Software and source-side upsampling

Upscaling and the M Scalers

Answers on this topic

• Should I upsample in software (HQPlayer, PGGB or my player) before a Chord DAC, and can it match an M Scaler?

Everything he said on this topic (47)

One-line summaries in our words, oldest first. Each link opens the original.

• 2014-08-30Watts says standard upsamplers have too few taps, so let Hugo do it; the exception is Red Reference's 18K WTA 4x upsampler, equivalent to 36,864 Hugo taps. Head-Fi, Aug 2014
• 2015-11-22Watts says not to upsample on a computer before his DACs; Dave's 166 DSP cores outpace a PC, and only the WTA algorithm targets transient timing errors. Head-Fi, Nov 2015
• 2016-08-09Watts explains why a PC cannot replicate DAVE's processing: FPGAs run filter taps in parallel (DAVE 166 DSP cores, Mojo 44 at 0.45 W), and algorithm design also matters. Head-Fi, Aug 2016
• 2017-02-07Watts says PC upsampling has limited taps and conventional algorithms, converting to DSD degrades depth and timing, and data to Mojo, Hugo or Dave should stay bit perfect. Head-Fi, Feb 2017
• 2017-09-06Watts says all Hugo 2 inputs are treated the same, but the source must not up-sample, since PC-based up-samplers lack Hugo 2's processing capability. Head-Fi, Sep 2017
• 2017-10-22Watts says CPUs struggle even with DSD512 DoP, 'input related' tap figures are meaningless pseudo numbers, cascaded FIR stages differ from one flat stage, and only real taps count. Head-Fi, Oct 2017
• 2017-11-15Watts says the M scaler's WTA coefficients match an ideal sinc to 16 bits, needing about 1 million taps and the right algorithm, which PC up-samplers lack. Head-Fi, Nov 2017
• 2018-01-15Watts says upsampling PCM to DSD256 will sound much worse than feeding the original PCM to Qutest, for several reasons. Head-Fi, Jan 2018
• 2018-03-17Watts says never upsample at the source into his DACs; cascading two DAVEs via DX helps slightly versus an M scaler's huge change; isolating FPGA noise was essential for DAVE. Head-Fi, Mar 2018
• 2018-07-26Watts says the M scaler needs bit-perfect input because PC upsamplers lack the processing power and a sinc-like WTA algorithm; its SPDIF output is asynchronous but the DPLL is transparent. Head-Fi, Jul 2018
• 2018-07-27Watts doubts vendors' software tap claims such as Audirvana's 500,000 taps, says PCs are noisy and poor at real time, and that filter structure alters sound. Head-Fi, Jul 2018
• 2018-08-03Watts says a standard PC lacks the CPU power for 1M taps though a GPU PC could, but it still would not be the WTA algorithm, which matters most. Head-Fi, Aug 2018
• 2018-09-16Watts says DSD sounds soft and warm because its delay depends on amplitude, by tens of uS; he advises against software PCM upsampling, and mentions F1 in Singapore. Head-Fi, Sep 2018
• 2018-09-25Watts says the warmth heard is softness from transients misaligned by crude upsampling versus Qutest's WTA filter, and advises a bit-perfect source with no upsampling. Head-Fi, Sep 2018
• 2018-12-17Watts says 2 V is 3.5 dB and 1 V 9.5 dB below 3 V; non-WTA upsamplers cannot predict the M Scaler, as longer conventional filters add near-zero coefficients. Head-Fi, Dec 2018
• 2019-01-20Watts says DAVE cannot switch out WTA2 since its timing errors add artificial softness, and other up-samplers lack taps or WTA-like algorithms and sound unnaturally soft. Head-Fi, Jan 2019
• 2019-03-25Watts comments on upsampling experiments: stopband beyond -150 dB matters little, output rate should suit tap length, his Davina decimation reaches 250 dB, and truncation needs aggressive noise shaping. Head-Fi, Mar 2019
• 2019-09-14Watts explains the trade-off between output rate and sinc accuracy at fixed taps, says 16FS is optimal for 1M taps, and rejects offline 16FS conversion over file size and metadata. Head-Fi, Sep 2019
• 2020-01-02Watts says let the DAC do it: PC algorithms are not WTA and processing is limited without a GPU; he has not listened, but respected owners report much worse sound. Passion for Sound 15:20
• 2020-01-02Watts says anyone who dislikes the WTA filter can use these upscale filters instead, or not buy Chord products. Passion for Sound 17:23
• 2020-11-11Watts compares 1M-tap rectangular sinc, Kaiser and WTA filters by measurement and listening, explains how the WTA equation was tuned, and concludes algorithm matters more than tap length. Head-Fi, Nov 2020
• 2020-11-12Watts, now 60, says WTA on a PC would degrade sound, a custom chip lacks a business case, and FPGAs are far short of a Hugo with built-in M scaler. Head-Fi, Nov 2020
• 2020-11-27Watts advises never upsampling before his DACs and says M Scaler research at 1M taps showed the WTA algorithm is 50 to 500 times more effective than conventional filters. Head-Fi, Nov 2020
• 2020-11-27Watts says to avoid all up-samplers unless they contain a WTA filter, since WTA optimization through listening tests matters much more than tap length. Head-Fi, Nov 2020
• 2021-05-23Watts says PGGB uses an apodizing, slow roll-off filter rather than sinc, so long taps cannot help its transient timing, and true offline sinc sounds like the M scaler. Head-Fi, May 2021
• 2021-05-23Watts quotes PGGB's FAQ on apodizing and says its windowing function turns sinc into something that should not be called sinc; true offline sinc needs no window. Head-Fi, May 2021
• 2021-06-02Watts says PGGB uses a slow roll-off apodizing filter for 44.1k, so many taps cannot fix transient timing; samples he heard sounded flatter, softer and less separated. Head-Fi, Jun 2021
• 2021-06-02Watts quotes PGGB's FAQ on apodizing and says a slow roll-off interpolation filter cannot reconstruct bandlimited signals and seriously degrades transient timing, making long taps pointless. Head-Fi, Jun 2021
• 2021-12-06Watts says HQPlayer and PGGB give no idea of M scaler performance, since many taps with the wrong algorithm give the wrong result, and only WTA minimizes transient timing error. Head-Fi, Dec 2021
• 2021-12-07Watts lists advantages of an FPGA up-sampler over a PC (lower RF noise, clock control, own DSP design) and says 64 bit floating point creates noise floor modulation. Head-Fi, Dec 2021
• 2022-07-14Watts says the M Scaler is essential, PGGB sounded like an apodizing filter, and the M Scaler reconstructs transient timing beyond 16 bits, passing -301 dB via dual BNC. Head-Fi, Jul 2022
• 2023-03-26Watts says licensing the algorithm is possible but a business decision; software like PGGB cannot run in real time, which matters for streaming and video, where he uses scalers. Eric's Hi-Fi 10:28
• 2023-03-26Watts says his NVIDIA streamer allows a one second video delay so he can use an M Scaler; original PGGB files sounded nothing like it, and apodizing cannot reconstruct transients. Eric's Hi-Fi 11:13
• 2023-03-26Watts says he has not heard the latest PGGB version, the original sounded nothing like an M Scaler, and he has no experience of HQPlayer. Eric's Hi-Fi 12:01
• 2023-08-22Watts disputes PGGB's pure-sinc claims, says coefficients must be windowed, and suggests listening for bass pitch, separation, focus and timbre variation. Head-Fi, Aug 2023
• 2023-08-25Watts says early Decca recordings have speed he attributes to aliasing and timing, suspects modern half-band ADC decimation, and says early PGGB sounded poorer than the Hugo M Scaler. Head-Fi, Aug 2023
• 2023-08-26Watts explains as a trade member he cannot rank rivals' filters, would never falsely call one worse, and criticizes PGGB's marketing claims and implementation. Head-Fi, Aug 2023
• 2023-09-14Watts explains why Mojo 2's EQ is loss-less rather than bit perfect, advises against source upsampling, and says an M scaler improves Mojo 2 though algorithm matters more than taps. Head-Fi, Sep 2023
• 2024-01-27Watts says the claim that SRC is inaudible is wrong, advises sending bit-perfect data without SRC or upsampling to his DACs, and says transient timing is vital for perception. Head-Fi, Jan 2024
• 2026-10-11Watts says he cannot judge others' undisclosed methods, but real-time convolution zero-pads buffers through an FFT and adds results, which suffers from FFT uncertainty. Distorsiuni Interioare 35:02
• 2026-10-11Watts says FFT uncertainty is a frequency resolution limit set by batch size, which translates into a transient timing error, and buffer sizes are not disclosed. Distorsiuni Interioare 35:49
• 2026-10-11Watts says his direct interpolation filter avoids this; 64-bit floating point creates small but audible noise floor modulation, compounded by FFT and IFFT, requiring much higher precision. Distorsiuni Interioare 36:27
• 2026-10-11Watts says doing even an M Scaler's processing this way takes huge computing time, so claimed results do not add up for him. Distorsiuni Interioare 37:16
• 2026-10-11Watts says the key issue is the algorithm; nobody discusses transient reconstruction or tuning filters thousands of times, and he gives customers one listening-optimized solution, not many options. Distorsiuni Interioare 38:03
• 2026-10-17Watts says companies claiming more taps use FFT convolution, whose Fourier frequency resolution degrades transient timing, and he is skeptical because FFT sizes are not published. article, Oct 2026
• 2026-10-17Watts says generating coefficients takes forever even on his very fast PC, which is why he is skeptical. article, Oct 2026
• 2026-10-17Watts says the FPGA gives needed precision and a direct FIR filter without FFT tricks, and the algorithm matters far more than tap count, as Blackbird and WTA show. article, Oct 2026