
Complex, as in Theta Gen V level complex. And the discrete I/V I did for the Theta Gen V had something like 250 parts on a 4 x 6” Teflon PCB—for a single channel. The total size of the Bifrost Multibit analog board is 4 x 5.”
Plus, going discrete in the Bifrost Multibit was completely different than our other DACs:
- The discrete stage in the Yggdrasil and GMB are both just buffers. The DACs used there have voltage output. So, they are very simple discrete stages, just four active devices per channel. However, as measurements clearly show, this simplicity does not compromise distortion performance (this is usually the penalty paid...simple discrete amps typically have high THD.) That's why when you see some "2-PPM wonder amp" it usually has about 80 active devices per channel. We can argue till the cows come home which sounds better.
- The discrete stage in the standard Bifrost and Gungnir is actually a small amp stage—not exactly a discrete op-amp, due to its very specific gain structure and open-loop bandwidth beyond the audio band--but it also takes a voltage output of the DAC, amplifies it a bit, and passes it on. No I/V necessary.
- The AD5547 is a multiplying DAC, which means it needs two very well-matched gain stages for exceptional accuracy, not just a single I/V converter. Which means it would need twice the circuitry. Fully discrete? As I mentioned, aha, no.
So…even with surface mount components, the math was simple. It all pointed to “nope, sorry, divide by zero, sums to nothin, system is unstable, ain’t gonna work around here.” Anyone looking for a comparison to a mythical discrete-output board for Bifrost Multibit will be disappointed…because
we simply never tried to do it.*
*And no, there is no conspiracy where we’re holding back a better analog board that will magically appear in 6-8 months, when sales start slowing down in the spring.**
**Now, that doesn’t say we’ll never have a better analog board for Bifrost. Hell, it’s possible that even better multibit DACs will be available in the future. Or a new delta-sigma technology will force Mike to recant and go 100% pro-DSD. Though the third sentence is, I think, about as likely as a meteor hitting me on the head before I finish this current sentence.***
***And, sentence finished. There you go.
So, was that it? Did Mike just hand me a board and call it a day? Is that all the development drama I could come up with?
Nope. Though Bifrost Multibit, like many of our latest products, was remarkably drama-free. The most fun we had was taking the finished Bifrost Multibit to a meet and the Schiit Show, 100% incognito, to see if anyone noticed.
And, you know what? Some
did.
“That’s not a stock Bifrost,” one of our early listeners told me at the Schiit Show.
“Sure it is,” I told him. “It says so right on the product mat.” (We’d had mats made up with product specs and pricing, so people new to the gear knew what they were listening to—one of the smartest things we’ve ever done.)
“No it’s not!” he insisted. “Is there, a…an upgrade, ah, coming?”
“You know we don’t talk about products before they’re released,” I teased, grinning.
“Come on! Tell me.”
And I did, after asking him not to repeat it on the forums. He smiled even wider. “October, huh?”
“We hope,” I reminded him. “Lots of things can change.”
And lots of things could change. We could possibly have introduced Bifrost Multibit earlier—well, maybe—but, more likely, it could have been pushed later. That’s why we don’t talk about new products. Because they aren’t products until they’re on the shelf.
Now, for the arms-crossed brigade out there saying, “Well, you heard it from the top, Schiit punks us at the shows, there’s no telling if what they’re showing is what it actually sounds like,” let me clarify:
- We took the prototype to at least one meet, where pretty much everyone knew what it was (we told them and asked them not to talk.)
- We took two prototypes to the Schiit Show, because, hey, we thought it would be fun to see if anyone notices.
Beyond that, here’s our super-secret method for choosing the best-sounding products to display at shows:
- Denise asks for what we need.
- Alex pulls it off a shelf and sends it.
Yes, that’s right. No special tweaks, no magic burn-in protocol, no hand-picked-for-best-measurements, no take-it-out-and-see-if-it-works. Just off the shelf and on to the table. As if you purchased it.
And…if you’re still miffed at us having a bit of fun at a meet and our own show, please accept my apologies. We were just having a little fun.
Now, while we were having fun at the shows and beating up on the Bifrost Multibit prototypes, lots of other wheels were turning. Because Bifrost Multibit wasn’t just a product launch, it was a product launch, a product change, an upgrade launch…and also the launch of a new AKM chip in the delta-sigma Bifrost.*
*Yes, we did a delta-sigma update too with the AK4490. And that’s all that needs to be said about that. Not that it’s bad. Just not super interesting. Except for the price drop.
Why is this a big deal? Because we couldn’t simply set up a single product page and make it live when the time came. We had to:
- Verify specs for the Bifrost Multibit and the Bifrost “4490”
- Write copy for the Bifrost Multibit, the Bifrost “4490”, the Bifrost Multibit Upgrade, and the Bifrost 4490 upgrade
- Get photos taken of both boards, both on the motherboard and off
- Set up the upgrade system to handle the coming Bifrost upgrades
- Buy tonnage of parts and make about a billion Bifrost Multibits and Multibit boards, so we’d be ready for the onslaught of upgrades—while we’d had some practice with Gungnir, Bifrost had about 12x the number of products in the field, so the potential volumes were daunting
- Argue about, and set, pricing—increases in our run size and lower parts cost meant that we could reduce the price of the Bifrost and have a very aggressive price set on the Bifrost Multibit*
- Get the tonnage of boards through the PCB assemblers
- Produce finished Bifrost Multibits to have on shelves for the launch
*Mike correctly predicted the consternation from the price drop. Yes, it’s an unusual thing to do in audio, but I’m still convinced it’s the
right thing. Even if the bitchfest that ensued made my cynical side say, “Ya shoulda taken tha money!”
Now, of course we can’t get into production without at least one little surprise. Considering the complexity and scale of this product,
one little surprise would be a minor miracle. So we ended up with two.
- Surprise the First. The prototypes (that we had listened to for hours, taken to shows, etc.) were one sample off between channels. This is what happens when you’re working on interfacing a DSP to never-before-used DACs. What this translates to is phase shift between channels that varies with frequency. No, we hadn’t heard it. No, nobody else had heard it, though some listeners had commented on “there being something discontinuous between the lows and highs” and “woolly bass.” Given the price point and musical performance, we didn’t discover the error until we were into the final qualification phase. This also goes to show why you shouldn’t bring prototypes to shows…the prototypes weren’t all they could be. Har har. In punking the showgoers, we’d punked ourselves.
- Surprise the Second. The production first articles had a singing voltage reference. As in, it was oscillating. It was at such a low level you’d probably never hear it, but it showed up on the Stanford right quick. We made a couple of component changes and fixed it up before the full run.
Which meant that the Bifrost Multibits that went to RMAF were not only the first production Bifrosts (with both Surprises addressed) but were also correctly labeled as Bifrost Multibits. On Friday, October 2
nd, Can-Jam attendees were able to hear Bifrost Multibit in its full glory for the first time.
And, I’m thrilled to say that not only did we announce on time, but we actually shipped on the day of the announcement. Yes. As in, the first Bifrost Multibits shipped within minutes of their orders.
Almost like we’re getting good at this, I sometimes think, then worry about the next intro…
Because you can always get caught off guard.
